WO2022052744A1 - 会话信息处理方法、装置、计算机可读存储介质及设备 - Google Patents
会话信息处理方法、装置、计算机可读存储介质及设备 Download PDFInfo
- Publication number
- WO2022052744A1 WO2022052744A1 PCT/CN2021/112732 CN2021112732W WO2022052744A1 WO 2022052744 A1 WO2022052744 A1 WO 2022052744A1 CN 2021112732 W CN2021112732 W CN 2021112732W WO 2022052744 A1 WO2022052744 A1 WO 2022052744A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- vector
- session
- information
- keyword
- conversation
- Prior art date
Links
- 238000003860 storage Methods 0.000 title claims abstract description 45
- 230000010365 information processing Effects 0.000 title claims abstract description 31
- 238000003672 processing method Methods 0.000 title claims abstract description 17
- 239000013598 vector Substances 0.000 claims abstract description 408
- 238000000034 method Methods 0.000 claims abstract description 61
- 239000011159 matrix material Substances 0.000 claims description 43
- 238000012545 processing Methods 0.000 claims description 38
- 238000013528 artificial neural network Methods 0.000 claims description 35
- 238000000926 separation method Methods 0.000 claims description 23
- 238000004364 calculation method Methods 0.000 claims description 19
- 230000009467 reduction Effects 0.000 claims description 17
- 230000000306 recurrent effect Effects 0.000 claims description 15
- 238000004590 computer program Methods 0.000 claims description 13
- 238000004458 analytical method Methods 0.000 claims description 12
- 238000004422 calculation algorithm Methods 0.000 claims description 7
- 230000015654 memory Effects 0.000 claims description 7
- 238000013473 artificial intelligence Methods 0.000 abstract description 3
- 230000006870 function Effects 0.000 description 56
- 238000010586 diagram Methods 0.000 description 20
- 230000003993 interaction Effects 0.000 description 16
- 238000005516 engineering process Methods 0.000 description 10
- 238000004891 communication Methods 0.000 description 8
- 238000000605 extraction Methods 0.000 description 8
- 230000009286 beneficial effect Effects 0.000 description 6
- 230000002452 interceptive effect Effects 0.000 description 5
- 230000008569 process Effects 0.000 description 5
- 241001489705 Aquarius Species 0.000 description 4
- 230000003287 optical effect Effects 0.000 description 4
- 238000011156 evaluation Methods 0.000 description 3
- 238000002372 labelling Methods 0.000 description 3
- 230000009471 action Effects 0.000 description 2
- 230000004913 activation Effects 0.000 description 2
- 210000004027 cell Anatomy 0.000 description 2
- 230000008094 contradictory effect Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 239000000835 fiber Substances 0.000 description 2
- 230000004927 fusion Effects 0.000 description 2
- 230000000644 propagated effect Effects 0.000 description 2
- 239000004065 semiconductor Substances 0.000 description 2
- 230000006978 adaptation Effects 0.000 description 1
- 230000002457 bidirectional effect Effects 0.000 description 1
- 239000003795 chemical substances by application Substances 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000013480 data collection Methods 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 238000013500 data storage Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 230000000670 limiting effect Effects 0.000 description 1
- 239000004973 liquid crystal related substance Substances 0.000 description 1
- 238000007477 logistic regression Methods 0.000 description 1
- 238000007726 management method Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 239000013307 optical fiber Substances 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 230000002441 reversible effect Effects 0.000 description 1
- 230000033764 rhythmic process Effects 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 230000000087 stabilizing effect Effects 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
- 238000012549 training Methods 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
- 230000001960 triggered effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
- G06F40/35—Discourse or dialogue representation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/33—Querying
- G06F16/332—Query formulation
- G06F16/3329—Natural language query formulation or dialogue systems
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
- G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/205—Parsing
Definitions
- the present application relates to the technical field of artificial intelligence, and in particular, to a session information processing method, a session information processing apparatus, a computer-readable storage medium, and an electronic device.
- a session information processing method a session information processing apparatus, a computer-readable storage medium, and an electronic device are provided.
- a session information processing method executed by an electronic device, the method comprising:
- the attribute structure and session structure corresponding to the target object are obtained; wherein, the session structure is used to represent the session information;
- the keyword set includes the keywords in the attribute structure and the keywords in the session structure;
- the matching degree between the target object and the session information is determined according to the joint semantic vector and the joint structure vector, and the session information is output if the matching degree satisfies a preset condition.
- a session information processing device comprising: a data acquisition unit, a vector calculation unit and a condition determination unit, wherein:
- a data acquisition unit configured to acquire the attribute structure and the session structure corresponding to the target object when the session information to be output is detected; wherein, the session structure is used to represent the session information;
- a vector calculation unit used for calculating the vector set corresponding to the keyword set; wherein, the keyword set includes the keywords in the attribute structure and the keywords in the session structure;
- the vector computing unit is also used to generate a joint semantic vector according to the vector set, and generate a joint structure vector according to the attribute structure and the session structure;
- the condition determination unit is configured to determine the matching degree between the target object and the session information according to the joint semantic vector and the joint structure vector, and output the session information if the matching degree satisfies a preset condition.
- An electronic device comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to execute the method described in any one of the above by executing the executable instructions .
- a computer program product or computer program comprising computer instructions stored in a computer-readable storage medium.
- the processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to cause the computer device to perform the methods provided in the various optional implementations described above.
- FIG. 1 shows a schematic diagram of an exemplary system architecture of a session information processing method and a session information processing apparatus to which embodiments of the present application can be applied.
- FIG. 2 shows a schematic structural diagram of a computer system suitable for implementing the electronic device according to the embodiment of the present application.
- FIG. 3 schematically shows a flowchart of a method for processing session information according to an embodiment of the present application.
- FIG. 4 schematically shows a schematic diagram of a joint semantic vector generation architecture according to an embodiment of the present application.
- FIG. 5 schematically shows a schematic frame diagram of a session information processing system according to an embodiment of the present application.
- FIG. 6 schematically shows a flowchart of a method for processing session information according to an embodiment of the present application.
- FIG. 7 schematically shows a schematic diagram of a user interface for displaying session information according to an embodiment of the present application.
- FIG. 8 schematically shows a schematic diagram of a user interface for displaying session information according to an embodiment of the present application.
- FIG. 9 schematically shows a schematic diagram of a user interface for displaying session information according to an embodiment of the present application.
- FIG. 10 schematically shows a structural block diagram of an apparatus for processing session information according to an embodiment of the present application.
- Example embodiments will now be described more fully with reference to the accompanying drawings.
- Example embodiments can be embodied in various forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this application will be thorough and complete, and will fully convey the concept of example embodiments to those skilled in the art.
- the described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.
- numerous specific details are provided in order to give a thorough understanding of the embodiments of the present application.
- those skilled in the art will appreciate that the technical solutions of the present application may be practiced without one or more of the specific details, or other methods, components, devices, steps, etc. may be employed.
- well-known solutions have not been shown or described in detail to avoid obscuring aspects of the application.
- FIG. 1 shows a schematic diagram of a system architecture of an exemplary application environment to which a session information processing method and a session information processing apparatus according to embodiments of the present application can be applied.
- the system architecture 100 may include one or more of terminal devices 101 , 102 , 103 , a network 104 and a server 105 .
- the network 104 is a medium used to provide a communication link between the terminal devices 101 , 102 , 103 and the server 105 .
- the network 104 may include various connection types, such as wired, wireless communication links, or fiber optic cables, among others.
- the terminal devices 101, 102, 103 may be various electronic devices with display screens, including but not limited to desktop computers, portable computers, smart phones, tablet computers, and the like. It should be understood that the numbers of terminal devices, networks and servers in FIG. 1 are merely illustrative. There can be any number of terminal devices, networks and servers according to implementation needs.
- the server 105 may be a server cluster composed of multiple servers, or the like.
- the server 105 may be an independent physical server, or a server cluster or a distributed system composed of multiple physical servers, or may provide cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud Cloud servers for basic cloud computing services such as communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms.
- the terminal may be a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited thereto.
- the terminal and the server may be directly or indirectly connected through wired or wireless communication, which is not limited in this application.
- Cloud technology refers to a hosting technology that unifies a series of resources such as hardware, software, and network in a wide area network or a local area network to realize the calculation, storage, processing and sharing of data.
- cloud computing included in the above-mentioned basic cloud computing services refers to the delivery and use mode of IT infrastructure, which refers to obtaining required resources in an on-demand and easily scalable manner through the network; cloud computing in a broad sense refers to the delivery of services and usage patterns, which refer to obtaining the required services in an on-demand and easily scalable manner through the network.
- Such services can be IT and software, Internet-related, or other services.
- Cloud computing is grid computing (Grid Computing), distributed computing (Distributed Computing), parallel computing (Parallel Computing), utility computing (Utility Computing), network storage (Network Storage Technologies), virtualization (Virtualization), load balancing (Load Balance) Balance) and other traditional computer and network technology development and integration products.
- cloud storage included in basic cloud computing services is a new concept extended and developed from the concept of cloud computing.
- Distributed cloud storage system (hereinafter referred to as storage system) refers to the Technology and distributed storage file system and other functions, integrate a large number of different types of storage devices (storage devices are also called storage nodes) in the network through application software or application interfaces to work together to provide external data storage and business access functions. of a storage system.
- the storage method of the storage system is as follows: creating a logical volume, and when creating a logical volume, a physical storage space is allocated to each logical volume, and the physical storage space may be composed of a storage device or disks of several storage devices.
- the client stores data on a logical volume, that is, the data is stored on the file system.
- the file system divides the data into many parts, each part is an object, and the object contains not only data but also data identification (ID, ID entity) and other additional information, the file system writes each object into the physical storage space of the logical volume, and the file system records the storage location information of each object, so that when the client requests to access data, the file system can The storage location information of the object allows the client to access the data.
- the process of allocating physical storage space by the storage system to the logical volume specifically: according to the capacity estimation of the objects stored in the logical volume (this estimation often has a large margin relative to the actual capacity of the objects to be stored) and independent redundant disks Array (RAID, Redundant Array of Independent Disk) group, which divides the physical storage space into strips in advance, and a logical volume can be understood as a stripe, thereby allocating physical storage space for the logical volume.
- RAID Redundant Array of Independent Disk
- the database (Database) included in the basic cloud computing service can be regarded as an electronic filing cabinet—a place where electronic files are stored, and users can perform operations such as adding, querying, updating, and deleting data in the files.
- the so-called “database” is a collection of data that is stored together in a certain way, can be shared with multiple users, has as little redundancy as possible, and is independent of applications.
- the big data included in the basic cloud computing service refers to the data collection that cannot be captured, managed and processed by conventional software tools within a certain time frame, and requires a new processing mode to have stronger decision-making power, Massive, high-growth and diverse information assets for insight discovery and process optimization capabilities.
- big data is also attracting more and more attention, and big data requires special technologies to efficiently process a large amount of data that tolerates elapsed time. Technologies applicable to big data, including massively parallel processing databases, data mining, distributed file systems, distributed databases, cloud computing platforms, the Internet, and scalable storage systems.
- the session information processing method provided by the embodiment of the present application is generally executed by the server 105 , and accordingly, the session information processing apparatus is generally set in the server 105 .
- the session information processing method provided by the embodiments of the present application can also be executed by the terminal equipment 101, 102 or 103, and correspondingly, the session information processing apparatus can also be set on the terminal equipment 101, 102 or 103.
- the session information processing apparatus can also be set on the terminal equipment 101, 102 or 103.
- no special limitation is made in this exemplary embodiment.
- the server 105 may acquire the attribute structure and the session structure corresponding to the target object when detecting the session information to be output; wherein, the session structure is used to represent the session information; The vector set corresponding to the set; wherein, the keyword set includes the keywords in the attribute structure and the keywords in the session structure; the joint semantic vector is generated according to the vector set, and the joint structure vector is generated according to the attribute structure and the session structure; according to the joint semantic vector and the joint structure vector to determine the matching degree between the target object and the session information, and output the session information if the matching degree satisfies the preset condition.
- FIG. 2 shows a schematic structural diagram of a computer system suitable for implementing the electronic device according to the embodiment of the present application.
- a computer system 200 includes a central processing unit (CPU) 201 that can be loaded into a random access memory (RAM) 203 according to a program stored in a read only memory (ROM) 202 or a program from a storage section 208 Instead, various appropriate actions and processes are performed.
- RAM 203 various programs and data required for system operation are also stored.
- the CPU 201, the ROM 202, and the RAM 203 are connected to each other through a bus 204.
- An input/output (I/O) interface 205 is also connected to the bus 204 .
- the following components are connected to the I/O interface 205: an input section 206 including a keyboard, a mouse, etc.; an output section 207 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 208 including a hard disk, etc. ; and a communication section 209 including a network interface card such as a LAN card, a modem, and the like. The communication section 209 performs communication processing via a network such as the Internet.
- a drive 210 is also connected to the I/O interface 205 as needed.
- a removable medium 211 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is mounted on the drive 210 as needed so that a computer program read therefrom is installed into the storage section 208 as needed.
- embodiments of the present application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the method illustrated in the flowchart.
- the computer program may be downloaded and installed from the network via the communication portion 209 and/or installed from the removable medium 211 .
- CPU central processing unit
- various functions defined in the method and apparatus of the present application are performed.
- users can chat with other users through a platform with a chat function.
- user B may be in a busy state and cannot meet user A's needs in time.
- different users since different users generally prefer different chat objects, when the preferred user B cannot chat, it may be that user C is in a non-busy state but user C is not the user preferred by user A. It can be seen that the situation that the user's chat needs cannot be met in a timely manner occurs from time to time.
- the network avatar can be designed so that the user can chat with the preferred network avatar at any time when there is a need for chatting, which can meet the user's real-time chat needs.
- the network avatar can correspond to a dialogue system, and the dialogue system can generate corresponding reply information according to the attribute information of the network avatar and the chat content sent by the user, so as to realize the instant chatting requirement of the user. Purpose.
- NLI Natural Language Interaction
- the attribute information of the network avatar is as follows:
- the above attribute information can be modeled.
- the dialogue information of Xiao H generated according to the chat content can include the following options R1 ⁇ R3: R1: I'm very happy You can come to City A; R2: I also want to go to City A one day to play; R3: After you come, I will take you to visit XX University. Among them, since the location of the small H is city A, the generated R2 does not match the attribute information of the small H.
- the present exemplary embodiment provides a session information processing method.
- the session information processing method may be applied to the foregoing server 105, and may also be applied to one or more of the foregoing terminal devices 101, 102, and 103, which is not particularly limited in this exemplary embodiment.
- the session information processing method may include the following steps S310 to S350.
- Step S310 When the session information to be output is detected, acquire the attribute structure and the session structure corresponding to the target object; wherein, the session structure is used to represent the session information.
- Step S320 Calculate a vector set corresponding to the keyword set; wherein, the keyword set includes keywords in the attribute structure and keywords in the conversation structure.
- Step S330 Generate a joint semantic vector according to the vector set.
- Step S340 Generate a joint structure vector according to the attribute structure and the session structure.
- Step S350 Determine the matching degree between the target object and the session information according to the joint semantic vector and the joint structure vector, and output the session information if the matching degree satisfies a preset condition.
- this application can be applied to the dialogue system between the user and the avatar to improve the matching degree between the information that the avatar replies to the user and the setting of the avatar, thereby making the avatar perceived by the user more vivid and avoiding the appearance of the avatar.
- Implementing the method shown in FIG. 3 can improve the accuracy of the matching degree between the calculated session information and the target object, so that the session information can be output under the condition that the matching degree satisfies the preset conditions, so that the response of the target object can be improved.
- Session information matches its attribute structure.
- the consistency between the target object and the output session information can also be improved, and the authenticity of the target object can be improved.
- step S310 when the session information to be output is detected, the attribute structure and session structure corresponding to the target object are acquired; wherein, the session structure is used to represent the session information.
- the target object may be a real user participating in the session, or may be a virtual user (eg, an AI host) participating in the session, which is not limited in this embodiment of the present application.
- the above-mentioned attribute structure can be used as a personalized description of the target object, and the attribute structure can include the following keywords: gender, address, constellation, age, occupation, and the like.
- the attribute structure can be expressed as: ⁇ gender: male, location: S city, constellation: Leo, age: 20, occupation: researcher ⁇ or ⁇ gender: female, location: A city, constellation: Leo ⁇ and so on.
- the attribute structure can be represented by multiple attribute keywords
- the session structure can be represented by multiple session keywords.
- the session information may be a dialogue message in which the target object communicates with other session objects.
- the session information is the reply to the chat message sent by the real user.
- conversation message Session keywords are derived from session information.
- the conversation structure can be expressed as: ⁇ I, very, happy, you, can, come, come, A city ⁇ , wherein, the conversation structure can also include the semantic relationship between the keywords.
- the above method before acquiring the attribute structure and the conversation structure corresponding to the target object, the above method further includes: when it is detected that the newly added chat message is displayed in the conversation window, determining the conversation information corresponding to the newly added chat message;
- the conversation window includes a target object and at least one conversation object, and the newly added chat message corresponds to the conversation object.
- the newly added chat message may be a newly added message for dialogue with the target object, such as a message sent when a real user dialogues with a virtual user.
- the newly added chat message may include: text information, text information + picture information, or text information + emoticons.
- the newly added chat message may include information corresponding to the conversation object (eg, @user1).
- the method may further include: detecting whether there is a newly added chat message in the session window according to a preset unit duration (eg, 10 ms).
- determining the session information corresponding to the newly added chat message includes: selecting session information from a preset information database according to the newly added chat message; or, generating session information according to the newly added chat message.
- the preset information library is used to store various types of preset session information that can be called.
- the method of selecting session information from the preset information database according to the newly added chat message may be: determining a chat keyword (eg, weather) in the newly added chat message, and selecting the session information from the preset information database according to the chat keyword (For example, "It's cloudy today", “It won't rain today", or "The sun won't come out today”).
- the method of generating the session information according to the newly added chat message may be: determining the chat keyword in the newly added chat message, and generating the session information according to the chat keyword and the attribute structure of the target object.
- the session information may include attribute keywords in the attribute structure.
- the corresponding session information can be determined according to the content of the newly added chat message input by the user, which can improve the user's chat experience and avoid unanswered questions to a certain extent.
- acquiring the attribute structure and the session structure corresponding to the target object includes: reading the attribute structure corresponding to the target object; and performing structured processing on the session information to obtain the session structure.
- the target object may correspond to multiple sets of attribute structures.
- the multiple sets of attribute structures corresponding to the target object H may be: ⁇ gender: female, age: 16, location: city A, constellation: Aries ⁇ , ⁇ gender: female , Age: 16, Location: City B, Constellation: Leo ⁇ and ⁇ Gender: Female, Age: 16, Location: City C, Constellation: Virgo ⁇ .
- the method of reading the attribute structure corresponding to the target object may be: determining the target object setting corresponding to the newly added chat message, and reading the attribute structure corresponding to the target object setting; wherein, different target object settings correspond to different Attribute structure, users can choose the target object setting according to their preferences.
- the following steps may also be performed: representing different target objects through different avatars to set and display avatars; when detecting an interactive operation acting on the target avatar, The target object setting corresponding to the target avatar is determined as the target object setting corresponding to the newly added chat message.
- the attribute structure and session structure of the target object can be determined, which is beneficial to determine the matching degree between the target object and the session information according to the attribute structure and the session structure, so as to avoid the output session information and the target object. Objects do not match.
- performing structured processing on the session information to obtain the session structure includes: determining a word vector corresponding to the session information according to a word vector model; performing text structure analysis on the session information according to the word vector, and obtaining the Semantic relationships between language units; conversational information is represented as conversational structures based on semantic relationships.
- the text structure analysis may include semantic dependency analysis, semantic dependency analysis (Semantic Dependency Parsing, SDP), which is used to analyze the semantic relationship between each language unit in the session information, and presents it as a dependency tree DT (dependency tree) Semantic relationship.
- the semantic relationship may include the following types: agent relationship (Agt), party relationship (Exp), affective relationship (Aft), consular relationship (Poss), subject relationship (Pat), guest relationship (Cont) and Success relationship (Prod) and so on.
- conversational structures can be used to represent language units and the semantic relationships between language units.
- the word vector model can be GloVe (Global Vectors for Word Representation). GloVe is a word representation tool based on global word frequency statistics, which can express a word as a vector composed of real numbers, which is used to characterize the semantic characteristics of the word .
- the text structure analysis is performed on the conversation information according to the word vector, and the semantic relationship between the language units in the conversation information can be obtained by: performing semantic dependency analysis on the conversation information according to the word vector, and obtaining the semantic relationship between the language units in the conversation information.
- the semantic relationship of ; among them, language units (such as, I, very, happy, you, can, come, come, A city) can be conversation keywords.
- abstract semantic representation (AMR) is performed on the conversation information according to the word vector, and the semantic relationship between the language units in the conversation information is obtained.
- Directed acyclic graph AMR can abstract real words into concepts as nodes on a single-root directed acyclic graph, and abstract function words without practical meaning into edges.
- the above method may further include the following steps:
- the model parameters of the word vector model are adjusted according to the loss function expression and the sample output result of the word vector model, so as to realize the pre-training of the word vector model.
- the element X ij of the co-occurrence matrix X may represent the number of co-occurrences of the word i and the context word j in a context window of a preset size.
- f(X ij ) can be expressed as
- ⁇ is a constant, eg, 0.75.
- a failure prompt is returned.
- the structured processing of the session information can be realized through the semantic analysis of the session information, which is beneficial to determine the matching degree between the target object and the session information.
- performing structured processing on conversation information to obtain a conversation structure includes: comparing the conversation information with a preset vocabulary set to determine keywords in the conversation information; Relationships build session structures that contain keywords in session information.
- the preset vocabulary set may include pre-stored vocabulary
- the preset vocabulary relationship may be a preset one-to-one correspondence, which is used to represent the relationship between adjacent language units.
- the session structure corresponding to the session information can be determined according to the preset information, so that the matching degree can be jointly determined based on the session structure and the attribute structure, and the accuracy of calculating the matching degree can be improved.
- step S320 a vector set corresponding to the keyword set is calculated; wherein, the keyword set includes keywords in the attribute structure and keywords in the conversation structure.
- the vector set may include: a keyword vector, a position vector, a category vector, and a separation vector.
- the attribute structure is ⁇ Gender: Female, Location: Beijing, Constellation: Leo ⁇
- the conversation structure is ⁇ I, Very, Happy, You, Can, Come, Arrived, City A ⁇
- the keyword set can be : ⁇ gender, female, location, Beijing, constellation, Leo, [SEP], I, very, happy, you, can, come, come, A city ⁇ ; where, [SEP] is a clause symbol, used to separate the key vocabulary material.
- calculating the vector set corresponding to the keyword set includes: determining the keyword vector of each keyword in the keyword set according to a preset vector relationship; determining the corresponding keyword according to the position information of each keyword
- the position vector of each keyword is determined according to the corresponding category of each keyword; the category vector of each keyword is determined according to the attribute structure and session structure; the separation vector of each keyword in the attribute structure is the same, and the separation vector of each keyword in the session structure is the same.
- the separation vectors of the keywords are the same; the keyword vector, position vector, category vector and separation vector corresponding to the same keyword are added to each other to obtain the vector to be processed corresponding to each keyword respectively;
- the set of vectors to be processed is determined as a vector set.
- the preset vector relationship may be used to represent a one-to-one correspondence between different keywords and keyword vectors
- the keyword vectors may be represented as a list of real numbers
- the keyword vectors corresponding to each keyword in the keyword set may correspond to in the same dimension (eg, 768).
- the above-mentioned parametric addition is used to represent: the sum of elements in the same position among the keyword vector, the position vector, the category vector and the separation vector, for example, the keyword vector corresponding to the keyword "weather” (1, 1, 1), the position vector (2, 2, 2), the category vector (3, 3, 3) and the separation vector (4, 4, 4) are added (1+2+3+4, 1+2+ 3+4, 1+2+3+4), the to-be-processed vector (10, 10, 10) corresponding to the keyword "weather” can be obtained.
- the manner of determining the position vector corresponding to each keyword according to the position information of each keyword may be as follows: if the position information is an even number, perform sinusoidal coding on the keyword corresponding to the position information; if the position information is an odd number, then Cosine coding is performed on the keywords corresponding to the position information, thereby obtaining position vectors corresponding to each keyword.
- the manner of determining the category vector of each keyword according to the category corresponding to each keyword may be: performing a unified category vector representation for keywords corresponding to the same category (for example, "gender” and "male"), Keywords of different categories correspond to different category vectors.
- the method of determining the separation vector of each keyword according to the attribute structure and the session structure may be: the keywords belonging to the attribute structure are represented by the same separation vector, and the keywords belonging to the session structure are represented by the same separation vector. Represents that the keywords of the attribute structure are different from the separation vectors corresponding to the keywords of the session structure.
- vector fusion can be performed to obtain the to-be-processed vector corresponding to each keyword, so as to calculate the matching between the target object and the session information according to the to-be-processed vector set to improve the calculation accuracy of the matching degree.
- the above method before determining the position vector corresponding to each keyword according to the position information of each keyword, the above method further includes: splicing the keywords in the attribute structure and the keywords in the session structure; The resulting keyword arrangement sequence marks the position information of each keyword in the keyword set in turn, and obtains the position information of each keyword.
- the splicing result includes attribute keywords of the attribute structure and session keywords of the session structure; the location information can be represented by numerical values.
- the method of splicing the keywords in the attribute structure and the keywords in the session structure may be: according to the clause symbols, splicing the keywords in the attribute structure and the keywords in the session structure in order (for example, ⁇ gender, female, location, Beijing, constellation, Leo, [SEP], I am, very, happy, you, can, come, come, A city ⁇ ).
- the location information marked in the order of the keywords can be represented as follows:
- the keyword can be marked with position information, which is beneficial to calculate the position vector of the keyword, so as to improve the accuracy of the subsequent calculation of the matching degree.
- step S330 a joint semantic vector is generated according to the vector set.
- the method of generating the joint semantic vector according to the vector set can be: input the vector set into the BERT (Bidirectional Encoder Representations from Transformers) model, so that BERT generates the joint semantic vector according to the vector set, and the joint semantic vector is used to represent each key.
- the output probability of the word; among them, BERT is a language analysis model.
- step S340 a joint structure vector is generated according to the attribute structure and the session structure.
- the method of generating the joint structure vector according to the attribute structure and the session structure may be: inputting the attribute structure and the session structure into a Long Short Term Memory network (LSTM) model for a component tree (Tree) (ie, Tree LSTM), so that Tree LSTM generates a joint structure vector according to the attribute structure and session structure, and the generated joint structure vector is a vectorized representation of the attribute structure and the session structure.
- LSTM Long Short Term Memory network
- Ree component tree
- the generated joint structure vector is a vectorized representation of the attribute structure and the session structure.
- the preset loss function can be a regression loss function, a squared error loss function, an absolute error loss function, and a Huber loss.
- Function, two-class loss function, two-class cross entropy, Hinge loss function, multi-class loss function, multi-class cross entropy loss, or KL divergence (Kullback Leibler Divergence Loss) the embodiments of the present application are not limited.
- generating a joint semantic vector according to the vector set includes: generating a matrix group corresponding to each vector to be processed according to the vector set, where the matrix group includes a query matrix, a key matrix, and a value matrix; Calculate the self-attention vector corresponding to each vector to be processed in the corresponding matrix groups; calculate the feature vector corresponding to the self-attention vector according to the feedforward neural network; calculate the joint semantics corresponding to the feature vector based on the self-attention algorithm and the feedforward neural network vector.
- the step of generating a joint semantic vector according to the vector set can be implemented based on the BERT model; wherein, the BERT model adopts the encoding and decoding mode of the machine translation (Transformer) model, and the Transformer model consists of an encoding component and a decoding component.
- the encoding component consists of It consists of multiple encoders, and the decoding component consists of multiple decoders, and the number of encoders and decoders is the same; among them, each encoder includes a feedforward neural network and a self-attention layer, and each decoder includes Feedforward neural networks, encoder-decoder attention layers, and self-attention layers.
- the method of generating the matrix groups corresponding to each to-be-processed vector according to the vector set may be: generating at least two reference matrix groups corresponding to each to-be-processed vector; The reference matrix groups corresponding to the same vector to be processed are spliced to obtain a matrix group corresponding to each vector to be processed; wherein, the number of reference matrix groups is consistent with the number of self-attention heads (eg, 8).
- the method of generating a matrix group corresponding to each vector to be processed according to the vector set may be: determine the query weight matrix ⁇ Q , the key weight matrix ⁇ K and the value weight matrix ⁇ V corresponding to each vector to be processed;
- the vector i is multiplied by ⁇ Q , ⁇ K and ⁇ V respectively to obtain a matrix group corresponding to the vector i to be processed including the query matrix Q i , the key matrix K i , and the value matrix V i ; cyclically execute the above steps until it is determined that The matrix groups corresponding to all the vectors to be processed;
- n is a positive integer.
- the method of calculating the self-attention vector corresponding to each vector to be processed according to the matrix group corresponding to each vector to be processed may be:
- the method of calculating the feature vector corresponding to the self-attention vector according to the feedforward neural network may be:
- the feature extraction is performed on the self-attention vector according to the following feedforward neural network, and the feature vector corresponding to the self-attention vector is obtained:
- FFN(Z) max(0, ZW 1 +b 1 )W 2 +b 2 ;
- W1, b1, W2 and b2 are adjustable parameters, which can be expressed as constants.
- the feedforward neural network includes an activation function layer and a rectifying unit function layer, which are used to normalize the elements in the vector to prevent the gradient from disappearing.
- the activation function layer may include a sigmoid function or a tanh function;
- the rectifying unit function layer may include a relu function.
- the method of calculating the joint semantic vector corresponding to the feature vector based on the self-attention algorithm and the feedforward neural network may be: processing the above-mentioned feature vector through multiple encoding and decoding, so as to calculate the joint semantic vector corresponding to the feature vector; wherein , the encoding and decoding process can rely on self-attention algorithms as well as feed-forward neural networks.
- the joint semantic vector generation architecture may include an encoder, a decoder and a fully connected layer 406; wherein, the encoder may include a feedforward neural network 401 and a self-attention layer 402, and the decoder may include a feedforward neural network network 403 , encode-decode attention layer 404 and self-attention layer 405 .
- the encoders and decoders in FIG. 4 are only schematically shown, and in practical application, the number of encoders and decoders is not limited.
- the self-attention layer 402 can calculate a matrix group corresponding to each keyword for the vector corresponding to each keyword in the vector set, and calculate the corresponding vectors of each to-be-processed vector according to the matrix group.
- the self-attention vector is used as the input of the feedforward neural network 401; further, the feedforward neural network 401 can perform feature extraction on the self-attention vector to obtain a feature vector corresponding to the self-attention vector.
- N encoders are included in the BERT model, and N is a positive integer greater than or equal to 2, feature encoding is performed on the feature vector through the N encoders in sequence.
- the feature encoding result is input to the decoder, so that the self-attention layer 405 , the encoding-decoding attention layer 404 , and the self-attention layer 405 in the decoder decode the encoding result and input it to the fully connected layer 406 .
- the fully connected layer 406 may output a joint semantic vector representing the output probability of each keyword.
- the attribute structure and session information can be combined with the multi-type vectors of each keyword, which is conducive to accurate calculation of the matching degree and avoids the occurrence of mismatch between the output session information and the attribute structure of the target object Case.
- generating the joint structure vector according to the attribute structure and the session structure includes: expressing the attribute structure as a string sequence; and calculating the joint structure vector according to the string sequence and the semantic relationship.
- strings in the string sequence are respectively used to represent the node positions of the keywords in the attribute structure.
- the property structure looks like this:
- the string sequence corresponding to the attribute structure including keywords ⁇ [KV], gender, female, location, city A, constellation, Aquarius ⁇ can be ⁇ 0 1 2 1 4 1 6 ⁇
- [KV] can correspond to the parent Node 0
- [KV] is used to represent the target object
- gender, location, and constellation can correspond to child node 1
- female, city A, and Aquarius can correspond to child nodes 2, 4, and 6 in turn.
- a vector required for matching degree determination can be calculated in combination with attribute keywords and session keywords, thereby helping to improve the accuracy of matching degree determination.
- calculating the joint structure vector according to the string sequence and the semantic relationship includes: calculating the first reference vector corresponding to the attribute structure based on the recurrent neural network and the string sequence; calculating the session based on the recurrent neural network and the semantic relationship The second reference vector corresponding to the structure; the joint structure vector is obtained by splicing the first reference vector and the second reference vector.
- the above-mentioned recurrent neural network can be a Tree LSTM.
- Each unit in the TreeLSTM includes an input gate, an output gate, a cell state and a hidden layer output.
- the update of the gate vector and cell state in the TreeLSTM unit depends on all related The state of the subunit.
- TreeLSTM has multiple forget gates, corresponding to each sub-unit of the current unit, and TreeLSTM can selectively obtain information from sub-nodes.
- the first reference vector is a vectorized representation of the attribute structure
- the second reference vector is a vectorized representation of the session structure.
- the method of calculating the first reference vector corresponding to the attribute structure based on the recurrent neural network and the string sequence may be: determining the attribute keyword vector corresponding to the attribute keywords constituting the attribute structure; The word vector is input to the recurrent neural network, so that the recurrent neural network constructs a keyword node tree according to the string sequence, and calculates a first reference vector corresponding to the attribute structure according to the keyword node tree.
- each node in the keyword node tree can be calculated according to the following expression:
- C j is the child node set of node j in the keyword node tree
- the first reference vector can be calculated according to h j
- the method of calculating the second reference vector corresponding to the session structure based on the recurrent neural network and the semantic relationship may be: determining the session keyword vector corresponding to the session keywords constituting the session structure; inputting the session keyword vector including the semantic relationship.
- a recurrent neural network so that the recurrent neural network constructs a keyword node tree according to the semantic relationship, and calculates a second reference vector corresponding to the session structure according to the keyword node tree.
- the vectorized representation of the session structure and the attribute structure can be realized, so that the structure vector can be used as a condition for calculating the matching degree, and the accuracy of calculating the matching degree can be improved.
- step S350 the matching degree between the target object and the session information is determined according to the joint semantic vector and the joint structure vector, and the session information is output if the matching degree satisfies a preset condition.
- the preset condition may be expressed as a numerical range, or may be expressed as a limit on a label/category, which is not limited in this embodiment of the present application.
- the preset condition may be: the matching degree is greater than 80%; or, the matching degree is expressed as text information "consistent".
- the matching degree does not meet the preset condition, the following steps may also be performed: discarding the session information.
- output session information if the matching degree satisfies the preset condition specifically including: if the matching degree satisfies the preset condition, adding the session information to the qualified information list;
- the target session information can be selected from the qualified information list and output.
- the qualified information list may include one or more session information, and the matching degrees corresponding to the session information in the qualified information list all satisfy preset conditions.
- the method of selecting the target session information from the qualified information list may be: selecting the top N session information with the highest matching degree from the qualified information list as the target session information. ; where N is a positive integer. If the matching degree is represented by tags/categories, the method of selecting the target session information from the qualified information list may be: randomly selecting at least one session message from the qualified information list as the target session information.
- determining the matching degree between the target object and the session information according to the joint semantic vector and the joint structure vector includes: splicing the joint semantic vector and the joint structure vector to obtain a comprehensive vector; performing dimension reduction processing on the comprehensive vector, And calculate the probability that the dimensionality reduction processing result belongs to each preset category; the matching degree is represented by the preset category corresponding to the highest probability.
- the preset category can be: consistent, contradictory or irrelevant.
- consistent refers to the fact that the session message contains attribute information and does not contain a fact that contradicts the attribute information.
- Contradiction refers to the fact that the session message contains attribute information and contains contradictory attribute information. Irrelevant, which means that attribute information is not included in the session message.
- the method of splicing the joint semantic vector and the joint structure vector to obtain the integrated vector may be: splicing the N1-dimensional joint semantic vector and the N2-dimensional joint structure vector to obtain a (N1+N2)-dimensional integrated vector.
- N1 and N2 are positive integers, and N1 and N2 may be the same or different.
- the method of performing dimensionality reduction processing on the comprehensive vector and calculating the probability that the result of the dimensionality reduction processing belongs to each preset category may be: mapping the comprehensive vector into a low-dimensional comprehensive vector through a fully connected layer, wherein the low-dimensional comprehensive vector The number of elements (eg, 3) in the vector is consistent with the number of preset categories (eg, 3) to achieve dimensionality reduction for the comprehensive vector; each element in the low-dimensional comprehensive vector is converted into a probability representation through the softmax logistic regression model, and the probability representation Corresponding one-to-one with the preset category, each probability representation is used to represent the possibility that the conversation message belongs to the preset category; wherein, the number of neurons (eg, 4096) included in the fully connected layer is not limited in the embodiment of the present application.
- the matching degree between the target object and the session information can be determined in combination with semantics and information structure, the accuracy of calculating the matching degree can be improved, and to a certain extent, the inconsistency between the target object and the output session information can be avoided.
- the problem can make the target object perceived by the user more vivid.
- FIG. 5 schematically shows a schematic frame diagram of a session information processing system according to an embodiment of the present application.
- the session information processing system may include: a semantic extraction network 510 , a semantic extraction network 520 and a fully connected layer. Specifically, when it is detected that the newly added chat message "Little H, I'm coming to City A" is displayed in the conversation window, determine the conversation information to be output corresponding to the newly added chat message "I'm glad you can come to City A" .
- the attribute structure ⁇ gender female, age: 16, location: city A, constellation: Aries ⁇ corresponding to the target object small H can be read, and the session information can be structured to obtain the session structure ⁇ I, very, happy , you, can, come, come to, city A ⁇ .
- a keyword set consisting of keywords in the attribute structure and keywords in the conversation structure can be determined according to the preset vector relationship ⁇ gender, female, location, Beijing, constellation, Leo, [SEP], I, very, happy , you, can, come, come, A city ⁇ ) the keyword vector of each keyword; according to the position information of each keyword ⁇ 1,2,3,4,5,6,7,8,...,14 ,15 ⁇ Determine the position vector corresponding to each keyword; determine the category vector ⁇ aa, bb, cc, dd, ..., dd ⁇ of each keyword according to the category corresponding to each keyword; determine each keyword according to the attribute structure and session structure. Separation vector of keywords ⁇ 0,0,0,0,0,0,1,1,...,1,1 ⁇ .
- the keyword vector, the position vector, the category vector and the separation vector corresponding to the same keyword can be added to each other to obtain the to-be-processed vector corresponding to each keyword, and the set of the to-be-processed vector corresponding to each keyword.
- the vector set includes: vector to be processed 1, vector to be processed 2, vector to be processed 3, vector to be processed 4, vector to be processed 5, vector to be processed 6, vector to be processed 7, vector to be processed 8,... ..., vector 14 to be processed, and vector 15 to be processed.
- the set of vectors can be input into the semantic extraction network 510, so that the semantic extraction network 510 calculates a joint semantic vector based on a self-attention algorithm and a feed-forward neural network.
- the attribute structure containing attribute keywords ⁇ [KV], gender, female, location, city A, constellation, Aquarius ⁇ and the attribute structure containing session keywords ⁇ [SEP], I am very, happy, you can , come, come, A city ⁇ and the conversation structure of the semantic relation are input into the semantic extraction network 520 in the form of the to-be-processed vector 16 and the to-be-processed vector 17, respectively, so that the semantic extraction network 520 calculates and is used for the attribute structure and the conversation structure.
- the joint structure vector represented by the vector; wherein, the vector to be processed 16 may include attribute keyword vector, and the to-be-processed vector 17 may include session keyword vector.
- the joint semantic vector and the joint structure vector can also be spliced to obtain a comprehensive vector, and the comprehensive vector is subjected to dimensionality reduction processing through the fully connected layer 530, and the probability [X% (consistent) that the result of the dimensionality reduction processing belongs to each preset category is calculated. , Y% (contradiction), Z% (irrelevant)]; wherein, the preset category corresponding to the highest probability can be used to represent the matching degree between the target object and the session information.
- FIG. 6 schematically shows a flowchart of a method for processing session information according to an embodiment of the present application.
- the session information processing method may include steps S600 to S680, wherein:
- Step S600 when it is detected that the newly added chat message is displayed in the conversation window, determine the conversation information to be output corresponding to the newly added chat message; wherein, the conversation window includes a target object and at least one conversation object, and the newly added chat message and conversation corresponding to the object.
- Step S610 Read the attribute structure corresponding to the target object, and determine the word vector corresponding to the conversation information according to the word vector model, perform semantic dependency analysis on the conversation information according to the word vector, and obtain the semantic relationship between the language units in the conversation information, and then Session information is represented as a session structure according to semantic relations.
- Step S620 Determine the keyword vector of each keyword in the keyword set according to the preset vector relationship, determine the position vector corresponding to each keyword according to the position information of each keyword, and determine the corresponding category of each keyword according to the category corresponding to each keyword.
- Category vector, the separation vector of each keyword is determined according to the attribute structure and session structure.
- Step S630 Add the keyword vector, the position vector, the category vector and the separation vector corresponding to the same keyword in parallel to obtain the to-be-processed vectors corresponding to each keyword respectively, and add the to-be-processed vectors corresponding to each keyword respectively.
- the set of is determined to be a vector set.
- Step S640 Generate a matrix group corresponding to each vector to be processed according to the vector set, the matrix group includes a query matrix, a key matrix, and a value matrix, and calculate the self-attention corresponding to each vector to be processed according to the matrix group corresponding to each vector to be processed. force vector.
- Step S650 Calculate the feature vector corresponding to the self-attention vector according to the feedforward neural network, and calculate the joint semantic vector corresponding to the feature vector based on the self-attention algorithm and the feedforward neural network.
- Step S660 Represent the attribute structure as a string sequence, and calculate the first reference vector corresponding to the attribute structure based on the recurrent neural network and the string sequence, and calculate the second reference vector corresponding to the session structure based on the recurrent neural network and the semantic relationship, and then The joint structure vector is obtained by concatenating the first reference vector and the second reference vector.
- Step S670 splicing the joint semantic vector and the joint structure vector to obtain a comprehensive vector, and then performing dimensionality reduction processing on the comprehensive vector, and calculating the probability that the result of the dimensionality reduction processing belongs to each preset category.
- Step S680 The matching degree is indicated by the preset category corresponding to the highest probability, and if the matching degree satisfies the preset condition, the session information is output.
- steps S600 to S680 correspond to the steps shown in FIG. 3 and their embodiments.
- steps S600 to S680 please refer to the steps and their embodiments shown in FIG. 3 . It is not repeated here.
- implementing the method shown in FIG. 6 can improve the accuracy of the matching degree between the calculated session information and the target object, so that the session information can be output under the condition that the matching degree satisfies the preset conditions, so that the target object has all
- the reply's session information matches its attribute structure.
- the consistency between the target object and the output session information can also be improved, and the authenticity of the target object can be improved.
- the display interface in the user may include a target object display area 710, an interaction area 720, and a session display area 730; wherein, the interaction area 720 may include: interaction function 721, interaction function 722, . . . Interactive function 723, the interactive function may include a song request function, a gift sending function, an evaluation function, and the like.
- the session display area 730 may be located on the right or left side of the target object display area 710, so that the user can also see the session content synchronously when viewing the dynamic display of the target object.
- Figures 7 to 9 show the multi-person conversation content of user 1, user 2 and target object H, as follows:
- Little H can make corresponding replies according to the chat messages of different users, keeping the content of the reply matched with the attributes of Little H.
- the display interface in the user may include a target object display area 810, an interaction area 820, and a session display area 830; wherein, the interaction area 820 may include: interaction function 821, interaction function 822, . . . Interaction function 823, the interaction function may include a song request function, a gift sending function, an evaluation function, and the like.
- the session display area 830 may be located on the upper side or the lower side of the target object display area 810, so that the user can also see the session content synchronously when viewing the dynamic display of the target object.
- FIG. 9 schematically shows a schematic diagram of a user interface for displaying session information according to an embodiment of the present application.
- the display interface in the user may include a target object display area 910, an interaction area 920, and a session display window 930; wherein, the interaction area 920 may include: an interaction function 921, an interaction function 922, . . .
- the interactive function 923, the interactive function may include a song request function, a gift sending function, an evaluation function, and the like.
- the session display window 930 can cover part of the target object display area 910 in the form of a pop-up window.
- the session display window 930 can be triggered to pop up by operations such as gestures, clicks, voice wake-up, etc., and The session display window 930 can be closed by closing the window control when it is not necessary to view the session display window 930, thereby realizing full-screen viewing of the target object.
- FIG. 7 to FIG. 9 respectively provide three different display manners.
- steps in the flowcharts of FIGS. 3 and 6 are sequentially displayed in accordance with the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, the execution of these steps is not strictly limited to the order, and these steps may be performed in other orders. Moreover, at least a part of the steps in FIGS. 3 and 6 may include multiple steps or multiple stages. These steps or stages are not necessarily executed and completed at the same time, but may be executed at different times. The execution of these steps or stages The order is also not necessarily sequential, but may be performed alternately or alternately with other steps or at least a portion of the steps or phases within the other steps.
- the session information processing apparatus 1000 may include: a data acquisition unit 1001, a vector calculation unit 1002, and a condition determination unit 1003, wherein:
- the data acquisition unit 1001 is configured to acquire the attribute structure and the session structure corresponding to the target object when the session information to be output is detected; wherein, the session structure is used to represent the session information;
- Vector calculation unit 1002 for calculating the vector set corresponding to the keyword set; wherein, the keyword set includes keywords in the attribute structure and keywords in the conversation structure;
- the vector computing unit 1002 is also used to generate a joint semantic vector according to the vector set, and generate a joint structure vector according to the attribute structure and the session structure;
- the condition determining unit 1003 is configured to determine the matching degree between the target object and the session information according to the joint semantic vector and the joint structure vector, and output the session information if the matching degree satisfies a preset condition.
- implementing the device shown in FIG. 10 can improve the accuracy of the matching degree between the calculated session information and the target object, so that the session information can be output under the condition that the matching degree satisfies the preset conditions, so that the target object has all
- the reply's session information matches its attribute structure.
- the consistency between the target object and the output session information can also be improved, and the authenticity of the target object can be improved.
- the above-mentioned apparatus further includes: an information determination unit (not shown), wherein:
- an information determination unit configured to determine the session information corresponding to the newly added chat message before the data acquisition unit 1001 acquires the attribute structure and the conversation structure corresponding to the target object, and when it is detected that the newly added chat message is displayed in the conversation window;
- the conversation window includes a target object and at least one conversation object, and the newly added chat message corresponds to the conversation object.
- the information determining unit determines the session information corresponding to the newly added chat message, including:
- the corresponding session information can be determined according to the content of the newly added chat message input by the user, which can improve the user's chat experience and avoid unanswered questions to a certain extent.
- the data acquisition unit 1001 acquires the attribute structure and session structure corresponding to the target object, including:
- the session information is structured to obtain the session structure.
- the attribute structure and session structure of the target object can be determined, which is beneficial to determine the matching degree between the target object and the session information according to the attribute structure and the session structure, so as to avoid the output session information and the target object. Objects do not match.
- the data acquisition unit 1001 performs structured processing on session information to obtain a session structure, including:
- Session information is represented as a session structure according to semantic relations.
- the structured processing of the session information can be realized through the semantic analysis of the session information, which is beneficial to determine the matching degree between the target object and the session information.
- the data acquisition unit 1001 performs structured processing on session information to obtain a session structure, including:
- the session structure corresponding to the session information can be determined according to the preset information, so that the matching degree can be jointly determined based on the session structure and the attribute structure, and the accuracy of calculating the matching degree can be improved.
- the vector calculation unit 1002 calculates the vector set corresponding to the keyword set, including:
- the keyword vector, the position vector, the category vector and the separation vector corresponding to the same keyword are added together to obtain the to-be-processed vector corresponding to each keyword;
- a set of to-be-processed vectors corresponding to each keyword is determined as a vector set.
- vector fusion can be performed to obtain the to-be-processed vector corresponding to each keyword, so as to calculate the matching between the target object and the session information according to the to-be-processed vector set to improve the calculation accuracy of the matching degree.
- the above-mentioned device further includes: a keyword splicing unit (not shown) and a position labeling unit (not shown), wherein:
- the keyword splicing unit is used for splicing the keywords in the attribute structure and the keywords in the conversation structure before the vector calculation unit 1002 determines the position vector corresponding to each keyword according to the position information of each keyword;
- the location labeling unit is used for labeling the location information of each keyword in the keyword set in turn according to the keyword arrangement order of the splicing result, so as to obtain the location information of each keyword.
- the keyword can be marked with position information, which is beneficial to calculate the position vector of the keyword, so as to improve the accuracy of the subsequent calculation of the matching degree.
- the vector calculation unit 1002 generates a joint semantic vector according to the vector set, including:
- the matrix group includes a query matrix, a key matrix, and a value matrix
- the joint semantic vector corresponding to the feature vector is calculated based on the self-attention algorithm and the feedforward neural network.
- the attribute structure and session information can be combined with the multi-type vectors of each keyword, which is conducive to accurate calculation of the matching degree and avoids the occurrence of mismatch between the output session information and the attribute structure of the target object Case.
- the vector calculation unit 1002 generates a joint structure vector according to the attribute structure and the session structure, including:
- a vector required for matching degree determination can be calculated in combination with attribute keywords and session keywords, thereby helping to improve the accuracy of matching degree determination.
- the vector calculation unit 1002 calculates the joint structure vector according to the string sequence and the semantic relationship, including:
- the joint structure vector is obtained by concatenating the first reference vector and the second reference vector.
- the vectorized representation of the session structure and the attribute structure can be realized, so that the structure vector is used as a condition for calculating the matching degree, and the accuracy of calculating the matching degree is improved.
- condition determining unit 1003 determines the matching degree between the target object and the session information according to the joint semantic vector and the joint structure vector, including:
- the matching degree is represented by the preset category corresponding to the highest probability.
- the matching degree between the target object and the session information can be determined in combination with semantics and information structure, the accuracy of calculating the matching degree can be improved, and to a certain extent, the inconsistency between the target object and the output session information can be avoided.
- the problem can make the target object perceived by the user more vivid.
- each functional module of the session information processing apparatus of the exemplary embodiment of the present application corresponds to the steps of the above-mentioned exemplary embodiment of the session information processing method, for details not disclosed in the embodiment of the apparatus of the present application, please refer to the above-mentioned session of the present application.
- Embodiments of information processing methods are described in detail below.
- the present application also provides a computer-readable medium.
- the computer-readable medium may be included in the electronic device described in the above embodiments; it may also exist alone without being assembled into the electronic device. middle.
- the above-mentioned computer-readable medium carries one or more programs, and when the above-mentioned one or more programs are executed by an electronic device, enables the electronic device to implement the methods described in the above-mentioned embodiments.
- the computer-readable medium shown in this application may be a computer-readable signal medium or a computer-readable storage medium, or any combination of the above two.
- the computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus or device, or a combination of any of the above. More specific examples of computer readable storage media may include, but are not limited to, electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read only memory (ROM), erasable Programmable read only memory (EPROM or flash memory), fiber optics, portable compact disk read only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.
- a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.
- a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code therein. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing.
- a computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device .
- Program code embodied on a computer readable medium may be transmitted using any suitable medium including, but not limited to, wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.
- each block in the flowchart or block diagrams may represent a module, segment, or portion of code that contains one or more logical functions for implementing the specified functions executable instructions.
- the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved.
- the units involved in the embodiments of the present application may be implemented in software or hardware, and the described units may also be provided in a processor. Among them, the names of these units do not constitute a limitation on the unit itself under certain circumstances.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Databases & Information Systems (AREA)
- Human Computer Interaction (AREA)
- Information Transfer Between Computers (AREA)
Abstract
一种会话信息处理方法、会话信息处理装置、计算机可读存储介质及电子设备;涉及人工智能技术领域;包括:当检测到待输出的会话信息时,获取目标对象对应的属性结构和会话结构(S310);计算关键词集合对应的向量集合(S320);根据向量集合生成联合语义向量(S330),并根据属性结构和会话结构生成联合结构向量(S340);根据联合语义向量和联合结构向量确定目标对象与会话信息的匹配度,若匹配度满足预设条件则输出会话信息(S350)。
Description
本申请要求于2020年09月14日提交中国专利局,申请号为2020109598124,发明名称为“会话信息处理方法、装置、计算机可读存储介质及设备”的中国专利申请的优先权,其全部内容通过引用结合在本申请中。
本申请涉及人工智能技术领域,具体而言,涉及一种会话信息处理方法、会话信息处理装置、计算机可读存储介质及电子设备。
随着计算机技术的不断发展,新兴的虚拟偶像得到越来越多人的喜爱。用户可以通过终端设备观看虚拟偶像唱歌跳舞,还可以与虚拟偶像聊天,实际上,虚拟偶像作为技术产物,其展示方式及交互方式均由程序控制,用户所感知的虚拟偶像对自已聊天信息的回复,实为计算机根据聊天信息生成的回复结果。一般来说,不同的虚拟偶像具有与之相对应的属性设定,用户可以根据偏好选择虚拟偶像进行观赏、聊天,为了打造更为真实的虚拟偶像,如何使虚拟对象所回复的信息与其属性设定相匹配成为了当前亟需解决的问题。
发明内容
根据本申请的各种实施例,提供了一种会话信息处理方法、会话信息处理装置、计算机可读存储介质及电子设备。
一种会话信息处理方法,由电子设备执行,所述方法包括:
当检测到待输出的会话信息时,获取目标对象对应的属性结构和会话结构;其中,会话结构用于表征会话信息;
计算关键词集合对应的向量集合;其中,关键词集合包括属性结构中的关键词和会话结构中的关键词;
根据向量集合生成联合语义向量,并根据属性结构和会话结构生成联合结构向量;
根据联合语义向量和联合结构向量确定目标对象与会话信息的匹配度,若匹配度满足预设条件则输出会话信息。
一种会话信息处理装置,包括:数据获取单元、向量计算单元以及条件判定单元,其中:
数据获取单元,用于当检测到待输出的会话信息时,获取目标对象对应的属性结构和会话结构;其中,会话结构用于表征会话信息;
向量计算单元,用于计算关键词集合对应的向量集合;其中,关键词集合包括属性结构中的关键词和会话结构中的关键词;
向量计算单元,还用于根据向量集合生成联合语义向量,并根据属性结构和会话结构生成联合结构向量;
条件判定单元,用于根据联合语义向量和联合结构向量确定目标对象与会话信息的匹配度,若匹配度满足预设条件则输出会话信息。
一种电子设备,包括:处理器;以及存储器,用于存储所述处理器的可执行指令;其中,所述处理器配置为经由执行所述可执行指令来执行上述任意一项所述的方法。
一种计算机可读存储介质,其上存储有计算机程序,所述计算机程序被处理器执行时实现上述任意一项所述的方法。
一种计算机程序产品或计算机程序,该计算机程序产品或计算机程序包括计算机指令,该计算机指令存储在计算机可读存储介质中。计算机设备的处理器从计算机可读存储介质读取该计算机指令,处理器执行该计算机指令,使得该计算机设备执行上述的各种可选实现方式中提供的方法。
应当理解的是,以上的一般描述和后文的细节描述仅是示例性和解释性的,并不能限制本申请。
此处的附图被并入说明书中并构成本说明书的一部分,示出了符合本申请的实施例,并与说明书一起用于解释本申请的原理。显而易见地,下面描述中的附图仅仅是本申请的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动的前提下,还可以根据这些附图获得其他的附图。
图1示出了可以应用本申请实施例的一种会话信息处理方法及会话信息处理装置的示例性***架构的示意图。
图2示出了适于用来实现本申请实施例的电子设备的计算机***的结构示意图。
图3示意性示出了根据本申请的一个实施例的会话信息处理方法的流程图。
图4示意性示出了根据本申请的一个实施例的联合语义向量生成架构示意图。
图5示意性示出了根据本申请的一个实施例的会话信息处理***的框架示意图。
图6示意性示出了根据本申请的一个实施例的会话信息处理方法的流程图。
图7示意性示出了根据本申请的一个实施例的用于展示会话信息的用户界面示意图。
图8示意性示出了根据本申请的一个实施例的用于展示会话信息的用户界面示意图。
图9示意性示出了根据本申请的一个实施例的用于展示会话信息的用户界面示意图。
图10示意性示出了根据本申请的一个实施例中的会话信息处理装置的结构框图。
现在将参考附图更全面地描述示例实施方式。然而,示例实施方式能够以多种形式实施,且不应被理解为限于在此阐述的范例;相反,提供这些实施方式使得本申请将更加全面和完整,并将示例实施方式的构思全面地传达给本领域的技术人员。所描述的特征、结构或特性可以以任何合适的方式结合在一个或更多实施方式中。在下面的描述中,提供许多具体细节从而给出对本申请的实施方式的充分理解。然而,本领域技术人员将意识到,可以实践本申请的技术方案而省略所述特定细节中的一个或更多,或者可以采用其它的方法、组元、装置、步骤等。在其它情况下,不详细示出或描述公知技术方案以避免喧宾夺主而使得本申请的各方面变得模糊。
此外,附图仅为本申请的示意性图解,并非一定是按比例绘制。图中相同的附图标记表示相同或类似的部分,因而将省略对它们的重复描述。附图中所示的一些方框图是功能实体,不一定必须与物理或逻辑上独立的实体相对应。可以采用软件形式来实现这些功能实体,或在一个或多个硬件模块或集成电路中实现这些功能实体,或在不同网络和/或处理器装置和/或微控制器装置中实现这些功能实体。
图1示出了可以应用本申请实施例的一种会话信息处理方法及会话信息处理装置的示例性应用环境的***架构的示意图。
如图1所示,***架构100可以包括终端设备101、102、103中的一个或多个,网络104和服务器105。网络104用以在终端设备101、102、103和服务器105之间提供通信链路的介质。网络104可以包括各种连接类型,例如有线、无线通信链路或者光纤电缆等等。终端设备101、102、103可以是具有显示屏的各种电子设备,包括但不限于台式计算机、便携式计算机、智能手机和平板电脑等等。应该理解,图1中的终端设备、网络和服务器的数目仅仅是示意性的。根据实现需要,可以具有任意数目的终端设备、网络和服务器。比如服务器105可以是多个服务器组成的服务器集群等。另外,服务器105可以是独立的物理服务器,也可以是多个物理服务器构成的服务器集群或者分布式***,还可以是提供云服务、云数据库、云计算、云函数、云存储、网络服务、云通信、中间件服务、域名服务、安全服务、CDN、以及大数据和人工智能平台等基础云计算服务的云服务器。终端可以是智能手机、平板电脑、笔记本电脑、台式计算机、智能音箱、智能手表等,但并不局限于此。终端以及服务器可以通过有线或无线通信方式进行直接或间接地连接,本申请在此不做限制。
需要说明的是,上述的云服务器可以利用云技术提供基础云计算服务。云技术(Cloud technology)是指在广域网或局域网内将硬件、软件、网络等系列资源统一起来,实现数据的计算、储存、处理和共享的一种托管技术。
另外,上述的基础云计算服务中包括的云计算(cloud computing)指IT基础设施的交付和使用模式,指通过网络以按需、易扩展的方式获得所需资源;广义云计算指服务的交付和使用模式,指通过网络以按需、易扩展的方式获得所需服务。这种服务可以是IT和软件、互联网相关,也可是其他服务。云计算是网格计算(Grid Computing)、分布式计算(DistributedComputing)、并行计算(Parallel Computing)、效用计算(Utility Computing)、网络存储(Network Storage Technologies)、虚拟化(Virtualization)、负载均衡(Load Balance)等传统计算机和网络技术发展融合的产物。
以及,基础云计算服务中包括的云存储(cloud storage)是在云计算概念上延伸和发展出来的一个新的概念,分布式云存储***(以下简称存储***)是指通过集群应用、网格技术以及分布存储文件***等功能,将网络中大量各种不同类型的存储设备(存储设备也称之为存储节点)通过应用软件或应用接口集合起来协同工作,共同对外提供数据存储和业务访问功能的一个存储***。目前,存储***的存储方法为:创建逻辑卷,在创建逻辑卷时,就为每个逻辑卷分配物理存储空间,该物理存储空间可能是某个存储设备或者某几个存储设备的磁盘组成。客户端在某一逻辑卷上存储数据,也就是将数据存储在文件***上,文件***将数据分成许多部分,每一部分是一个对象,对象不仅包含数据而且还包含数据标识(ID,ID entity)等额外的信息,文件***将每个对象分别写入该逻辑卷的物理存储空间,且文件***会记录每个对象的存储位置信 息,从而当客户端请求访问数据时,文件***能够根据每个对象的存储位置信息让客户端对数据进行访问。存储***为逻辑卷分配物理存储空间的过程,具体为:按照对存储于逻辑卷的对象的容量估量(该估量往往相对于实际要存储的对象的容量有很大余量)和独立冗余磁盘阵列(RAID,Redundant Array of Independent Disk)的组别,预先将物理存储空间划分成分条,一个逻辑卷可以理解为一个分条,从而为逻辑卷分配了物理存储空间。
以及,基础云计算服务中包括的数据库(Database)可视为电子化的文件柜——存储电子文件的处所,用户可以对文件中的数据进行新增、查询、更新、删除等操作。所谓“数据库”是以一定方式储存在一起、能与多个用户共享、具有尽可能小的冗余度、与应用程序彼此独立的数据集合。
以及,基础云计算服务中包括的大数据(Big data)是指无法在一定时间范围内用常规软件工具进行捕捉、管理和处理的数据集合,是需要新处理模式才能具有更强的决策力、洞察发现力和流程优化能力的海量、高增长率和多样化的信息资产。随着云时代的来临,大数据也吸引了越来越多的关注,大数据需要特殊的技术,以有效地处理大量的容忍经过时间内的数据。适用于大数据的技术,包括大规模并行处理数据库、数据挖掘、分布式文件***、分布式数据库、云计算平台、互联网和可扩展的存储***。
本申请实施例所提供的会话信息处理方法一般由服务器105执行,相应地,会话信息处理装置一般设置于服务器105中。但本领域技术人员容易理解的是,本申请实施例所提供的会话信息处理方法也可以由终端设备101、102或103执行,相应的,会话信息处理装置也可以设置于终端设备101、102或103中,本示例性实施例中对此不做特殊限定。举例而言,在一种示例性实施例中,服务器105可以当检测到待输出的会话信息时,获取目标对象对应的属性结构和会话结构;其中,会话结构用于表征会话信息;计算关键词集合对应的向量集合;其中,关键词集合包括属性结构中的关键词和会话结构中的关键词;根据向量集合生成联合语义向量,并根据属性结构和会话结构生成联合结构向量;根据联合语义向量和联合结构向量确定目标对象与会话信息的匹配度,若匹配度满足预设条件则输出会话信息。
图2示出了适于用来实现本申请实施例的电子设备的计算机***的结构示意图。
需要说明的是,图2示出的电子设备的计算机***200仅是一个示例,不应对本申请实施例的功能和使用范围带来任何限制。
如图2所示,计算机***200包括中央处理单元(CPU)201,其可以根据存储在只读存储器(ROM)202中的程序或者从储存部分208加载到随机访问存储器(RAM)203中的程序而执行各种适当的动作和处理。在RAM 203中,还存储有***操作所需的各种程序和数据。CPU 201、ROM 202以及RAM 203通过总线204彼此相连。输入/输出(I/O)接口205也连接至总线204。
以下部件连接至I/O接口205:包括键盘、鼠标等的输入部分206;包括诸如阴极射线管(CRT)、液晶显示器(LCD)等以及扬声器等的输出部分207;包括硬盘等的储存部分208;以及包括诸如LAN卡、调制解调器等的网络接口卡的通信部分209。通信部分209经由诸如因特网的网络执行通信处理。驱动器210也根据需要连接至I/O接口205。可拆卸介质211,诸如磁盘、光盘、磁光盘、半导体存储器等等,根据需要 安装在驱动器210上,以便于从其上读出的计算机程序根据需要被安装入储存部分208。
特别地,根据本申请的实施例,下文参考流程图描述的过程可以被实现为计算机软件程序。例如,本申请的实施例包括一种计算机程序产品,其包括承载在计算机可读介质上的计算机程序,该计算机程序包含用于执行流程图所示的方法的程序代码。在这样的实施例中,该计算机程序可以通过通信部分209从网络上被下载和安装,和/或从可拆卸介质211被安装。在该计算机程序被中央处理单元(CPU)201执行时,执行本申请的方法和装置中限定的各种功能。
在现有技术中,用户可以通过具备聊天功能的平台与其他用户聊天。但是,随着现代生活的节奏越来越快,当用户A存在聊天需求时,可能用户B正处于忙碌状态,无法及时满足用户A的需求。并且,由于不同用户偏好的聊天对象一般不同,所偏好的用户B无法聊天时,可能用户C处于非忙碌状态但是用户C却非用户A所偏好的用户。可见,无法及时满足用户聊天需求的情况时有发生。
申请人想到,可以设计网络虚拟角色,使得用户在有聊天需求时可以随时与所偏好的网络虚拟角色进行聊天,这样可以满足用户的即时聊天需求。具体地,申请人还想到,网络虚拟角色可以对应于一种对话***,该对话***可以根据网络虚拟角色的属性信息和用户发送的聊天内容生成对应的回复信息,以实现满足用户即时聊天需求的目的。
基于现有技术可以想到,自然语言语义互动(Natural Language Interaction,NLI)管理***可以应用于上述的对话***,但是,基于NLI的时序性特点,一般在对网络虚拟角色的属性信息(如,姓名、性别、所在地等)进行建模时容易因属性间无时序关系而忽略掉其中的重要信息,从而可能造成生成的回复信息与属性信息不匹配的问题,即,回复信息与网络虚拟角色的设定不相符的问题,进而容易影响用户的沉浸式聊天体验。
举例来说,网络虚拟角色的属性信息如下:
姓名 | 小H |
性别 | 女 |
所在地 | A市 |
星座 | 水瓶座 |
年龄 | 16 |
根据NLI可以对上述属性信息进行建模,若用户的聊天内容为“我要去A市了”,依据该聊天内容生成的小H的对话信息可以包括以下可选项R1~R3:R1:很高兴你能来A市;R2:我也想有一天能去A市玩;R3:你来之后我带你去XX大学逛逛。其中,由于小H的所在地为A市,因此,生成的R2与小H的属性信息是不相符的。
可见,NLI基于时序关系的建模结果导致所生成的R2与上述属性信息不相符,若针对用户的聊天内容输出R2,则容易造成回复内容与属性信息不匹配的问题,影响用户的沉浸式聊天体验。
基于上述内容中提出的技术问题,本示例实施方式提供了一种会话信息处理方法。该会话信息处理方法可以应用于上述服务器105,也可以应用于上述终端设备101、102、103中的一个或多个,本示例性实施例中对此不做特殊限定。参考图3所示,该 会话信息处理方法可以包括以下步骤S310至步骤S350。
步骤S310:当检测到待输出的会话信息时,获取目标对象对应的属性结构和会话结构;其中,会话结构用于表征会话信息。
步骤S320:计算关键词集合对应的向量集合;其中,关键词集合包括属性结构中的关键词和会话结构中的关键词。
步骤S330:根据向量集合生成联合语义向量。
步骤S340:根据属性结构和会话结构生成联合结构向量。
步骤S350:根据联合语义向量和联合结构向量确定目标对象与会话信息的匹配度,若匹配度满足预设条件则输出会话信息。
可以理解的是,本申请可以应用于用户与虚拟人物之间的对话***,提升虚拟人物回复用户的信息与虚拟人物设定的匹配度,从而使得用户感知到的虚拟人物更为生动,避免出现因回复内容与虚拟人物设定不匹配而造成的违和感。
实施图3所示的方法,可以提升所计算的会话信息与目标对象之间的匹配度的准确率,从而可以在匹配度满足预设条件的情况下输出会话信息,以使得目标对象所回复的会话信息与其属性结构相匹配。此外,还可以提升目标对象与输出的会话信息的一致性,提升目标对象的真实性。
下面,对于本示例实施方式的上述步骤进行更加详细的说明。
在步骤S310中,当检测到待输出的会话信息时,获取目标对象对应的属性结构和会话结构;其中,会话结构用于表征会话信息。
具体地,目标对象可以为参与会话的真实用户,也可以为参与会话的虚拟用户(如,AI主播),本申请实施例不作限定。此外,上述的属性结构可以作为目标对象的个性化描述,属性结构可以包括如下关键词:性别、地址、星座、年龄以及职业等。举例来说,属性结构可以表示为:{性别:男,地点:S市,星座:狮子座,年龄:20,职业:研究员}或{性别:女,地点:A市,星座:狮子座}等。不仅属性结构可以通过多个属性关键词进行表示,会话结构也可以通过多个会话关键词进行表示。
其中,会话信息可以是目标对象与其他的会话对象进行沟通的对话消息,例如,假设目标对象为虚拟用户,其他的会话对象为真实用户,则会话信息即为对真实用户发出的聊天消息进行回复的对话消息。会话关键词源自会话信息。举例来说,会话结构可以表示为:{我,很,高兴,你,能,来,到,A市},其中,会话结构中还可以包括各关键词之间的语义关系。
作为一种可选的实施例,获取目标对象对应的属性结构和会话结构之前,上述方法还包括:当检测到会话窗口中展示新增聊天消息时,确定与新增聊天消息对应的会话信息;其中,会话窗口中包括目标对象和至少一个会话对象,新增聊天消息与会话对象相对应。其中,新增聊天消息可以是新增的用于与目标对象进行对话的消息,如真实用户与虚拟用户进行对话时所发出的消息。
具体地,新增聊天消息可以包括:文本信息、文本信息+图片信息或者文本信息+表情符号。新增聊天消息中可以包括会话对象对应的信息(如,@用户1)。此外,可选的,确定与新增聊天消息对应的会话信息之前,还可以包括:根据预设单位时长(如,10ms)检测会话窗口中是否存在新增聊天消息。
可见,实施该可选的实施例,能够在检测到用户输入的对应于目标对象的聊天消 息时,按照目标对象的属性对其进行回应,提供了一种聊天功能,以满足用户的即时聊天需求。
作为一种可选的实施例,确定与新增聊天消息对应的会话信息,包括:根据新增聊天消息从预设信息库中选择会话信息;或者,根据新增聊天消息生成会话信息。
具体地,预设信息库用于存储可调用的多种预设会话信息。进而,根据新增聊天消息从预设信息库中选择会话信息的方式可以为:确定新增聊天消息中的聊天关键词(如,天气),根据聊天关键词从预设信息库中选择会话信息(如,“今天是阴天”、“今天不会下雨”或者“今天不会出太阳”)。
可选的,根据新增聊天消息生成会话信息的方式可以为:确定新增聊天消息中的聊天关键词,根据聊天关键词和目标对象的属性结构生成会话信息。其中,会话信息中可以包括属性结构中的属性关键词。
可见,实施该可选的实施例,能够根据用户输入的新增聊天消息的内容确定相对应的会话信息,这样可以提升用户的聊天体验,一定程度地避免答非所问的情况。
作为一种可选的实施例,获取目标对象对应的属性结构和会话结构,包括:读取目标对象对应的属性结构;对会话信息进行结构化处理,得到会话结构。
具体地,目标对象可以对应多组属性结构,例如,目标对象小H对应的多组属性结构可以为:{性别:女,年龄:16,地点:A市,星座:白羊座}、{性别:女,年龄:16,地点:B市,星座:狮子座}以及{性别:女,年龄:16,地点:C市,星座:处女座}。
进而,读取目标对象对应的属性结构的方式可以为:确定新增聊天消息对应的目标对象设定,读取目标对象设定对应的属性结构;其中,不同的目标对象设定对应于不同的属性结构,用户可以根据偏好个性化选择目标对象设定。
进一步地,读取目标对象对应的属性结构之前,还可以执行以下步骤:通过不同的虚拟形象表征不同的目标对象设定并展示虚拟形象;当检测到作用于目标虚拟形象的交互操作时,将目标虚拟形象对应的目标对象设定确定为新增聊天消息对应的目标对象设定。
可见,实施该可选的实施例,能够确定出目标对象的属性结构和会话结构,从而有利于根据属性结构和会话结构判定目标对象与会话信息的匹配度,以避免所输出的会话信息与目标对象不匹配的情况。
作为一种可选的实施例,对会话信息进行结构化处理,得到会话结构,包括:根据词向量模型确定会话信息对应的词向量;根据词向量对会话信息进行文本结构分析,得到会话信息中各语言单位之间的语义关系;根据语义关系将会话信息表示为会话结构。
具体地,文本结构分析中可以包括语义依存分析,语义依存分析(Semantic Dependency Parsing,SDP),用于分析会话信息中各个语言单位之间的语义关系,并以依存关系树DT(dependency tree)呈现语义关系。其中,语义关系可以包括如下类型:施事关系(Agt)、当事关系(Exp)、感事关系(Aft)、领事关系(Poss)、受事关系(Pat)、客事关系(Cont)以及成事关系(Prod)等。此外,会话结构可以用于表示语言单位以及各语言单位之间的语义关系。此外,词向量模型可以为GloVe(Global Vectors for Word Representation),GloVe是一个基于全局词频统计的词表征工具,可以将一个词表达成 一个由实数组成的向量,该向量用于表征词的语义特性。
进而,根据词向量对会话信息进行文本结构分析,得到会话信息中各语言单位之间的语义关系的方式可以为:根据词向量对会话信息进行语义依存分析,得到会话信息中各语言单位之间的语义关系;其中,语言单位(如,我、很、高兴、你、能、来、到、A市)可以为会话关键词。或者,根据词向量对会话信息进行抽象语义表示(Abstract meaning represention,AMR),得到会话信息中各语言单位之间的语义关系;其中,AMR是一种语义表示方法,基本结构为“单根有向无环图”,AMR可以将实词抽象为概念作为单根有向无环图上的节点,将没有实际意义的虚词抽象为边。
可选的,上述方法还可以包括以下步骤:
根据语料库构建共现矩阵X;
构建用于表示词向量和X之间近似关系的词向量模型:
根据词向量模型确定损失函数表达式:J
f(X
ij)为非递减的权重函数,f(0)=0;
根据损失函数表达式以及词向量模型的样本输出结果调整词向量模型的模型参数,以实现对于词向量模型的预训练。其中,共现矩阵X的元素X
ij可以表示单词i和上下文单词j在预设尺寸的上下文窗口内共同出现的次数。例如,f(X
ij)可以表示为
可选的,若根据词向量对会话信息进行文本结构分析之后,未得到会话信息中各语言单位之间的语义关系,则可以执行以下步骤:返回失败提示。
可见,实施该可选的实施例,能够通过对会话信息的语义分析,实现对会话信息的结构化处理,从而有利于确定目标对象与会话信息之间的匹配度。
作为另一种可选的实施例,对会话信息进行结构化处理,得到会话结构,包括:将会话信息与预设词汇集合进行比对,以确定出会话信息中的关键词;根据预设词汇关系构建包含会话信息中的关键词的会话结构。
具体地,预设词汇集合中可以包括预存储的词汇,预设词汇关系可以为预设的一一对应关系,用于表征相邻语言单位之间的关系。
可见,实施该可选的实施例,能够根据预设信息确定出会话信息对应的会话结构,从而可以基于会话结构和属性结构共同确定匹配度,提升计算匹配度的准确率。
在步骤S320中,计算关键词集合对应的向量集合;其中,关键词集合包括属性结 构中的关键词和会话结构中的关键词。
具体地,向量集合中可以包括:关键词向量、位置向量、类别向量以及分隔向量。举例来说,若属性结构为{性别:女,地点:北京,星座:狮子座},会话结构为{我,很,高兴,你,能,来,到,A市},则关键词集合可以为:{性别,女,地点,北京,星座,狮子座,[SEP],我,很,高兴,你,能,来,到,A市};其中,[SEP]为分句符号,用于分隔关键词语料。
作为一种可选的实施例,计算关键词集合对应的向量集合,包括:根据预设向量关系确定关键词集合中各关键词的关键词向量;根据各关键词的位置信息确定各关键词对应的位置向量;根据各关键词分别对应的类别确定各关键词的类别向量;根据属性结构和会话结构确定各关键词的分隔向量;其中,属性结构中的关键词的分隔向量相同,会话结构中的关键词的分隔向量相同;将对应于同一关键词的关键词向量、位置向量、类别向量以及分隔向量进行对位相加,得到各关键词分别对应的待处理向量;将各关键词分别对应的待处理向量的集合确定为向量集合。
具体地,预设向量关系可以用于表征不同的关键词与关键词向量之间的一一对应关系,关键词向量可以表示为实数列表,关键词集合中各关键词对应的关键词向量可以对应于同一维度(如,768)。另外,上述的对位相加用于表示:关键词向量、位置向量、类别向量以及分隔向量之间相同位置的元素加和,例如,关键词“天气”对应的关键词向量(1,1,1)、位置向量(2,2,2)、类别向量(3,3,3)以及分隔向量(4,4,4)进行对位相加(1+2+3+4,1+2+3+4,1+2+3+4)之后,可以得到关键词“天气”对应的待处理向量(10,10,10)。
可选的,根据各关键词的位置信息确定各关键词对应的位置向量的方式可以为:若位置信息为偶数,则对该位置信息对应的关键词进行正弦编码;若位置信息为奇数,则对该位置信息对应的关键词进行余弦编码,从而,得到各关键词对应的位置向量。
可选的,根据各关键词分别对应的类别确定各关键词的类别向量的方式可以为:将对应于同一类别的关键词(如,“性别”和“男”)进行统一的类别向量表示,不同类别的关键词对应于不同的类别向量。
可选的,根据属性结构和会话结构确定各关键词的分隔向量的方式可以为:将属于属性结构的关键词通过相同的分隔向量进行表示,将属于会话结构的关键词通过相同的分隔向量进行表示,属性结构的关键词与会话结构的关键词对应的分隔向量不同。
可见,实施该可选的实施例,能够计算各关键词对应的不同类型向量,并进行向量融合,得到各关键词对应的待处理向量,以便根据待处理向量集合计算目标对象与会话信息的匹配度,提升匹配度的计算准确率。
作为一种可选的实施例,根据各关键词的位置信息确定各关键词对应的位置向量之前,上述方法还包括:将属性结构中的关键词和会话结构中的关键词进行拼接;根据拼接结果的关键词排列顺序依次对关键词集合中各关键词标注位置信息,得到各关键词的位置信息。
其中,拼接结果中包括属性结构的属性关键词和会话结构中的会话关键词;位置信息可以通过数值进行表示。
可选的,将属性结构中的关键词和会话结构中的关键词进行拼接的方式可以为:根据分句符号将属性结构中的关键词和会话结构中的关键词按照顺序进行拼接(如, {性别,女,地点,北京,星座,狮子座,[SEP],我,很,高兴,你,能,来,到,A市})。举例来说,按照关键词排列顺序所标注的位置信息可以表示如下:
可见,实施该可选的实施例,能够对关键词进行位置信息标注,从而有利于计算关键词的位置向量,以提升后续计算匹配度的准确率。
在步骤S330中,根据向量集合生成联合语义向量。
可选的,根据向量集合生成联合语义向量的方式可以为:将向量集合输入BERT(Bidirectional Encoder Representations from Transformers)模型,以使得BERT根据向量集合生成联合语义向量,联合语义向量用于表示每个关键词的输出概率;其中,BERT是一种语言分析模型。
在步骤S340中,根据属性结构和会话结构生成联合结构向量。
可选的,根据属性结构和会话结构生成联合结构向量的方式可以为:将属性结构和会话结构输入针对成分树(Tree)的长短期记忆网络(Long Short Term Memory networks,LSTM)模型(即,Tree LSTM),以使得Tree LSTM根据属性结构和会话结构生成联合结构向量,生成联合结构向量是属性结构和会话结构的向量化表示。
可选的,在将向量集合输入BERT模型之前,还可以执行:根据预设损失函数训练BERT模型;其中,预设损失函数可以为回归损失函数、平方误差损失函数、绝对误差损失函数、Huber损失函数、二分类损失函数、二分类交叉熵、Hinge损失函数、多分类损失函数、多分类交叉熵损失或KL散度(Kullback Leibler Divergence Loss),本申请实施例不作限定。同理,在将属性结构和会话结构输入Tree LSTM模型之前,还可以执行:根据预设损失函数训练Tree LSTM模型。
作为一种可选的实施例,根据向量集合生成联合语义向量,包括:根据向量集合生成各待处理向量分别对应的矩阵组,矩阵组包括查询矩阵、键矩阵、值矩阵;根据各待处理向量分别对应的矩阵组计算各待处理向量分别对应的自注意力向量;根据前馈神经网络计算自注意力向量对应的特征向量;基于自注意力算法以及前馈神经网络计算特征向量对应的联合语义向量。
基于前述限定,根据向量集合生成联合语义向量的步骤可以基于BERT模型实现;其中,BERT模型中采用了机器翻译(Transformer)模型的编解码模式,Transformer模型由编码组件和解码组件构成,编码组件由多个编码器构成,解码组件由多个解码器构成,编码器和解码器的数量相同;其中,每个编码器中均包括前馈神经网络和自注意力层,每个解码器中均包括前馈神经网络、编码-解码注意力层以及自注意力层。
可选的,若自注意力层中存在多头自注意力,根据向量集合生成各待处理向量分别对应的矩阵组的方式可以为:生成与各待处理向量对应的至少两个参考矩阵组;将对应于同一待处理向量的参考矩阵组进行拼接,得到各待处理向量对应的矩阵组;其中,参考矩阵组的数量与自注意力的头数(如,8)一致。
可选的,根据向量集合生成各待处理向量分别对应的矩阵组的方式可以为:确 定各待处理向量对应的查询权重矩阵ω
Q、键权重矩阵ω
K以及值权重矩阵ω
V;将待处理向量i分别与ω
Q、ω
K以及ω
V相乘,得到包括查询矩阵Q
i、键矩阵K
i、值矩阵V
i的与待处理向量i对应的矩阵组;循环执行上述步骤,直到确定出所有待处理向量分别对应的矩阵组;
其中,i∈[1、2、......、n],n为正整数。
进一步地,根据各待处理向量分别对应的矩阵组计算各待处理向量分别对应的自注意力向量的方式可以为:
依据下述表达式计算各待处理向量对应的自注意力向量Z
i;
可选的,根据前馈神经网络计算自注意力向量对应的特征向量的方式可以为:
根据如下前馈神经网络对自注意力向量进行特征提取,得到自注意力向量对应的特征向量:
FFN(Z)=max(0,ZW
1+b
1)W
2+b
2;
其中,W1、b1、W2以及b2为可调参数,可以表示为常数。
前馈神经网络包括激活函数层和整流性单元函数层,用于对向量中的元素进行归一化,防止梯度消失。具体地,激活函数层可以包括sigmoid函数或tanh函数;整流性单元函数层可以包括relu函数。
可选的,基于自注意力算法以及前馈神经网络计算特征向量对应的联合语义向量的方式可以为:通过多次编解码处理上述的特征向量,以计算得到特征向量对应的联合语义向量;其中,编解码处理可以依赖于自注意力算法以及前馈神经网络。
请参阅图4,图4示意性示出了根据本申请的一个实施例的联合语义向量生成架构示意图。如图4所示,联合语义向量生成架构中可以包括编码器、解码器和全连接层406;其中,编码器可以包括前馈神经网络401和自注意力层402,解码器可以包括前馈神经网络403、编码-解码注意力层404以及自注意力层405。需要说明的是,图4中的编码器和解码器仅为示意性示出,在实际应用过程中,对于编码器和解码器的数量不作限定。
具体地,自注意力层402在接收到输入BERT模型的向量集合后,可以针对向量集合中各关键词对应的向量计算各关键词对应的矩阵组,并根据矩阵组计算各待处理向量分别对应的自注意力向量,以作为前馈神经网络401的输入;进而,前馈神经网络401可以对自注意力向量进行特征提取,得到自注意力向量对应的特征向量。进一步地,若BERT模型中包括N个编码器,N为大于等于2的正整数,则依次通过N个编码器对特征向量进行特征编码。进而,将特征编码结果输入解码器,以使得解码器中的自注意力层405、编码-解码注意力层404以及自注意力层405对编码结果进行解码并输入全连接层406。全连接层406可以输出用于表示每个关键词输出概率的联合语义向量。
可见,实施该可选的实施例,能够结合各个关键词的多类型向量将属性结构和会话信息进行联合,从而有利于精准计算匹配度,避免出现输出的会话信息与目标对象的属性结构不匹配的情况。
作为一种可选的实施例,根据属性结构和会话结构生成联合结构向量,包括:将属性结构表示为字符串序列;根据字符串序列以及语义关系计算联合结构向量。
具体地,字符串序列中的字符串分别用于表征关键词在属性结构中的节点位置。举例来说,属性结构如下所示:
其中,包括关键词{[KV]、性别、女、地点、A市、星座、水瓶座}的属性结构对应的字符串序列可以为{0 1 2 1 4 1 6},[KV]可以对应父节点0,[KV]用于表示目标对象,性别、地点和星座可以对应子节点1,女、A市和水瓶座可以依次对应子节点2、4、6。
可见,实施该可选的实施例,能够结合属性关键词和会话关键词计算出匹配度判定所需的向量,从而有利于提升匹配度判定的准确率。
作为一种可选的实施例,根据字符串序列以及语义关系计算联合结构向量,包括:基于递归神经网络和字符串序列计算属性结构对应的第一参考向量;基于递归神经网络和语义关系计算会话结构对应的第二参考向量;将第一参考向量和第二参考向量拼接得到联合结构向量。
具体地,上述的递归神经网络可以为Tree LSTM,TreeLSTM中每个单元都包括输入门、输出门、单元状态和隐层输出,TreeLSTM单元中门向量和细胞状态的更新依赖于所有与之相关的子单元的状态。另外,TreeLSTM拥有多个遗忘门,分别对应当前单元的每个子单元,TreeLSTM可以选择性地从子节点中获取信息。另外,第一参考向量是对属性结构的向量化表示,第二参考向量是对会话结构的向量化表示。
可选的,基于递归神经网络和字符串序列计算属性结构对应的第一参考向量的方式可以为:确定构成属性结构的属性关键词所对应的属性关键词向量;将包含字符串序列的属性关键词向量输入递归神经网络,以使得递归神经网络根据字符串序列构建关键词节点树,并根据关键词节点树计算属性结构对应的第一参考向量。
具体地,可以根据下述表达式对关键词节点树中每个节点进行计算:
遗忘门f
jk=σ(W
(f)x
j+U
(f)h
k+b
(f));
进而,可以得到如下表达式:
C
j=i
j⊙u
j+∑
k∈c(j)f
jk⊙c
k;
C
j为关键词节点树中的节点j的子节点集合;
根据C
j可以计算得到隐藏状态h
j=o
j⊙tanh(C
j);
同理,基于递归神经网络和语义关系计算会话结构对应的第二参考向量的方式可以为:确定构成会话结构的会话关键词所对应的会话关键词向量;将包含语义关系的会话关键词向量输入递归神经网络,以使得递归神经网络根据语义关系构建关键词节点树,并根据关键词节点树计算会话结构对应的第二参考向量。
可见,实施该可选的实施例,能够实现对于会话结构和属性结构的向量化表示,以将结构向量作为匹配度计算的条件,提升计算匹配度的准确率。
在步骤S350中,根据联合语义向量和联合结构向量确定目标对象与会话信息的匹配度,若匹配度满足预设条件则输出会话信息。
具体地,预设条件可以表示为数值范围、也可以表示为对标签/类别的限定,本申请实施例不作限定。举例来说,预设条件可以为:匹配度大于80%;或者,匹配度表示为文本信息“一致”。此外,若匹配度不满足预设条件,还可以执行以下步骤:丢弃该会话信息。
可选的,若匹配度满足预设条件则输出会话信息,具体包括:若匹配度满足预设条件,则将该会话信息添加至合格信息列表;当检测到单位时间内生成的所有待输出的会话消息均完成匹配度计算时,可以从合格信息列表中选取目标会话信息并输出。其中,合格信息列表中可以包括一个或多个会话信息,合格信息列表中的会话信息对应的匹配度均满足于预设条件。
进一步地,若匹配度通过数值(如,80%)进行表示,从合格信息列表中选取目标会话信息的方式可以为:从合格信息列表中选取匹配度最高的前N个会话信息作为目标会话信息;其中,N为正整数。若匹配度通过标签/类别进行表示,从合格信息列表中选取目标会话信息的方式可以为:从合格信息列表中随机选取至少一个会话消息作为目标会话信息。
作为一种可选的实施例,根据联合语义向量和联合结构向量确定目标对象与会话信息的匹配度,包括:拼接联合语义向量和联合结构向量,得到综合向量;对综合向量进行降维处理,并计算降维处理结果属于各预设类别的概率;通过最高概率对应的预设类别表示匹配度。
具体地,预设类别可以为:一致、矛盾或无关。其中,一致,指的是会话消息中包含属性信息且不含有与属性信息相矛盾的事实。矛盾,指的是会话消息中包含属性信息且含有与属性信息相矛盾的事实。无关,指的是会话消息中未包含属性信息。
可选的,拼接联合语义向量和联合结构向量,得到综合向量的方式可以为:将N1维联合语义向量和N2维的联合结构向量进行拼接,得到(N1+N2)维的综合向量。其中,N1和N2均为正整数,N1和N2可以相同也可以不同。
可选的,对综合向量进行降维处理,并计算降维处理结果属于各预设类别的概率的方式可以为:通过全连接层将综合向量映射为低维综合向量,其中,低维综合向量中的元素数量(如,3)与预设类别数量(如,3)一致,以实现对于综合向量的降维; 通过softmax逻辑回归模型将低维综合向量中各元素转换为概率表示,概率表示与预设类别一一对应,各概率表示分别用于表征会话消息属于该预设类别的可能性;其中,全连接层中包含的神经元数量(如,4096),本申请实施例不作限定。
可见,实施该可选的实施例,能够结合语义和信息结构确定目标对象与会话信息之间的匹配度,可以提升计算匹配度的准确率,一定程度地避免目标对象与输出的会话信息不一致的问题,可以使得用户感知到的目标对象更为生动。
请参阅图5,图5示意性示出了根据本申请的一个实施例的会话信息处理***的框架示意图。如图5所示,该会话信息处理***可以包括:语义提取网络510、语义提取网络520以及全连接层。具体地,当检测到会话窗口中展示新增聊天消息“小H,我来A市了”时,确定与新增聊天消息对应的待输出的会话信息“我很高兴你能来到A市”。进而,可以读取目标对象小H对应的属性结构{性别:女,年龄:16,地点:A市,星座:白羊座},并对会话信息进行结构化处理,得到会话结构{我,很,高兴,你,能,来,到,A市}。
进而,可以根据预设向量关系确定由属性结构中的关键词和会话结构中的关键词构成的关键词集合{性别,女,地点,北京,星座,狮子座,[SEP],我,很,高兴,你,能,来,到,A市})中各关键词的关键词向量;根据各关键词的位置信息{1,2,3,4,5,6,7,8,……,14,15}确定各关键词对应的位置向量;根据各关键词分别对应的类别确定各关键词的类别向量{aa,bb,cc,dd,……,dd};根据属性结构和会话结构确定各关键词的分隔向量{0,0,0,0,0,0,1,1,……,1,1}。
进而,可以将对应于同一关键词的关键词向量、位置向量、类别向量以及分隔向量进行对位相加,得到各关键词分别对应的待处理向量,各关键词分别对应的待处理向量的集合确定为向量集合,向量集合包括:待处理向量1、待处理向量2、待处理向量3、待处理向量4、待处理向量5、待处理向量6、待处理向量7、待处理向量8、……、待处理向量14和待处理向量15。进而,可以将向量集合输入语义提取网络510,以使得语义提取网络510基于自注意力算法以及前馈神经网络计算联合语义向量。以及,还可以将包含属性关键词{[KV]、性别、女、地点、A市、星座、水瓶座}的属性结构和包含会话关键词{[SEP],我,很,高兴,你,能,来,到,A市}和语义关系的会话结构分别以待处理向量16和待处理向量17的形式输入语义提取网络520,以使得语义提取网络520计算得到用于对属性结构和会话结构进行向量表示的联合结构向量;其中,待处理向量16可以包括属性关键词向量,待处理向量17可以包括会话关键词向量。
进而,还可以拼接联合语义向量和联合结构向量,得到综合向量,并通过全连接层530对综合向量进行降维处理,并计算降维处理结果属于各预设类别的概率[X%(一致),Y%(矛盾),Z%(无关)];其中,最高概率对应的预设类别可以用于表示目标对象与会话信息之间的匹配度。
请参阅图6,图6示意性示出了根据本申请的一个实施例的会话信息处理方法的流程图。如图6所示,该会话信息处理方法可以包括:步骤S600~步骤S680,其中:
步骤S600:当检测到会话窗口中展示新增聊天消息时,确定与新增聊天消息对应的待输出的会话信息;其中,会话窗口中包括目标对象和至少一个会话对象,新增聊天消息与会话对象相对应。
步骤S610:读取目标对象对应的属性结构,并根据词向量模型确定会话信息对应的词向量,根据词向量对会话信息进行语义依存分析,得到会话信息中各语言单位之间的语义关系,进而根据语义关系将会话信息表示为会话结构。
步骤S620:根据预设向量关系确定关键词集合中各关键词的关键词向量、根据各关键词的位置信息确定各关键词对应的位置向量、根据各关键词分别对应的类别确定各关键词的类别向量、根据属性结构和会话结构确定各关键词的分隔向量。
步骤S630:将对应于同一关键词的关键词向量、位置向量、类别向量以及分隔向量进行对位相加,得到各关键词分别对应的待处理向量,并将各关键词分别对应的待处理向量的集合确定为向量集合。
步骤S640:根据向量集合生成各待处理向量分别对应的矩阵组,矩阵组包括查询矩阵、键矩阵、值矩阵,并根据各待处理向量分别对应的矩阵组计算各待处理向量分别对应的自注意力向量。
步骤S650:根据前馈神经网络计算自注意力向量对应的特征向量,基于自注意力算法以及前馈神经网络计算特征向量对应的联合语义向量。
步骤S660:将属性结构表示为字符串序列,并基于递归神经网络和字符串序列计算属性结构对应的第一参考向量,以及基于递归神经网络和语义关系计算会话结构对应的第二参考向量,进而将第一参考向量和第二参考向量拼接得到联合结构向量。
步骤S670:拼接联合语义向量和联合结构向量,得到综合向量,进而对综合向量进行降维处理,并计算降维处理结果属于各预设类别的概率。
步骤S680:通过最高概率对应的预设类别表示匹配度,若匹配度满足预设条件则输出会话信息。
需要说明的是,步骤S600~步骤S680与图3所示的各步骤及其实施例相对应,针对步骤S600~步骤S680的实施方式,请参阅图3所示的各步骤及其实施例,此处不再赘述。
可见,实施图6所示的方法,可以提升所计算的会话信息与目标对象之间的匹配度的准确率,从而可以在匹配度满足预设条件的情况下输出会话信息,以使得目标对象所回复的会话信息与其属性结构相匹配。此外,还可以提升目标对象与输出的会话信息的一致性,提升目标对象的真实性。
请参阅图7,图7示意性示出了根据本申请的一个实施例的用于展示会话信息的用户界面示意图。如图7所示,在用户中的展示界面中,可以包括目标对象展示区域710、交互区域720以及会话展示区域730;其中,交互区域720可以包括:交互功能721、交互功能722、……、交互功能723,交互功能可以包括点歌功能、送礼物功能以及评价功能等。会话展示区域730可以处于目标对象展示区域710的右侧或左侧,使得用户在观看目标对象动态展示时还可以同步看到会话内容。
需要说明的是,图7~图9中展示了用户1、用户2以及目标对象小H的多人会话内容,如下所示:
用户1:@小H你是谁。小H:@用户1你好,这里是你的专属歌姬小H!用户1:@小H你来自哪里?小H:@用户1我来自A市。用户1:@小H你几岁了?小H:@用户1我已经16岁了。用户2:@小H给我唱首《小芒果》吧。小H:@用户2好的!
可见,小H可以根据不同的用户的聊天消息进行相应的回复,保持所回复的内容与小H的属性相匹配。
请参阅图8,图8示意性示出了根据本申请的一个实施例的用于展示会话信息的用户界面示意图。如图8所示,在用户中的展示界面中,可以包括目标对象展示区域810、交互区域820以及会话展示区域830;其中,交互区域820可以包括:交互功能821、交互功能822、……、交互功能823,交互功能可以包括点歌功能、送礼物功能以及评价功能等。会话展示区域830可以处于目标对象展示区域810的上侧或下侧,使得用户在观看目标对象动态展示时还可以同步看到会话内容。
请参阅图9,图9示意性示出了根据本申请的一个实施例的用于展示会话信息的用户界面示意图。如图9所示,在用户中的展示界面中,可以包括目标对象展示区域910、交互区域920以及会话展示窗口930;其中,交互区域920可以包括:交互功能921、交互功能922、……、交互功能923,交互功能可以包括点歌功能、送礼物功能以及评价功能等。在图9中,会话展示窗口930可以以弹窗的形式覆盖部分目标对象展示区域910,用户可以在需要查看会话展示窗口930时通过手势、点击、语音唤醒等操作触发会话展示窗口930弹出,并且可以在不需要查看会话展示窗口930时通过关闭窗口控件关闭会话展示窗口930,进而实现对于目标对象的全屏观赏。需要说明的是,图7~图9分别提供了三种不同的展示方式。
应该理解的是,虽然图3、6的流程图中的各个步骤按照箭头的指示依次显示,但是这些步骤并不是必然按照箭头指示的顺序依次执行。除非本文中有明确的说明,这些步骤的执行并没有严格的顺序限制,这些步骤可以以其它的顺序执行。而且,图3、6中的至少一部分步骤可以包括多个步骤或者多个阶段,这些步骤或者阶段并不必然是在同一时刻执行完成,而是可以在不同的时刻执行,这些步骤或者阶段的执行顺序也不必然是依次进行,而是可以与其它步骤或者其它步骤中的步骤或者阶段的至少一部分轮流或者交替地执行。
进一步的,本示例实施方式中,还提供了一种会话信息处理装置。参考图10所示,该会话信息处理装置1000可以包括:数据获取单元1001、向量计算单元1002以及条件判定单元1003,其中:
数据获取单元1001,用于当检测到待输出的会话信息时,获取目标对象对应的属性结构和会话结构;其中,会话结构用于表征会话信息;
向量计算单元1002,用于计算关键词集合对应的向量集合;其中,关键词集合包括属性结构中的关键词和会话结构中的关键词;
向量计算单元1002,还用于根据向量集合生成联合语义向量,并根据属性结构和会话结构生成联合结构向量;
条件判定单元1003,用于根据联合语义向量和联合结构向量确定目标对象与会话信息的匹配度,若匹配度满足预设条件则输出会话信息。
可见,实施图10所示的装置,可以提升所计算的会话信息与目标对象之间的匹配度的准确率,从而可以在匹配度满足预设条件的情况下输出会话信息,以使得目标对象所回复的会话信息与其属性结构相匹配。此外,还可以提升目标对象与输出的会话信息的一致性,提升目标对象的真实性。
在本申请的一种示例性实施例中,上述装置还包括:信息确定单元(未图示),其中:
信息确定单元,用于在数据获取单元1001获取目标对象对应的属性结构和会话结构之前,以及当检测到会话窗口中展示新增聊天消息时,确定与新增聊天消息对应的会话信息;
其中,会话窗口中包括目标对象和至少一个会话对象,新增聊天消息与会话对象相对应。
可见,实施该可选的实施例,能够在检测到用户输入的对应于目标对象的聊天消息时,按照目标对象的属性对其进行回应,提供了一种聊天功能,以满足用户的即时聊天需求。
在本申请的一种示例性实施例中,信息确定单元确定与新增聊天消息对应的会话信息,包括:
根据新增聊天消息从预设信息库中选择会话信息;
或者,根据新增聊天消息生成会话信息。
可见,实施该可选的实施例,能够根据用户输入的新增聊天消息的内容确定相对应的会话信息,这样可以提升用户的聊天体验,一定程度地避免答非所问的情况。
在本申请的一种示例性实施例中,数据获取单元1001获取目标对象对应的属性结构和会话结构,包括:
读取目标对象对应的属性结构;
对会话信息进行结构化处理,得到会话结构。
可见,实施该可选的实施例,能够确定出目标对象的属性结构和会话结构,从而有利于根据属性结构和会话结构判定目标对象与会话信息的匹配度,以避免所输出的会话信息与目标对象不匹配的情况。
在本申请的一种示例性实施例中,数据获取单元1001对会话信息进行结构化处理,得到会话结构,包括:
根据词向量模型确定会话信息对应的词向量;
根据词向量对会话信息进行文本结构分析,得到会话信息中各语言单位之间的语义关系;
根据语义关系将会话信息表示为会话结构。
可见,实施该可选的实施例,能够通过对会话信息的语义分析,实现对会话信息的结构化处理,从而有利于确定目标对象与会话信息之间的匹配度。
在本申请的一种示例性实施例中,数据获取单元1001对会话信息进行结构化处理,得到会话结构,包括:
将会话信息与预设词汇集合进行比对,以确定出会话信息中的关键词;
根据预设词汇关系构建包含会话信息中的关键词的会话结构。
可见,实施该可选的实施例,能够根据预设信息确定出会话信息对应的会话结构,从而可以基于会话结构和属性结构共同确定匹配度,提升计算匹配度的准确率。
在本申请的一种示例性实施例中,向量计算单元1002计算关键词集合对应的向量集合,包括:
根据预设向量关系确定关键词集合中各关键词的关键词向量;
根据各关键词的位置信息确定各关键词对应的位置向量;
根据各关键词分别对应的类别确定各关键词的类别向量;
根据属性结构和会话结构确定各关键词的分隔向量;其中,属性结构中的关键词的分隔向量相同,会话结构中的关键词的分隔向量相同;
将对应于同一关键词的关键词向量、位置向量、类别向量以及分隔向量进行对位相加,得到各关键词分别对应的待处理向量;
将各关键词分别对应的待处理向量的集合确定为向量集合。
可见,实施该可选的实施例,能够计算各关键词对应的不同类型向量,并进行向量融合,得到各关键词对应的待处理向量,以便根据待处理向量集合计算目标对象与会话信息的匹配度,提升匹配度的计算准确率。
在本申请的一种示例性实施例中,上述装置还包括:关键词拼接单元(未图示)和位置标注单元(未图示),其中:
关键词拼接单元,用于在向量计算单元1002根据各关键词的位置信息确定各关键词对应的位置向量之前,将属性结构中的关键词和会话结构中的关键词进行拼接;
位置标注单元,用于根据拼接结果的关键词排列顺序依次对关键词集合中各关键词标注位置信息,得到各关键词的位置信息。
可见,实施该可选的实施例,能够对关键词进行位置信息标注,从而有利于计算关键词的位置向量,以提升后续计算匹配度的准确率。
在本申请的一种示例性实施例中,向量计算单元1002根据向量集合生成联合语义向量,包括:
根据向量集合生成各待处理向量分别对应的矩阵组,矩阵组包括查询矩阵、键矩阵、值矩阵;
根据各待处理向量分别对应的矩阵组计算各待处理向量分别对应的自注意力向量;
根据前馈神经网络计算自注意力向量对应的特征向量;
基于自注意力算法以及前馈神经网络计算特征向量对应的联合语义向量。
可见,实施该可选的实施例,能够结合各个关键词的多类型向量将属性结构和会话信息进行联合,从而有利于精准计算匹配度,避免出现输出的会话信息与目标对象的属性结构不匹配的情况。
在本申请的一种示例性实施例中,向量计算单元1002根据属性结构和会话结构生成联合结构向量,包括:
将属性结构表示为字符串序列;
根据字符串序列以及语义关系计算联合结构向量。
可见,实施该可选的实施例,能够结合属性关键词和会话关键词计算出匹配度判定所需的向量,从而有利于提升匹配度判定的准确率。
在本申请的一种示例性实施例中,向量计算单元1002根据字符串序列以及语义关系计算联合结构向量,包括:
基于递归神经网络和字符串序列计算属性结构对应的第一参考向量;
基于递归神经网络和语义关系计算会话结构对应的第二参考向量;
将第一参考向量和第二参考向量拼接得到联合结构向量。
可见,实施该可选的实施例,能够实现对于会话结构和属性结构的向量化表示, 以将结构向量作为匹配度计算的条件,提升计算匹配度的准确率。
在本申请的一种示例性实施例中,条件判定单元1003根据联合语义向量和联合结构向量确定目标对象与会话信息的匹配度,包括:
拼接联合语义向量和联合结构向量,得到综合向量;
对综合向量进行降维处理,并计算降维处理结果属于各预设类别的概率;
通过最高概率对应的预设类别表示匹配度。
可见,实施该可选的实施例,能够结合语义和信息结构确定目标对象与会话信息之间的匹配度,可以提升计算匹配度的准确率,一定程度地避免目标对象与输出的会话信息不一致的问题,可以使得用户感知到的目标对象更为生动。
应当注意,尽管在上文详细描述中提及了用于动作执行的设备的若干模块或者单元,但是这种划分并非强制性的。实际上,根据本申请的实施方式,上文描述的两个或更多模块或者单元的特征和功能可以在一个模块或者单元中具体化。反之,上文描述的一个模块或者单元的特征和功能可以进一步划分为由多个模块或者单元来具体化。
由于本申请的示例实施例的会话信息处理装置的各个功能模块与上述会话信息处理方法的示例实施例的步骤对应,因此对于本申请装置实施例中未披露的细节,请参照本申请上述的会话信息处理方法的实施例。
作为另一方面,本申请还提供了一种计算机可读介质,该计算机可读介质可以是上述实施例中描述的电子设备中所包含的;也可以是单独存在,而未装配入该电子设备中。上述计算机可读介质承载有一个或者多个程序,当上述一个或者多个程序被一个该电子设备执行时,使得该电子设备实现上述实施例中所述的方法。
需要说明的是,本申请所示的计算机可读介质可以是计算机可读信号介质或者计算机可读存储介质或者是上述两者的任意组合。计算机可读存储介质例如可以是——但不限于——电、磁、光、电磁、红外线、或半导体的***、装置或器件,或者任意以上的组合。计算机可读存储介质的更具体的例子可以包括但不限于:具有一个或多个导线的电连接、便携式计算机磁盘、硬盘、随机访问存储器(RAM)、只读存储器(ROM)、可擦式可编程只读存储器(EPROM或闪存)、光纤、便携式紧凑磁盘只读存储器(CD-ROM)、光存储器件、磁存储器件、或者上述的任意合适的组合。在本申请中,计算机可读存储介质可以是任何包含或存储程序的有形介质,该程序可以被指令执行***、装置或者器件使用或者与其结合使用。而在本申请中,计算机可读的信号介质可以包括在基带中或者作为载波一部分传播的数据信号,其中承载了计算机可读的程序代码。这种传播的数据信号可以采用多种形式,包括但不限于电磁信号、光信号或上述的任意合适的组合。计算机可读的信号介质还可以是计算机可读存储介质以外的任何计算机可读介质,该计算机可读介质可以发送、传播或者传输用于由指令执行***、装置或者器件使用或者与其结合使用的程序。计算机可读介质上包含的程序代码可以用任何适当的介质传输,包括但不限于:无线、电线、光缆、RF等等,或者上述的任意合适的组合。
附图中的流程图和框图,图示了按照本申请各种实施例的***、方法和计算机程序产品的可能实现的体系架构、功能和操作。在这点上,流程图或框图中的每个方框可以代表一个模块、程序段、或代码的一部分,上述模块、程序段、或代码的一部分包含一个或多个用于实现规定的逻辑功能的可执行指令。也应当注意,在有些作为替 换的实现中,方框中所标注的功能也可以以不同于附图中所标注的顺序发生。例如,两个接连地表示的方框实际上可以基本并行地执行,它们有时也可以按相反的顺序执行,这依所涉及的功能而定。也要注意的是,框图或流程图中的每个方框、以及框图或流程图中的方框的组合,可以用执行规定的功能或操作的专用的基于硬件的***来实现,或者可以用专用硬件与计算机指令的组合来实现。
描述于本申请实施例中所涉及到的单元可以通过软件的方式实现,也可以通过硬件的方式来实现,所描述的单元也可以设置在处理器中。其中,这些单元的名称在某种情况下并不构成对该单元本身的限定。
本领域技术人员在考虑说明书及实践这里公开的发明后,将容易想到本申请的其它实施方案。本申请旨在涵盖本申请的任何变型、用途或者适应性变化,这些变型、用途或者适应性变化遵循本申请的一般性原理并包括本申请未公开的本技术领域中的公知常识或惯用技术手段。说明书和实施例仅被视为示例性的,本申请的真正范围和精神由下面的权利要求指出。
应当理解的是,本申请并不局限于上面已经描述并在附图中示出的精确结构,并且可以在不脱离其范围进行各种修改和改变。本申请的范围仅由所附的权利要求来限制。
Claims (20)
- 一种会话信息处理方法,其特征在于,由电子设备执行,所述方法包括:当检测到待输出的会话信息时,获取目标对象对应的属性结构和会话结构;其中,所述会话结构用于表征所述会话信息;计算关键词集合对应的向量集合;其中,所述关键词集合包括所述属性结构中的关键词和所述会话结构中的关键词;根据所述向量集合生成联合语义向量;根据所述属性结构和所述会话结构生成联合结构向量;根据所述联合语义向量和所述联合结构向量确定所述目标对象与所述会话信息的匹配度,若所述匹配度满足预设条件则输出所述会话信息。
- 根据权利要求1所述的方法,其特征在于,所述获取目标对象对应的属性结构和会话结构之前,所述方法还包括:当检测到会话窗口中展示新增聊天消息时,确定与所述新增聊天消息对应的所述会话信息;其中,所述会话窗口中包括所述目标对象和至少一个会话对象,所述新增聊天消息与所述会话对象相对应。
- 根据权利要求2所述的方法,其特征在于,所述确定与所述新增聊天消息对应的所述会话信息,包括:根据所述新增聊天消息从预设信息库中选择所述会话信息;或者,根据所述新增聊天消息生成所述会话信息。
- 根据权利要求1所述的方法,其特征在于,所述根据所述联合语义向量和所述联合结构向量确定所述目标对象与所述会话信息的匹配度,包括:拼接所述联合语义向量和所述联合结构向量,得到综合向量;对所述综合向量进行降维处理,得到降维处理结果;计算所述降维处理结果属于各预设类别的概率;通过最高概率对应的预设类别表示所述匹配度。
- 根据权利要求1所述的方法,其特征在于,所述获取目标对象对应的属性结构和会话结构,包括:读取所述目标对象对应的属性结构;对所述会话信息进行结构化处理,得到所述会话结构。
- 根据权利要求5所述的方法,其特征在于,所述对所述会话信息进行结构化处理,得到所述会话结构,包括:将所述会话信息与预设词汇集合进行比对,以确定出所述会话信息中的关键词;根据预设词汇关系构建包含所述会话信息中的关键词的会话结构。
- 根据权利要求5所述的方法,其特征在于,所述对所述会话信息进行结构化处理,得到所述会话结构,包括:根据词向量模型确定所述会话信息对应的词向量;根据所述词向量对所述会话信息进行文本结构分析,得到所述会话信息中各语言单位之间的语义关系;根据所述语义关系将所述会话信息表示为所述会话结构。
- 根据权利要求7所述的方法,其特征在于,所述根据所述属性结构和所述会话结构生成联合结构向量,包括:将所述属性结构表示为字符串序列;根据所述字符串序列以及所述语义关系计算所述联合结构向量。
- 根据权利要求8所述的方法,其特征在于,所述根据所述字符串序列以及所述语义关系计算所述联合结构向量,包括:基于递归神经网络和所述字符串序列计算所述属性结构对应的第一参考向量;基于所述递归神经网络和所述语义关系计算所述会话结构对应的第二参考向量;将所述第一参考向量和所述第二参考向量拼接得到所述联合结构向量。
- 根据权利要求1所述的方法,其特征在于,所述计算关键词集合对应的向量集合,包括:根据预设向量关系确定所述关键词集合中各关键词的关键词向量;根据所述各关键词的位置信息确定所述各关键词对应的位置向量;根据所述各关键词分别对应的类别确定所述各关键词的类别向量;根据所述属性结构和所述会话结构确定所述各关键词的分隔向量;其中,所述属性结构中的关键词的分隔向量相同,所述会话结构中的关键词的分隔向量相同;将对应于同一关键词的关键词向量、位置向量、类别向量以及分隔向量进行对位相加,得到所述各关键词分别对应的待处理向量;将所述各关键词分别对应的待处理向量的集合确定为所述向量集合。
- 根据权利要求10所述的方法,其特征在于,所述根据所述各关键词的位置信息确定所述各关键词对应的位置向量之前,所述方法还包括:将所述属性结构中的关键词和所述会话结构中的关键词进行拼接,得到拼接结果;根据所述拼接结果的关键词排列顺序依次对所述关键词集合中各关键词标注位置信息,得到所述各关键词的位置信息。
- 根据权利要求10所述的方法,其特征在于,所述根据所述向量集合生成联合语义向量,包括:根据所述向量集合生成各所述待处理向量分别对应的矩阵组,所述矩阵组包括查询矩阵、键矩阵、值矩阵;根据各所述待处理向量分别对应的矩阵组计算各所述待处理向量分别对应的自注意力向量;根据前馈神经网络计算所述自注意力向量对应的特征向量;基于自注意力算法以及所述前馈神经网络计算所述特征向量对应的联合语义向量。
- 一种会话信息处理装置,其特征在于,包括:数据获取单元,用于当检测到待输出的会话信息时,获取目标对象对应的属性结构和会话结构;其中,所述会话结构用于表征所述会话信息;向量计算单元,用于计算关键词集合对应的向量集合;其中,所述关键词集合包括所述属性结构中的关键词和所述会话结构中的关键词;所述向量计算单元,还用于根据所述向量集合生成联合语义向量,并根据所述属性结构和所述会话结构生成联合结构向量;条件判定单元,用于根据所述联合语义向量和所述联合结构向量确定所述目标对 象与所述会话信息的匹配度,若所述匹配度满足预设条件则输出所述会话信息。
- 根据权利要求13所述的装置,其特征在于,所述装置还包括:信息确定单元,用于当检测到会话窗口中展示新增聊天消息时,确定与所述新增聊天消息对应的所述会话信息;其中,所述会话窗口中包括所述目标对象和至少一个会话对象,所述新增聊天消息与所述会话对象相对应。
- 根据权利要求14所述的装置,其特征在于,所述信息确定单元,具体用于根据新增聊天消息从预设信息库中选择会话信息;或者,根据新增聊天消息生成会话信息。
- 根据权利要求13所述的装置,其特征在于,所述条件判定单元,具体用于拼接所述联合语义向量和所述联合结构向量,得到综合向量;对所述综合向量进行降维处理,得到降维处理结果;计算所述降维处理结果属于各预设类别的概率;通过最高概率对应的预设类别表示所述匹配度。
- 根据权利要求13所述的装置,其特征在于,所述数据获取单元,具体用于读取目标对象对应的属性结构;对会话信息进行结构化处理,得到会话结构。
- 根据权利要求17所述的装置,其特征在于,所述数据获取单元,具体用于将会话信息与预设词汇集合进行比对,以确定出会话信息中的关键词;根据预设词汇关系构建包含会话信息中的关键词的会话结构。
- 一种计算机可读存储介质,其上存储有计算机程序,其特征在于,所述计算机程序被处理器执行时实现权利要求1-12任一项所述的方法。
- 一种电子设备,其特征在于,包括:处理器;以及存储器,用于存储所述处理器的可执行指令;其中,所述处理器配置为经由执行所述可执行指令来执行权利要求1-12任一项所述的方法。
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US17/974,798 US20230061778A1 (en) | 2020-09-14 | 2022-10-27 | Conversation information processing method, apparatus, computer- readable storage medium, and device |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010959812.4A CN111930915B (zh) | 2020-09-14 | 2020-09-14 | 会话信息处理方法、装置、计算机可读存储介质及设备 |
CN202010959812.4 | 2020-09-14 |
Related Child Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US17/974,798 Continuation US20230061778A1 (en) | 2020-09-14 | 2022-10-27 | Conversation information processing method, apparatus, computer- readable storage medium, and device |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2022052744A1 true WO2022052744A1 (zh) | 2022-03-17 |
Family
ID=73333428
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2021/112732 WO2022052744A1 (zh) | 2020-09-14 | 2021-08-16 | 会话信息处理方法、装置、计算机可读存储介质及设备 |
Country Status (3)
Country | Link |
---|---|
US (1) | US20230061778A1 (zh) |
CN (1) | CN111930915B (zh) |
WO (1) | WO2022052744A1 (zh) |
Families Citing this family (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111930915B (zh) * | 2020-09-14 | 2021-01-01 | 腾讯科技(深圳)有限公司 | 会话信息处理方法、装置、计算机可读存储介质及设备 |
CN113760568A (zh) * | 2021-01-04 | 2021-12-07 | 北京沃东天骏信息技术有限公司 | 数据处理的方法和装置 |
JP7329585B2 (ja) * | 2021-05-24 | 2023-08-18 | ネイバー コーポレーション | ペルソナチャットボット制御方法及びシステム |
CN115733632B (zh) * | 2021-08-26 | 2024-06-25 | 腾讯科技(深圳)有限公司 | 目标对象检测方法、装置、计算机设备及存储介质 |
US20230306203A1 (en) * | 2022-03-24 | 2023-09-28 | International Business Machines Corporation | Generating semantic vector representation of natural language data |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150058014A1 (en) * | 2012-05-17 | 2015-02-26 | Postech Academy-Industry Foundation | System and method for managing conversation |
CN106649746A (zh) * | 2016-12-26 | 2017-05-10 | 北京云知声信息技术有限公司 | 一种答案获取方法及装置 |
CN110083693A (zh) * | 2019-04-28 | 2019-08-02 | 腾讯科技(深圳)有限公司 | 机器人对话回复方法及装置 |
CN110209897A (zh) * | 2018-02-12 | 2019-09-06 | 腾讯科技(深圳)有限公司 | 智能对话方法、装置、存储介质及设备 |
CN111930915A (zh) * | 2020-09-14 | 2020-11-13 | 腾讯科技(深圳)有限公司 | 会话信息处理方法、装置、计算机可读存储介质及设备 |
Family Cites Families (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110597968A (zh) * | 2019-04-28 | 2019-12-20 | 河北省讯飞人工智能研究院 | 一种回复选择方法及装置 |
CN110069612B (zh) * | 2019-04-28 | 2022-02-01 | 河北省讯飞人工智能研究院 | 一种回复生成方法及装置 |
CN111079985B (zh) * | 2019-11-26 | 2021-07-06 | 昆明理工大学 | 一种基于bert并融合可区分属性特征的刑事案件刑期预测方法 |
-
2020
- 2020-09-14 CN CN202010959812.4A patent/CN111930915B/zh active Active
-
2021
- 2021-08-16 WO PCT/CN2021/112732 patent/WO2022052744A1/zh active Application Filing
-
2022
- 2022-10-27 US US17/974,798 patent/US20230061778A1/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150058014A1 (en) * | 2012-05-17 | 2015-02-26 | Postech Academy-Industry Foundation | System and method for managing conversation |
CN106649746A (zh) * | 2016-12-26 | 2017-05-10 | 北京云知声信息技术有限公司 | 一种答案获取方法及装置 |
CN110209897A (zh) * | 2018-02-12 | 2019-09-06 | 腾讯科技(深圳)有限公司 | 智能对话方法、装置、存储介质及设备 |
CN110083693A (zh) * | 2019-04-28 | 2019-08-02 | 腾讯科技(深圳)有限公司 | 机器人对话回复方法及装置 |
CN111930915A (zh) * | 2020-09-14 | 2020-11-13 | 腾讯科技(深圳)有限公司 | 会话信息处理方法、装置、计算机可读存储介质及设备 |
Non-Patent Citations (1)
Title |
---|
ANONYMOUS: "Introduction Introduction to Language Cloud and the Problems It Can Solve", 10 January 2016 (2016-01-10), pages 1 - 14, XP055910859, Retrieved from the Internet <URL:https://web.archive.org/web/20160110230027/http://www.ltp-cloud.com/intro> [retrieved on 20220408] * |
Also Published As
Publication number | Publication date |
---|---|
CN111930915B (zh) | 2021-01-01 |
CN111930915A (zh) | 2020-11-13 |
US20230061778A1 (en) | 2023-03-02 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2022052744A1 (zh) | 会话信息处理方法、装置、计算机可读存储介质及设备 | |
US20230206087A1 (en) | Techniques for building a knowledge graph in limited knowledge domains | |
Qi et al. | Finding all you need: web APIs recommendation in web of things through keywords search | |
US11250839B2 (en) | Natural language processing models for conversational computing | |
TW202009749A (zh) | 人機對話方法、裝置、電子設備及電腦可讀媒體 | |
US20180322188A1 (en) | Automatic conversation creator for news | |
EP3400557A1 (en) | Webinterface generation and testing using artificial neural networks | |
WO2022007438A1 (zh) | 情感语音数据转换方法、装置、计算机设备及存储介质 | |
WO2017210634A1 (en) | Iterative alternating neural attention for machine reading | |
JP2022050379A (ja) | 意味検索方法、装置、電子機器、記憶媒体およびコンピュータプログラム | |
US11874798B2 (en) | Smart dataset collection system | |
CN114201682A (zh) | 融合社交关系和语义关系的图神经网络推荐方法及*** | |
US20230088445A1 (en) | Conversational recommendation method, method of training model, device and medium | |
CN110275962B (zh) | 用于输出信息的方法和装置 | |
CN113779225B (zh) | 实体链接模型的训练方法、实体链接方法及装置 | |
JP2023017921A (ja) | コンテンツ推薦とソートモデルトレーニング方法、装置、機器、記憶媒体及びコンピュータプログラム | |
CN112214601A (zh) | 一种社交短文本情感分类方法、装置及存储介质 | |
JP2022545335A (ja) | 新語分類技術 | |
CN113569017A (zh) | 一种模型处理方法、装置、电子设备及存储介质 | |
CN113420212A (zh) | 基于深度特征学习的推荐方法、装置、设备及存储介质 | |
CN113641805A (zh) | 结构化问答模型的获取方法、问答方法及对应装置 | |
CN113779186A (zh) | 一种文本生成方法和装置 | |
KR20200041199A (ko) | 챗봇 구동 방법, 장치 및 컴퓨터 판독가능 매체 | |
US11558339B2 (en) | Stepwise relationship cadence management | |
US11893990B2 (en) | Audio file annotation |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 21865800 Country of ref document: EP Kind code of ref document: A1 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
32PN | Ep: public notification in the ep bulletin as address of the adressee cannot be established |
Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 01.08.2023) |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 21865800 Country of ref document: EP Kind code of ref document: A1 |