CN111191034B - Human-computer interaction method, related device and readable storage medium - Google Patents
Human-computer interaction method, related device and readable storage medium Download PDFInfo
- Publication number
- CN111191034B CN111191034B CN201911392035.3A CN201911392035A CN111191034B CN 111191034 B CN111191034 B CN 111191034B CN 201911392035 A CN201911392035 A CN 201911392035A CN 111191034 B CN111191034 B CN 111191034B
- Authority
- CN
- China
- Prior art keywords
- representation
- topic
- theme
- question
- determining
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- 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
- G06F16/35—Clustering; Classification
-
- 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
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Databases & Information Systems (AREA)
- Mathematical Physics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Human Computer Interaction (AREA)
- Computational Linguistics (AREA)
- Machine Translation (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The application discloses a human-computer interaction method, related equipment and a readable storage medium, wherein theme representations corresponding to each theme are configured in a human-computer interaction system, and as each theme comprises a plurality of problems, the number of the themes is far smaller than that of the problems, when human-computer interaction is required, the time spent for determining the similarity between a user input problem and each theme is far smaller than the time spent for determining the similarity between the user input problem and each problem, and therefore, the response efficiency of the human-computer interaction system can be improved.
Description
Technical Field
The present application relates to the field of natural language processing technologies, and in particular, to a human-computer interaction method, a related device, and a readable storage medium.
Background
With the rapid development of human-computer interaction technology, more and more human-computer interaction systems are widely applied. Currently, a commonly used human-computer interaction system responds to a user request based on a search method, and in such a human-computer interaction system, a problem library is usually stored, and feature vectors of all problems are stored in the problem library. When the human-computer interaction system is used for human-computer interaction, after a user inputs a question, the human-computer interaction system determines the feature vector of the question, performs similarity calculation on the feature vector of the question and the feature vectors of all questions in a question bank, and takes the answer corresponding to the question with high similarity as the response to the question.
However, the number of questions in the question bank is large, and the response efficiency is low by calculating the similarity between the feature vector of the question and the feature vectors of all the questions in the question bank.
Therefore, an optimized human-computer interaction method is needed to improve the response efficiency of the human-computer interaction system based on the retrieval method.
Disclosure of Invention
In view of the above problems, the present application is proposed to provide a human-computer interaction method, a related device and a readable storage medium. The specific scheme is as follows:
a human-computer interaction method, comprising:
acquiring a user input problem;
determining a target topic matched with the user input question based on the configured topic representations corresponding to each topic; the topic representation corresponding to any topic is determined by the representation of the sentence level and the representation of the word level of the question set under the topic;
responding to the user-entered question based on the answer to the target topic.
Preferably, the step of determining a target topic matching the user input question based on the configured topic representations corresponding to each topic comprises:
determining similarity of the user input question and each topic based on the configured topic representation corresponding to each topic;
and determining a target theme matched with the user input question from all themes according to the similarity between the user input question and each theme.
Preferably, the determining the similarity of the user input question to each topic based on the configured topic representations corresponding to each topic comprises:
determining a problem characterization for the user input problem;
and calculating the similarity of the problem representation and the topic representation corresponding to the topic as the similarity of the user input problem and the topic aiming at each topic.
Preferably, the method further comprises the following steps:
when determining that the theme representation corresponding to the theme to be updated needs to be updated, acquiring a newly-added problem of the theme to be updated;
storing the newly-added questions into a question set under a subject to be updated to obtain a new question set under the subject to be updated;
and determining the representation of the sentence level and the representation of the word level of the new question set, and determining the updated topic representation of the topic to be updated according to the representation of the sentence level and the representation of the word level.
Preferably, the topic characterization of any topic is determined as follows:
determining a question set under the theme and keywords of the question set;
determining a representation of a sentence hierarchy of the problem set;
determining a representation of a term hierarchy of the problem set based on the keywords;
and determining the topic representation of the topic based on the representation of the sentence hierarchy and the representation of the word hierarchy.
Preferably, determining a set of questions under the topic comprises:
acquiring all the problems under the theme;
and determining a problem set under the theme based on all the problems under the theme.
Preferably, the process of determining the sentence-level and word-level representations of the problem set and determining the topic representation of the topic based thereon comprises:
processing the problem set under the theme and the keywords of the problem set by using a theme representation model to obtain the theme representation of the theme output by the theme representation model;
the topic representation model is provided with a representation for determining sentence level of the question set; determining a representation of a term hierarchy of the problem set based on the keywords; determining a capability of topic characterization of the topic based on the characterization of the sentence hierarchy and the characterization of the word hierarchy.
Preferably, the processing, by using the topic representation model, the problem set under the topic and the keywords of the problem set to obtain the topic representation of the topic output by the topic representation model includes:
screening the representations of the problems in the problem set under the theme based on an attention mechanism by using a sentence level processing module of a theme representation model to obtain the representations of the sentence levels of the problem set;
determining the representation of the word hierarchy of the problem set based on the keywords of the problem set and the weight of each keyword serving as the contribution information of the theme by using a word hierarchy processing module of a theme representation model;
and utilizing a fusion module of a topic representation model to fuse the representation of the sentence level of the question set and the representation of the word level of the question set to obtain the topic representation of the topic.
Preferably, the sentence level processing module of the topic representation model processes the question set under the topic, and the sentence level representation of the question set is obtained in the following manner:
acquiring an initial vector of the problem set;
determining the weight of the vector of each screened problem on the initial vector based on an attention mechanism;
and carrying out weighted addition on the initial vectors based on the determined weight values of the vectors of each screened problem on the initial vectors to obtain vector representations of each screened problem, wherein the vector representations are used as the representations of sentence levels of the problem set.
Preferably, the word hierarchy processing module of the topic representation model processes the keywords of the problem set to obtain the word hierarchy representation of the problem set in the following manner:
acquiring an initial vector of the keyword;
determining the weight of the keyword;
and carrying out weighting processing on the initial vector of the keyword based on the weight of the keyword to obtain the representation of the word hierarchy of the problem set.
Preferably, the method for obtaining the topic representation of the topic includes the following steps of fusing the sentence level representation of the question set and the word level representation of the question set by using a fusion module of a topic representation model:
converting the representation of the word hierarchy of the question set into a target vector, wherein the target vector is the same as the representation dimension of the sentence hierarchy of the question set;
and splicing the target vector and the representation of the sentence level of the question set to obtain the topic representation of the topic.
A human-computer interaction device, comprising:
an input question acquisition unit for acquiring a user input question;
the target theme determining unit is used for determining a target theme matched with the user input question based on the configured theme representation corresponding to each theme; the topic representation corresponding to any topic is determined by the representation of sentence level and the representation of word level of the topic set under the topic;
and the question response unit is used for responding to the user input question based on the answer of the target theme.
Preferably, the target subject determination unit includes:
the similarity determining unit is used for determining the similarity of the user input question and each topic based on the configured topic representation corresponding to each topic;
and the similarity using unit is used for determining a target theme matched with the user input question from all themes according to the similarity between the user input question and each theme.
Preferably, the similarity determination unit includes:
the input problem representation determining unit is used for determining the problem representation of the user input problem;
and the problem representation calculating unit is used for calculating the similarity of the problem representation and the topic representation corresponding to the topic as the similarity of the user input problem and the topic.
Preferably, the method further comprises the following steps:
the first updating unit is used for acquiring a newly-added problem of the theme to be updated when the theme representation corresponding to the theme to be updated is determined to be required to be updated;
the second updating unit is used for storing the newly increased problems into a problem set under a theme to be updated to obtain a new problem set under the theme to be updated;
and the third updating unit is used for determining the representation of the sentence level and the representation of the word level of the new question set and determining the updated topic representation of the topic to be updated according to the representation of the sentence level and the representation of the word level.
Preferably, the method further comprises the following steps: the theme representation determining unit is used for determining the theme representation corresponding to each theme, wherein the theme representation of any theme is determined in the following way:
determining a problem set under the theme and keywords of the problem set;
determining a representation of a sentence hierarchy of the problem set;
determining a representation of a term hierarchy of the problem set based on the keywords;
and determining the topic representation of the topic based on the representation of the sentence hierarchy and the representation of the word hierarchy.
Preferably, the process of determining the sentence level and word level of the problem set and determining the topic representation of the topic based on the sentence level and word level, which comprises the following steps:
processing the problem set under the theme and the keywords of the problem set by using a theme representation model to obtain the theme representation of the theme output by the theme representation model;
the topic representation model is provided with a representation for determining sentence level of the question set; determining a representation of a term hierarchy of the problem set based on the keywords; determining a capability of topic representation of the topic based on the representation of sentence hierarchy and the representation of word hierarchy.
Preferably, the process of processing, by the topic representation determining unit, the problem set under the topic and the keyword of the problem set by using the topic representation model to obtain the topic representation of the topic output by the topic representation model includes:
screening the representations of the problems in the problem set under the theme based on an attention mechanism by using a sentence level processing module of a theme representation model to obtain the representations of the sentence levels of the problem set;
determining the representation of the word hierarchy of the problem set based on the keywords of the problem set and the weight of each keyword serving as the contribution information of the theme by using a word hierarchy processing module of a theme representation model;
and utilizing a fusion module of a topic representation model to fuse the representation of the sentence level of the question set and the representation of the word level of the question set to obtain the topic representation of the topic.
Preferably, the process of screening the representations of the questions in the question set under the topic based on an attention mechanism by the topic representation determining unit using a sentence level processing module of the topic representation model to obtain the sentence level representations of the question set includes:
acquiring an initial vector of the problem set;
determining the weight of the vector of each screened problem on the initial vector based on an attention mechanism;
and carrying out weighted addition on the initial vectors based on the determined weight values of the vectors of each screened problem on the initial vectors to obtain vector representations of each screened problem, wherein the vector representations are used as the representations of sentence levels of the problem set.
Preferably, the process of determining, by the topic representation determining unit, the representation of the term hierarchy of the problem set based on the keywords of the problem set and the weight of each keyword contributing information to the topic by using the term hierarchy processing module of the topic representation model includes:
acquiring an initial vector of the keyword;
determining the weight of the keyword;
and carrying out weighting processing on the initial vector of the keyword based on the weight of the keyword to obtain the representation of the word hierarchy of the problem set.
Preferably, the topic representation determining unit utilizes a fusion module of a topic representation model to fuse the representation of the sentence hierarchy of the question set and the representation of the word hierarchy of the question set, and the topic representation of the topic is obtained in the following manner:
converting the representation of the word hierarchy of the question set into vector representation of the same dimension based on the dimension of the representation of the sentence hierarchy of the question set;
and splicing the converted word level representation and the sentence level representation of the question set to obtain the topic representation of the topic.
A human-computer interaction device, comprising: a memory and a processor;
the memory is used for storing programs;
the processor is configured to execute the program, and implement the steps of the human-computer interaction method.
A readable storage medium having stored thereon a computer program which, when executed by a processor, carries out the steps of the human-computer interaction method as described above.
By means of the technical scheme, the application discloses a human-computer interaction method, related equipment and a readable storage medium, theme representations corresponding to all themes are configured in a human-computer interaction system, and because each theme comprises a plurality of problems, the number of the themes is far smaller than that of the themes, when human-computer interaction is needed, the time spent on determining the similarity between a user input question and each theme is far smaller than the time spent on determining the similarity between the user input question and each question, and therefore the response efficiency of the human-computer interaction system can be improved.
Furthermore, the representation of the sentence level of the question set can represent each question in the question set from the sentence level. And the representation of the word hierarchy of the problem set can represent each problem in the problem set from the word hierarchy. The method and the device have the advantages that the representation of sentence levels and word levels is considered, the problem set under the theme can be represented more accurately based on the determined theme representation, the determined target theme matched with the user input problem is more accurate, and the response output to the user is more accurate.
Drawings
Various additional advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The drawings are only for purposes of illustrating the preferred embodiments and are not to be construed as limiting the application. Also, like reference numerals are used to refer to like parts throughout the drawings. In the drawings:
fig. 1 is a schematic flow chart of a human-computer interaction method according to an embodiment of the present disclosure;
fig. 2 is a schematic view of several problems under the subject of a "kentucky cone" in an example of an embodiment of the present application;
FIG. 3 is a schematic diagram of a training target of a topic characterization model according to an example of the present application;
FIG. 4 is a schematic diagram of an exemplary topic characterization model architecture according to an embodiment of the present application;
fig. 5 is a schematic structural diagram of a human-computer interaction device according to an embodiment of the present application;
fig. 6 is a schematic structural diagram of a human-computer interaction device according to an embodiment of the present application.
Detailed Description
The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are only a part of the embodiments of the present application, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present application.
The application provides a human-computer interaction scheme, which can be applied to human-computer interaction equipment, wherein the human-computer interaction equipment can receive input operation of a user and give a response. Generally, the human-computer interaction device can be a common electronic device with input capability and output capability, such as a mobile phone, a computer, an IPAD, and other electronic devices for realizing interactive entertainment with a user.
An alternative applicable scenario is that the user inputs the question to the human-computer interaction device by voice, keyboard or other input means: "how the kendiry sweet cylinder tastes", the human-computer interaction device may determine the response information to the input question based on the scheme provided in the present application, for example, the determined response information is: the sweet cones have strawberry, chocolate and vanilla flavors for selection, and then the response information can be output in modes of voice broadcasting, character display and the like, so that man-machine interaction is realized.
It can be understood that the response rate and accuracy of the machine to the user input problem in the human-computer interaction are important factors influencing the human-computer interaction, and the present application provides a human-computer interaction scheme from the direction of improving the response rate and accuracy of the machine, and the following embodiments are used to introduce the human-computer interaction method provided by the present application.
As described in conjunction with fig. 1, the human-computer interaction method of the present application may include the following steps:
and step S100, acquiring a user input question.
Specifically, according to different supports of the human-computer interaction equipment on the user input modes, the method and the device can acquire the problems input by the user in different input modes. Common input modes include voice mode, keyboard input, handwriting input, and the like.
And step S110, determining a target topic matched with the user input question based on the configured topic representation corresponding to each topic.
Specifically, for a large number of questions stored in the question bank, the questions can be divided according to the topics to which the questions belong, and then the question set under each topic can be obtained.
It will be appreciated that the subject matter of the questions in the question sets under each topic is the same, and that the questions in the question sets are meant the same and correspond to the subject matter of the question sets, although the questions in the question sets may be similarly or distinctly different in their presentation.
Based on this, the subject characterization of the corresponding subject of each problem set can be determined. The topic representation can represent the meaning of the corresponding topic, and the topic representation can represent each question in the question set. In this case, the system may only store topic representations corresponding to the topics of each question set, and the number of the question sets is obviously much smaller than the number of the questions, that is, the number of the stored topic representations is also much smaller than the number of the question representations of each question stored in the prior art, thereby greatly reducing the storage pressure.
Further, after determining the topic characterization corresponding to each topic, a target topic matching the user input question may be determined from the respective topics based on the topic characterization. Obviously, the method and the device only need to perform matching calculation on the topic representation of each topic and the user input problems, and avoid the problem that the prior art needs to perform matching calculation on the user input problems and each problem in a problem library, thereby greatly reducing the calculation amount and greatly improving the response rate of the machine to the user.
It should be further noted that, in the present application, topic representations corresponding to any topic may be determined based on the sentence-level representations and the word-level representations of the topic set under the topic.
The representation of the sentence level of the question set can represent each question in the question set from the sentence level. And the representation of the word hierarchy of the problem set can represent each problem in the problem set from the word hierarchy. The topic representation determined based on the sentence level and the word level can more accurately represent the problem set under the topic, namely the topic representation can contain the multi-level meaning of the problem set.
In this case, the topic representation determined based on the multi-level representation is more accurate, and thus the target topic determined based on the multi-level representation and matched with the user input problem is more accurate.
And step S120, responding to the question input by the user based on the answer of the target theme.
Specifically, the response answer corresponding to each topic can be configured, and on the basis, the answer of the target topic can be output, so that the response to the user input question is completed.
According to the human-computer interaction method, the theme representations corresponding to the themes are configured in the human-computer interaction system, and as each theme comprises a plurality of problems, the number of the themes is far smaller than that of the problems, when human-computer interaction is required, the time spent on determining the similarity between the user input problem and each theme is far smaller than the time spent on determining the similarity between the user input problem and each problem, and therefore the response efficiency of the human-computer interaction system can be improved.
Furthermore, the representation of the sentence level of the question set can represent each question in the question set from the sentence level. And the representation of the word hierarchy of the problem set can represent each problem in the problem set from the word hierarchy. The method and the device have the advantages that the representation of sentence levels and word levels is considered, the problem set under the theme can be represented more accurately based on the determined theme representation, the determined target theme matched with the user input problem is more accurate, and the response output to the user is more accurate.
In another embodiment of the present application, a process of determining a target topic matching the user input question based on the configured topic characterization corresponding to each topic in step S110 is described.
In an alternative embodiment, the process of determining the target topic may include:
s1, determining similarity of the user input question and each theme based on configured theme representations corresponding to the themes.
Specifically, the topic representation corresponding to each topic can represent the topic, and then the similarity between each topic and the user input question can be determined based on the topic representations, wherein the similarity represents the semantic similarity between the user input question and each topic.
It will be appreciated that the higher the similarity, the more similar the semantics of the corresponding topic are to the user input question.
When the similarity calculation is performed, the problem representation of the user input problem may be determined first, and then, for each topic, the similarity between the problem representation and the topic representation corresponding to the topic is calculated as the similarity between the user input problem and the topic.
The similarity can be in a numerical form, and the similarity can be determined by the numerical size. In addition, the similarity may also be in a category form, and if the similarity is divided into three categories, namely high, medium, and low, in advance, the size of the similarity may be determined according to the category of the similarity.
The problem representation of the user input problem can be determined in various ways, such as by a sentence semantic model modeling sentence semantics.
S2, determining a target theme matched with the user input question from all themes according to the similarity between the user input question and each theme.
Specifically, the similarity between the user input question and each topic is determined, and the topic with the highest similarity may be selected as the target topic matched with the user input question according to the similarity. Or, a similarity threshold value may be configured, and the topic with the highest similarity is selected from the topics with the similarity exceeding the threshold value as the target topic for matching the user input question. Of course, if there is no theme whose similarity exceeds the threshold, it may be considered that there is no theme matching the user input problem, and a user response cannot be given, and a prompt may be output.
It will be appreciated that problems are increasing progressively over time, and it is clear that it is not possible to cover all problems in the problem bank, i.e. there are cases where new problems are added to the problem bank.
In the prior art, the problem representation corresponding to each problem in the problem library is determined through a unified sentence representation model. When a newly added problem is added into the problem library, the sentence representation model needs to be retrained and updated aiming at the newly added problem, and the problem representation of each problem in the problem library is determined again, so that the correct response of man-machine interaction is influenced, and unpredictable risks are generated.
In the application, for the newly added problem, only the problem set under the topic to which the newly added problem belongs needs to be updated, and the topic representation corresponding to the problem set is determined again, the influence range of the operation is limited under the problem set corresponding to one topic, and the problem sets of other topics cannot be influenced, so that the robustness of human-computer interaction is improved.
Next, the update process is described, which may specifically include:
s1, when the theme representation corresponding to the theme to be updated needs to be updated, acquiring a newly-added problem of the theme to be updated.
Specifically, when a newly added problem is obtained, a theme to which the newly added problem belongs may be determined, and the theme is used as a theme to be updated.
And S2, storing the newly-added questions into a question set under the subject to be updated to obtain a new question set under the subject to be updated.
For newly added problems, the newly added problems can be stored in the problem set under the theme to be updated, and completing the expansion of the problem set.
And S3, determining the sentence level representation and the word level representation of the new problem set, and determining the updated theme representation of the theme to be updated according to the sentence level representation and the word level representation.
For the new problem set after the updating is finished, the sentence level representation and the word level representation of the new problem set can be determined, and the updated theme representation of the theme to be updated is determined according to the sentence level representation and the word level representation of the new problem set.
And at this point, the updating of the problem set of the subject to be updated and the updating of the subject representation are completed.
In yet another embodiment of the present application, a process for determining a subject characterization of the subject matter described in the above scheme is described.
The topic characterization for any topic may be determined by:
s1, determining a problem set under the theme and keywords of the problem set.
Wherein, the question set under the subject can be all the questions under the subject in the question bank. In addition, a plurality of questions can be selected from all the questions under the subject in the question bank to form a question set.
It will be appreciated that there are a number of questions in the question bank that belong to the subject, which questions are of the same meaning but may be of similar or distinct expressions. As shown in fig. 2, which illustrates several problems that belong to the topic of "kentucky cones". Among them, the expression of the problem "how sweet cylindrical taste is peculiar" and "sweet cylindrical taste peculiar" are similar to each other, but the expression of the problem "whether the taste of your sweet cylindrical is peculiar or not" is two very different expressions.
The method can adopt random sampling or other modes to select M questions from all the questions under the theme in the question bank to form a question set under the theme.
Further, a keyword of the problem set under the topic needs to be determined, the keyword is a class of words with a large contribution degree to the current topic, and the keyword is used as important information for distinguishing different topics.
The method can determine the keywords of the problem set in various ways, for example, the importance degree of each word in the problem set is represented by the word frequency-inverse document frequency TF-IDF, and the topK words with the maximum TF-IDF value are selected to form the keyword set of the problem set under the theme.
Still referring to fig. 2, keywords under the topic of "kentucky cone" may include: "cone", "taste", "mouthfeel", etc.
And S2, determining the representation of the sentence hierarchy of the question set.
Specifically, the representation of the sentence level of the question set can represent each question in the question set from the sentence level.
In this step, the representation of the question set may be determined from the sentence level based on the questions included in the question set, and used as the representation of the sentence level of the question set.
And S3, determining the representation of the word hierarchy of the question set based on the keywords.
Specifically, the representation of the term hierarchy of the problem set can represent each problem in the problem set from the term hierarchy.
In this step, the representation of the problem set may be determined from the term hierarchy based on each keyword corresponding to the problem set, and used as the representation of the term hierarchy of the problem set.
And S4, determining the topic representation of the topic based on the representation of the sentence hierarchy and the representation of the word hierarchy.
According to the scheme, the representation of sentence levels and word levels is considered at the same time, the topic representation determined based on the representation can more accurately represent the problem set under the topic, namely the topic representation can contain the multi-level meaning of the problem set.
In another embodiment of the present application, the above S2-S4: an alternative implementation of the process of determining sentence-level and word-level tokens of a problem set and determining topic tokens therefrom.
Specifically, in this embodiment, the above process may be implemented by a neural network model. The method can pre-train a topic representation model, wherein the trained topic representation model is used for processing a problem set under an input topic and keywords of the problem set so as to determine the representation of sentence levels of the problem set; determining the representation of the word hierarchy of the question set based on the keywords; determining a capability of topic characterization of the topic based on the characterization of the sentence hierarchy and the characterization of the word hierarchy.
The topic representation model training process may use a question set and keywords of the question set under each topic in the topic library as training data, and the training targets may include the following:
topic representation r of topic t of model output t Approaching to the problem representation of each problem in the problem set under the subject t, and outputting the subject representation r of the subject t by the model t The problem characterization of each problem in the problem set under the other topics than the topic t becomes far away.
Specifically, referring to the training target diagram of the topic representation model illustrated in fig. 3:
each topic is contained in FIG. 3 t Under (t epsilon (1,x), x is the total number of topics) there is a corresponding question q (the total number of questions under each topic is M).
Q in FIG. 3 t Showing the individual questions under the topic t,representing individual questions under non-topic t.
The training objective is that the topic representation of the topic t is similar to the representation of each question under the topic t and is far away from each question under other topics than the topic t.
And updating the training parameters of the theme representation model according to the training mode to finally obtain the trained theme representation model.
Further, the embodiment of the present application illustrates an alternative architecture of the topic characterization model.
Turning to FIG. 4, an architecture of a topic characterization model is illustrated.
The topic characterization model may include: a sentence level processing module, a word level processing module and a fusion module. Each module is described separately below.
First, a,
And the sentence level processing module is used for screening the representation of the problem set problem under the theme based on an attention mechanism to obtain the representation of the sentence level of the problem set.
Specifically, because the problem set under the theme contains a plurality of problems with the same meaning but different expression modes, in order to obtain some representative problems under the theme, the method and the device can express the characterization of each problem in the problem set through N groups of attention through the attention mechanism, so as to obtain problem representations of N perspectives, namely, the characterization of the problem in the problem set under the theme is screened based on the attention mechanism, so as to obtain the sentence-level characterization of the problem set.
The specific implementation process can comprise the following steps:
s1, obtaining an initial vector of the problem set.
Specifically, defining the number of problems contained in the problem set as M, determining the problem representation of the M problems in the problem set, wherein the problem representation can be in a vector form, and then obtaining the problem representation
Wherein h is m Problem characterization vectors representing m problems of the ground, d 0 Is the dimension of the vector.
And S2, determining the weight of the vector of each screened problem on the initial vector based on the attention mechanism.
Specifically, a fixed-value parameter N of the attention mechanism may be configured to represent that M questions are extracted into N representative questions based on the attention mechanism, and the implementation is performed by performing N times of self-attention transformations on initial vectors of the M questions in the question set, that is, mapping the initial vectors to obtain a set of non-normalized weights [ e ] according to the following formula (1) jk ]Wherein e is jk The screened j (1 ≦ j ≦ N) problem vector is represented as the original initial vector h k (k is more than or equal to 1 and less than or equal to M) and the weight, U, of the unnormalized value j 、W j 、b j Are all model training parameters.
Further, the weight normalization is performed by the following formula (2) to obtain a group of weight vectors, which are expressed as [ alpha ] jk ] N×M In which α is jk The screened j (1 ≦ j ≦ N) th problem vector is represented in the original initial vector h k And (k is more than or equal to 1 and less than or equal to M) is normalized.
[e jk ]=U j *tanh(W j *H+b j ) (1)
And S3, carrying out weighted addition on the initial vectors based on the determined weight values of the vectors of each screened problem on the initial vectors to obtain vector representations of each screened problem, wherein the vector representations are used as the representations of sentence levels of the problem set.
Specifically, the problem representation Q at N views is obtained by weighting the vectors: [ q ] of 1 ,q 2 ...q j ...,q N ],q j ∈R 1*d Wherein q is j Expressed as the jth weighted by the initial vector of questions in the original question setThe problem vector of the new problem, the formula is as follows:
the second step,
And the word hierarchy processing module is used for determining the representation of the word hierarchy of the problem set based on the keywords of the problem set and the weight of the contribution information of each keyword as the theme.
Specifically, the word-level processing module may obtain the word-level representation as follows:
s1, obtaining an initial vector of the keyword.
The initial vector of keywords may be derived by random initialization and may be denoted as w 1 ,w 2 ...w k ...,w K ],w k ∈R 1*l Wherein w is k And representing an initial word vector of the keyword k, wherein the dimension is l dimension.
And S2, determining the weight of the keyword.
As described above, in the model training phase, the application may determine the keywords of the problem set in various ways, for example, the importance degree of each word in the problem set is represented by the word frequency-inverse document frequency TF-IDF, and the topK words with the largest TF-IDF value are selected to form the keyword set of the problem set under the topic. Therefore, the TF-IDF value of the keyword may be used as the initial weight of the keyword. That is, the TF-IDF values of the keywords are used as the weight vector [ beta ] in the training phase t1 ,β t2 ,...,β tK ]For representing the weight of each keyword contributing information to the topic t. The weight vector [ beta ] t1 ,β t2 ,...,β tK ]Continuously optimizing and updating in the training stage of the theme representation model, and determining the final weight vector [ beta ] after the model is trained t1 ,β t2 ,...,β tK ]As a weight for the keyword.
And S3, carrying out weighting processing on the initial vector of the keyword based on the weight of the keyword to obtain the representation of the word hierarchy of the problem set.
Specifically, the final word vector of the keyword is an initial vector and a weight vector [ beta ] t1 ,β t2 ,...,β tK ]By a corresponding multiplication of v k =β tk *w k . Finally, obtaining a word hierarchy representation V of the question set under the theme: [ v ] of 1 ,v 2 ...v k ...,v K ],v k ∈R 1*l 。
Third a (c),
And the fusion module is used for fusing the representation of the sentence level of the problem set and the representation of the word level of the problem set to obtain the theme representation of the theme.
Specifically, the fusion module may obtain the topic representation of the topic as follows:
s1, converting the representation of the sentence level of the question set into vector representation of the same dimension based on the dimension of the representation of the sentence level of the question set.
Specifically, the sentence level representation Q = [ Q ] of the question set can be obtained through the sentence level processing module 1 ,q 2 ,...,q N ]The word hierarchy processing module can obtain the word hierarchy representation V = [ V ] of the problem set 1 ,v 2 ,...,v K ]。
In order to realize the fusion of the word-level representation and the sentence-level representation, as shown in fig. 4, the present application may introduce a feature mapping layer f (V), that is, a word-level representation is mapped to a vector at the same latitude as the sentence-level representation by matrix transformation.
And S2, splicing the converted word level representation and the sentence level representation of the question set to obtain the topic representation of the topic.
Specifically, through dimension conversion, the representation of the sentence level and the representation of the word level are in the same dimension, so that the sentence level and the word level can be directly spliced to obtain the spliced topic representation.
With reference to FIG. 4, after Q and f (V) are spliced, a topic representation r can be obtained through a full connection layer g (-) output t The formula is as follows:
r t =g([Q,f(V)]) (4)
it should be noted that, in the above S1, based on the dimension of the representation of the sentence level of the question set, the process of converting the representation of the word level of the question set into the vector representation of the same dimension may be implemented by a fusion module. In addition, the method can be implemented by a word hierarchy processing module, that is, after the word hierarchy processing module obtains the representation of the word hierarchy, the word hierarchy processing module can perform dimension conversion on the representation of the word hierarchy to convert the representation of the word hierarchy into the same dimension as the representation of the sentence hierarchy of the question set. On the basis, the fusion module can directly perform fusion without dimension conversion. Based on this, the module division of the topic representation model illustrated in fig. 4 is only an optional way.
The following describes the human-computer interaction device provided in the embodiments of the present application, and the human-computer interaction device described below and the human-computer interaction method described above may be referred to correspondingly.
Referring to fig. 5, fig. 5 is a schematic structural diagram of a human-computer interaction device disclosed in the embodiment of the present application.
As shown in fig. 5, the apparatus may include:
an input question acquisition unit 11 for acquiring a user input question;
a target topic determination unit 12, configured to determine a target topic matching the user input question based on the configured topic tokens corresponding to each topic; the topic representation corresponding to any topic is determined by the representation of the sentence level and the representation of the word level of the question set under the topic;
a question response unit 13, configured to respond to the user input question based on the answer of the target topic.
Further, the target subject determination unit may include:
the similarity determining unit is used for determining the similarity of the user input question and each topic based on the configured topic representation corresponding to each topic;
and the similarity using unit is used for determining a target theme matched with the user input question from all themes according to the similarity between the user input question and each theme.
Further, the similarity determination unit may include:
the input problem representation determining unit is used for determining the problem representation of the user input problem;
and the problem representation calculating unit is used for calculating the similarity of the problem representation and the topic representation corresponding to the topic as the similarity of the user input problem and the topic.
Further, the apparatus of the present application may further include:
the first updating unit is used for acquiring a newly-added problem of the theme to be updated when the theme representation corresponding to the theme to be updated is determined to be required to be updated;
the second updating unit is used for storing the newly increased problems into a problem set under a theme to be updated to obtain a new problem set under the theme to be updated;
and the third updating unit is used for determining the representation of the sentence level and the representation of the word level of the new question set and determining the updated topic representation of the topic to be updated according to the representation of the sentence level and the representation of the word level.
Further, the apparatus of the present application may further include: the theme representation determining unit is used for determining the theme representation corresponding to each theme, wherein the theme representation of any theme is determined in the following way:
determining a problem set under the theme and keywords of the problem set;
determining a representation of a sentence hierarchy of the problem set;
determining a representation of a term hierarchy of the problem set based on the keywords;
and determining the topic representation of the topic based on the representation of the sentence hierarchy and the representation of the word hierarchy.
Further, the process of determining the sentence level and word level of the question set by the topic characterization determination unit and determining the topic characterization of the topic based on the sentence level and word level may include:
processing the problem set under the theme and the keywords of the problem set by using a theme representation model to obtain the theme representation of the theme output by the theme representation model;
the topic representation model is provided with a representation for determining the sentence level of the question set; determining a representation of a term hierarchy of the problem set based on the keywords; determining a capability of topic characterization of the topic based on the characterization of the sentence hierarchy and the characterization of the word hierarchy.
Further, the process of processing the problem set and the keywords of the problem set under the topic by the topic representation determining unit using the topic representation model to obtain the topic representation of the topic output by the topic representation model may include:
screening the representations of the problems in the problem set under the theme based on an attention mechanism by utilizing a sentence level processing module of a theme representation model to obtain the representations of the sentence levels of the problem set;
determining the representation of the word hierarchy of the problem set based on the keywords of the problem set and the weight of each keyword serving as the contribution information of the theme by using a word hierarchy processing module of a theme representation model;
and utilizing a fusion module of a topic representation model to fuse the representation of the sentence level of the question set and the representation of the word level of the question set to obtain the topic representation of the topic.
Further, the process of using the sentence level processing module of the topic representation model to screen the representations of the questions in the question set under the topic based on the attention mechanism by the topic representation determining unit to obtain the representations of the sentence levels of the question set may include:
acquiring an initial vector of the problem set;
determining the weight of the vector of each screened problem on the initial vector based on an attention mechanism;
and performing weighted addition on the initial vectors based on the determined weight values of the vectors of each screened problem on the initial vectors to obtain the vector representation of each screened problem as the representation of the sentence hierarchy of the problem set.
Further, the process of determining the word hierarchy representation of the problem set by the topic representation determining unit using the word hierarchy processing module of the topic representation model based on the keywords of the problem set and the weight of each keyword as the contribution information of the topic may include:
acquiring an initial vector of the keyword;
determining the weight of the keyword;
and carrying out weighting processing on the initial vector of the keyword based on the weight of the keyword to obtain the representation of the word hierarchy of the problem set.
Further, the topic representation determining unit uses a fusion module of a topic representation model to fuse the representation of the sentence hierarchy of the question set and the representation of the word hierarchy of the question set, and the way of obtaining the topic representation of the topic may be as follows:
converting the representation of the word level of the problem set into vector representation with the same dimension based on the dimension of the representation of the sentence level of the problem set;
and splicing the converted word level representation and the sentence level representation of the question set to obtain the topic representation of the topic.
The human-computer interaction device provided by the embodiment of the application can be applied to human-computer interaction equipment such as mobile phones, computers and the like. Optionally, fig. 6 shows a block diagram of a hardware structure of the human-computer interaction device, and referring to fig. 6, the hardware structure of the human-computer interaction device may include: at least one processor 1, at least one communication interface 2, at least one memory 3 and at least one communication bus 4;
in the embodiment of the application, the number of the processor 1, the communication interface 2, the memory 3 and the communication bus 4 is at least one, and the processor 1, the communication interface 2 and the memory 3 complete mutual communication through the communication bus 4;
the processor 1 may be a central processing unit CPU, or an Application Specific Integrated Circuit ASIC (Application Specific Integrated Circuit), or one or more Integrated circuits configured to implement embodiments of the present invention, etc.;
the memory 3 may include a high-speed RAM memory, and may further include a non-volatile memory (non-volatile memory) or the like, such as at least one disk memory;
wherein the memory stores a program and the processor can call the program stored in the memory, the program for:
acquiring a user input problem;
determining a target topic matched with the user input question based on the configured topic representations corresponding to each topic; the topic representation corresponding to any topic is determined by the representation of the sentence level and the representation of the word level of the question set under the topic;
responding to the user-entered question based on the answer to the target topic.
Alternatively, the detailed function and the extended function of the program may refer to the above description.
Embodiments of the present application further provide a readable storage medium, where a program suitable for being executed by a processor may be stored, where the program is configured to:
acquiring a user input problem;
determining a target topic matched with the user input question based on the configured topic representations corresponding to each topic; the topic representation corresponding to any topic is determined by the representation of the sentence level and the representation of the word level of the question set under the topic;
responding to the user-entered question based on the answer to the target topic.
Alternatively, the detailed function and the extended function of the program may be as described above.
Finally, it should also be noted that, herein, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising a … …" does not exclude the presence of another identical element in a process, method, article, or apparatus that comprises the element.
The embodiments in the present description are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other.
The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the application. Thus, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims (13)
1. A human-computer interaction method, comprising:
acquiring a user input problem;
determining a target topic matched with the user input question based on the configured topic representations corresponding to each topic; the topic representation corresponding to any topic is determined by the representation of the sentence level and the representation of the word level of the question set under the topic; the representation of the sentence level of the question set under the theme is obtained by screening the representation of the questions in the question set under the theme based on an attention mechanism, and the representation of the word level of the question set under the theme is determined based on the keywords of the question set and the weight of the contribution information of each keyword as the theme;
responding to the user-entered question based on the answer to the target topic.
2. The method of claim 1, wherein determining a target topic that matches the user input question based on the configured topic tokens corresponding to each topic comprises:
determining similarity of the user input question and each topic based on the configured topic representation corresponding to each topic;
and determining a target theme matched with the user input question from all themes according to the similarity between the user input question and each theme.
3. The method of claim 2, wherein determining the similarity of the user input question to each topic based on the configured topic tokens corresponding to each topic comprises:
determining a problem characterization for the user input problem;
and calculating the similarity of the problem representation and the topic representation corresponding to the topic as the similarity of the user input problem and the topic aiming at each topic.
4. The method of claim 1, further comprising:
when determining that a theme representation corresponding to a theme to be updated needs to be updated, acquiring a newly-added problem of the theme to be updated;
storing the newly-added questions into a question set under a theme to be updated to obtain a new question set under the theme to be updated;
and determining the representation of the sentence level and the representation of the word level of the new question set, and determining the updated topic representation of the topic to be updated according to the representation of the sentence level and the representation of the word level.
5. The method of claim 1, wherein the topic characterization of any topic is determined by:
determining a problem set under the theme and keywords of the problem set;
determining a representation of a sentence hierarchy of the problem set;
determining a representation of a term hierarchy of the problem set based on the keywords;
and determining the topic representation of the topic based on the representation of the sentence hierarchy and the representation of the word hierarchy.
6. The method of claim 5, wherein determining a sentence-level and word-level representation of a problem set and determining a topic representation of a topic based thereon comprises:
processing the problem set under the theme and the keywords of the problem set by using a theme representation model to obtain the theme representation of the theme output by the theme representation model;
the topic representation model is provided with a representation for determining sentence level of the question set; determining a representation of a term hierarchy of the problem set based on the keywords; determining a capability of topic characterization of the topic based on the characterization of the sentence hierarchy and the characterization of the word hierarchy.
7. The method according to claim 6, wherein the processing the question set and the keywords of the question set under the topic by using the topic representation model to obtain the topic representation of the topic output by the topic representation model comprises:
screening the representations of the problems in the problem set under the theme based on an attention mechanism by utilizing a sentence level processing module of a theme representation model to obtain the representations of the sentence levels of the problem set;
determining the representation of the word hierarchy of the problem set based on the keywords of the problem set and the weight of each keyword serving as the contribution information of the theme by using a word hierarchy processing module of a theme representation model;
and utilizing a fusion module of a topic representation model to fuse the representation of the sentence level of the question set and the representation of the word level of the question set to obtain the topic representation of the topic.
8. The method according to claim 6, wherein a topic characterization model fusion module is used to fuse the sentence-level characterization of the problem set and the word-level characterization of the problem set, and the topic characterization of the topic is obtained as follows:
converting the representation of the word hierarchy of the question set into vector representation of the same dimension based on the dimension of the representation of the sentence hierarchy of the question set;
and splicing the converted word level representation and the sentence level representation of the question set to obtain the topic representation of the topic.
9. A human-computer interaction device, comprising:
an input question acquisition unit for acquiring a user input question;
the target theme determining unit is used for determining a target theme matched with the user input question based on the configured theme representation corresponding to each theme; the topic representation corresponding to any topic is determined by the representation of sentence level and the representation of word level of the topic set under the topic; the representation of the sentence level of the question set under the theme is obtained by screening the representation of the questions in the question set under the theme based on an attention mechanism, and the representation of the word level of the question set under the theme is determined based on the keywords of the question set and the weight of the contribution information of each keyword as the theme;
and the question response unit is used for responding to the user input question based on the answer of the target theme.
10. The apparatus of claim 9, wherein the target subject determination unit comprises:
the similarity determining unit is used for determining the similarity of the user input question and each topic based on the configured topic representation corresponding to each topic;
and the similarity using unit is used for determining a target theme matched with the user input question from all themes according to the similarity between the user input question and each theme.
11. The apparatus of claim 9, further comprising:
the first updating unit is used for acquiring a newly-added problem of the theme to be updated when the theme representation corresponding to the theme to be updated is determined to be required to be updated;
the second updating unit is used for storing the newly increased problems into a problem set under a theme to be updated to obtain a new problem set under the theme to be updated;
and the third updating unit is used for determining the representation of the sentence level and the representation of the word level of the new question set and determining the updated topic representation of the topic to be updated according to the representation of the sentence level and the representation of the word level.
12. A human-computer interaction device, comprising: a memory and a processor;
the memory is used for storing programs;
the processor, which is used for executing the program, realizes the steps of the human-computer interaction method according to any one of claims 1-8.
13. A readable storage medium, having stored thereon a computer program, characterized in that the computer program, when being executed by a processor, carries out the steps of the human-computer interaction method according to any one of claims 1 to 8.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201911392035.3A CN111191034B (en) | 2019-12-30 | 2019-12-30 | Human-computer interaction method, related device and readable storage medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201911392035.3A CN111191034B (en) | 2019-12-30 | 2019-12-30 | Human-computer interaction method, related device and readable storage medium |
Publications (2)
Publication Number | Publication Date |
---|---|
CN111191034A CN111191034A (en) | 2020-05-22 |
CN111191034B true CN111191034B (en) | 2023-01-17 |
Family
ID=70711087
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201911392035.3A Active CN111191034B (en) | 2019-12-30 | 2019-12-30 | Human-computer interaction method, related device and readable storage medium |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111191034B (en) |
Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103810218A (en) * | 2012-11-14 | 2014-05-21 | 北京百度网讯科技有限公司 | Problem cluster-based automatic asking and answering method and device |
CN104899188A (en) * | 2015-03-11 | 2015-09-09 | 浙江大学 | Problem similarity calculation method based on subjects and focuses of problems |
WO2016173326A1 (en) * | 2015-04-30 | 2016-11-03 | 北京贝虎机器人技术有限公司 | Subject based interaction system and method |
CN106294341A (en) * | 2015-05-12 | 2017-01-04 | 阿里巴巴集团控股有限公司 | A kind of Intelligent Answer System and theme method of discrimination thereof and device |
CN106815311A (en) * | 2016-12-21 | 2017-06-09 | 杭州朗和科技有限公司 | A kind of problem matching process and device |
CN109145099A (en) * | 2018-08-17 | 2019-01-04 | 百度在线网络技术(北京)有限公司 | Answering method and device based on artificial intelligence |
CN109710744A (en) * | 2018-12-28 | 2019-05-03 | 合肥讯飞数码科技有限公司 | A kind of data matching method, device, equipment and storage medium |
CN110083692A (en) * | 2019-04-22 | 2019-08-02 | 齐鲁工业大学 | A kind of the text interaction matching process and device of finance knowledge question |
CN110209790A (en) * | 2019-06-06 | 2019-09-06 | 阿里巴巴集团控股有限公司 | Question and answer matching process and device |
-
2019
- 2019-12-30 CN CN201911392035.3A patent/CN111191034B/en active Active
Patent Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103810218A (en) * | 2012-11-14 | 2014-05-21 | 北京百度网讯科技有限公司 | Problem cluster-based automatic asking and answering method and device |
CN104899188A (en) * | 2015-03-11 | 2015-09-09 | 浙江大学 | Problem similarity calculation method based on subjects and focuses of problems |
WO2016173326A1 (en) * | 2015-04-30 | 2016-11-03 | 北京贝虎机器人技术有限公司 | Subject based interaction system and method |
CN106294341A (en) * | 2015-05-12 | 2017-01-04 | 阿里巴巴集团控股有限公司 | A kind of Intelligent Answer System and theme method of discrimination thereof and device |
CN106815311A (en) * | 2016-12-21 | 2017-06-09 | 杭州朗和科技有限公司 | A kind of problem matching process and device |
CN109145099A (en) * | 2018-08-17 | 2019-01-04 | 百度在线网络技术(北京)有限公司 | Answering method and device based on artificial intelligence |
CN109710744A (en) * | 2018-12-28 | 2019-05-03 | 合肥讯飞数码科技有限公司 | A kind of data matching method, device, equipment and storage medium |
CN110083692A (en) * | 2019-04-22 | 2019-08-02 | 齐鲁工业大学 | A kind of the text interaction matching process and device of finance knowledge question |
CN110209790A (en) * | 2019-06-06 | 2019-09-06 | 阿里巴巴集团控股有限公司 | Question and answer matching process and device |
Non-Patent Citations (3)
Title |
---|
"Attention-based Hierarchical LSTM Model for Document Sentiment Classification";Bo Wang;《Materials Science and Engineering》;20181231;第1-6页 * |
"Attention-based LSTM network for cross-lingual sentiment classification";Zhou Xinjie;《Proc of Conference on Empirical Methods in Natural Language Processing》;20161231;第247-256页 * |
远程教育中自动问答***相关技术探究;杨亚菲;《中国教育信息化》;20181010(第20期);第98-100页 * |
Also Published As
Publication number | Publication date |
---|---|
CN111191034A (en) | 2020-05-22 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108959246B (en) | Answer selection method and device based on improved attention mechanism and electronic equipment | |
CN108829822B (en) | Media content recommendation method and device, storage medium and electronic device | |
CN109829104B (en) | Semantic similarity based pseudo-correlation feedback model information retrieval method and system | |
RU2678716C1 (en) | Use of autoencoders for learning text classifiers in natural language | |
CN108846077B (en) | Semantic matching method, device, medium and electronic equipment for question and answer text | |
CN109948149B (en) | Text classification method and device | |
CN103870001B (en) | A kind of method and electronic device for generating candidates of input method | |
AU2016256764A1 (en) | Semantic natural language vector space for image captioning | |
US20230386238A1 (en) | Data processing method and apparatus, computer device, and storage medium | |
CN111125520B (en) | Event line extraction method based on deep clustering model for news text | |
CN109447273A (en) | Model training method, advertisement recommended method, relevant apparatus, equipment and medium | |
EP3377983A1 (en) | Generating feature embeddings from a co-occurrence matrix | |
CN116601626A (en) | Personal knowledge graph construction method and device and related equipment | |
CN113297410A (en) | Image retrieval method and device, computer equipment and storage medium | |
JP6973255B2 (en) | Word vector changing device, method, and program | |
CN112650842A (en) | Human-computer interaction based customer service robot intention recognition method and related equipment | |
CN116186326A (en) | Video recommendation method, model training method, electronic device and storage medium | |
CN117390450A (en) | Large language model training method, device and related equipment | |
CN114490926A (en) | Method and device for determining similar problems, storage medium and terminal | |
CN112527967A (en) | Text matching method, device, terminal and storage medium | |
CN111191034B (en) | Human-computer interaction method, related device and readable storage medium | |
CN111858860B (en) | Search information processing method and system, server and computer readable medium | |
CN111897943A (en) | Session record searching method and device, electronic equipment and storage medium | |
CN116932735A (en) | Text comparison method, device, medium and equipment | |
US20230368003A1 (en) | Adaptive sparse attention pattern |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |