CN111708869A

CN111708869A - Man-machine conversation processing method and device

Info

Publication number: CN111708869A
Application number: CN202010397838.4A
Authority: CN
Inventors: 张茂洪
Original assignee: Beijing Mininglamp Software System Co ltd
Current assignee: Beijing Mininglamp Software System Co ltd
Priority date: 2020-05-12
Filing date: 2020-05-12
Publication date: 2020-09-25
Anticipated expiration: 2040-05-12
Also published as: CN111708869B

Abstract

The embodiment of the invention discloses a method and a device for processing man-machine conversation, wherein the method comprises the following steps: receiving a problem submitted by a user, and identifying the type of the problem according to a preset intention identification binary model; when the type of the question is related to the field, judging whether the question is a map question-answer or an FAQ question-answer; when the question is a map question-answer, inquiring based on a knowledge map database of a vertical field to which the question belongs; when the query is successful, outputting a query result to the user; wherein the intent recognition bi-classification model is a bi-classification model for recognizing whether the type of the input question is domain-dependent or non-domain-dependent. Therefore, the knowledge graph is introduced into the man-machine conversation, and the accuracy and the working efficiency of the man-machine conversation are improved.

Description

Man-machine conversation processing method and device

Technical Field

The embodiment of the invention relates to Artificial Intelligence (AI) technology, in particular to a method and a device for processing man-machine conversation.

Background

In recent years, with the continuous expansion of customer service demands, the improvement of service complexity and the continuous increase of user quantity, customer service departments bear huge pressure, and the traditional manual customer service cannot completely meet the customer service demands of enterprises. Statistical data shows that for enterprises, the cost of intelligent customer service is only 10% of that of manual work, the service efficiency can be improved by 86% after the intelligent customer service is used, the customer satisfaction can reach 96%, and the order conversion rate is improved by about 20%. Artificial Intelligence (AI) is becoming the mainstream trend of information processing instead of the conventional manual operation. Natural Language Processing (NLP) is a key field of artificial intelligence and one of the most difficult fields. NLP is divided into three main directions of Natural Language identification, Natural Language Understanding, and Natural Language generation, wherein Natural Language Understanding (NLU) is particularly difficult, but the demand is also broad.

How to deepen understanding of natural language problems in the process of man-machine conversation so as to improve the processing efficiency of intelligent customer service is a problem which needs to be solved urgently in the prior art.

Disclosure of Invention

In view of this, an embodiment of the present invention provides a method for processing a human-computer conversation, including:

receiving a problem submitted by a user, and identifying the type of the problem according to a preset intention identification binary model;

when the type of the question is related to the field, judging whether the question is a map question-answer or an FAQ question-answer;

when the question is a map question-answer, inquiring based on a knowledge map database of a vertical field to which the question belongs;

when the query is successful, outputting a query result to the user;

wherein the intent recognition bi-classification model is a bi-classification model for recognizing whether the type of the input question is domain-dependent or non-domain-dependent.

An embodiment of the present invention further provides an electronic device, including:

the identification unit is used for receiving the questions submitted by the user and identifying the type of the questions according to a preset intention identification binary model;

a judging unit configured to judge whether the question is a map question-answer or an FAQ question-answer when the type of the question is domain-related;

the query single member is set to query based on a knowledge map database of a vertical field to which the question belongs when the question is a map question and answer;

the output unit is set to output the query result to the user when the query is successful;

An embodiment of the present invention further provides an electronic device, including: the device comprises a memory, a processor and a computer program which is stored on the memory and can run on the processor, wherein the computer program realizes the processing method of the man-machine conversation when being executed by the processor.

The embodiment of the invention also provides a computer readable storage medium, wherein an information processing program is stored on the computer readable storage medium, and the information processing program is used for realizing the man-machine conversation processing method when being executed by a processor.

According to the technical scheme provided by the embodiment of the invention, the knowledge graph is introduced into the man-machine conversation, so that the accuracy and the working efficiency of the man-machine conversation are improved.

Additional features and advantages of the application will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by the practice of the application. Other advantages of the present application may be realized and attained by the instrumentalities and combinations particularly pointed out in the specification and the drawings.

Drawings

The accompanying drawings are included to provide an understanding of the present disclosure and are incorporated in and constitute a part of this specification, illustrate embodiments of the disclosure and together with the examples serve to explain the principles of the disclosure and not to limit the disclosure.

Fig. 1 is a schematic flowchart illustrating a processing method of a man-machine conversation according to an embodiment of the present invention;

fig. 2 is a flowchart illustrating a processing method of a man-machine conversation according to another embodiment of the present invention;

FIG. 3 is a flowchart illustrating a processing method of a human-machine interaction according to another embodiment of the present invention;

FIG. 4 is a query graph diagram of an example query in an embodiment of the invention;

fig. 5 is a schematic structural diagram of an electronic device according to an embodiment of the invention.

Detailed Description

The present application describes embodiments, but the description is illustrative rather than limiting and it will be apparent to those of ordinary skill in the art that many more embodiments and implementations are possible within the scope of the embodiments described herein. Although many possible combinations of features are shown in the drawings and discussed in the detailed description, many other combinations of the disclosed features are possible. Any feature or element of any embodiment may be used in combination with or instead of any other feature or element in any other embodiment, unless expressly limited otherwise.

The present application includes and contemplates combinations of features and elements known to those of ordinary skill in the art. The embodiments, features and elements disclosed in this application may also be combined with any conventional features or elements to form a unique inventive concept as defined by the claims. Any feature or element of any embodiment may also be combined with features or elements from other inventive aspects to form yet another unique inventive aspect, as defined by the claims. Thus, it should be understood that any of the features shown and/or discussed in this application may be implemented alone or in any suitable combination. Accordingly, the embodiments are not limited except as by the appended claims and their equivalents. Furthermore, various modifications and changes may be made within the scope of the appended claims.

Further, in describing representative embodiments, the specification may have presented the method and/or process as a particular sequence of steps. However, to the extent that the method or process does not rely on the particular order of steps set forth herein, the method or process should not be limited to the particular sequence of steps described. Other orders of steps are possible as will be understood by those of ordinary skill in the art. Therefore, the particular order of the steps set forth in the specification should not be construed as limitations on the claims. Further, the claims directed to the method and/or process should not be limited to the performance of their steps in the order written, and one skilled in the art can readily appreciate that the sequences may be varied and still remain within the spirit and scope of the embodiments of the present application.

Fig. 1 is a schematic flowchart of a processing method of a man-machine conversation according to an embodiment of the present invention, as shown in fig. 1, the method includes:

step 101, receiving a question submitted by a user, and identifying the type of the question according to a preset intention identification binary model;

step 102, when the type of the question is related to the field, judging whether the question is a map question-answer or an FAQ question-answer;

103, when the question is a map question and answer, inquiring a knowledge map database of a vertical field to which the question belongs;

104, when the query is successful, outputting a query result to a user;

In one example, the determining whether the question is a profile question-answer or an FAQ question-answer includes:

identifying a vertical field to which the problem belongs and a subordinate classification to which the vertical field belongs according to a preset intention identification multi-classification model;

determining a corresponding word bank according to the subordinate classification of the vertical field to which the problem belongs, and extracting an entity and an entity relation corresponding to the problem by using an entity extraction algorithm based on the word bank;

carrying out sentence pattern classification based on the entity and entity relation corresponding to the question, and judging whether the question is a map question-answer or an FAQ question-answer based on the sentence pattern classification;

wherein the intent recognition multi-classification model is a multi-classification model for identifying a vertical domain to which a problem belongs and a subordinate classification to which the vertical domain belongs.

In one example, the querying based on the vertical domain knowledge graph database to which the problem belongs includes:

checking whether the preset slot information of the conversation type is complete; when the slot information is incomplete, the user is asked for the slot information, and the slot information is filled according to the supplement of the user, and the rest is done until the slot information is complete; and when the slot information is complete, querying on the basis of a knowledge map database of the vertical field to which the problem belongs by using a corresponding application program interface API.

In one example, the method further comprises:

when the query is unsuccessful, carrying out similarity matching on the question and a preset common question answer (FAQ), and when the matching result is greater than or equal to a preset similarity threshold value, outputting the matching result to a user; and when the matching result is smaller than a preset similarity threshold value, the problem is manually processed.

In one example, the method further comprises:

when the question is an FAQ question and answer, carrying out similarity matching on the question and a pre-configured FAQ by using a similarity matching algorithm, and when the matching result is greater than or equal to a preset similarity threshold value, outputting the matching result to a user; and when the matching result is smaller than a preset similarity threshold value, the problem is manually processed.

In one example, after the identifying the type of the problem according to the pre-set intent identification binary model, the method further comprises:

and when the type of the question is non-domain related and is chatting, inputting the question into a pre-trained deep learning model, outputting an answer, and returning the answer to the user.

and when the type of the question is non-domain related and is self-defined chatting, inputting the question into a preset self-defined rule base, outputting an answer, and returning the answer to the user.

According to the technical scheme provided by the embodiment of the invention, the knowledge graph is introduced into the man-machine conversation, so that the understanding of the natural language problem can be deepened, and the accuracy and the working efficiency of the man-machine conversation are improved.

Fig. 2 is a schematic flow chart of a processing method of a man-machine conversation according to another embodiment of the present invention, as shown in fig. 2, the method includes:

step 201, receiving a question submitted by a user, and identifying the type of the question according to a preset intention identification binary model;

wherein the type includes two categories of domain-related and non-domain-related. The intention recognition binary model is a binary model, the question is input, and the type of the question is output. In this example, the intent identifies a binary classification model, outputting the type of the question as either domain-related or non-domain-related (e.g., chatting) for both domain-related and non-domain-related classifications, i.e., input questions. Domain-related means that the problem belongs to the vertical domain, e.g. "latest price of iphone apple" relates to the field of mobile phones and thus belongs to the domain-related and vertical domain as mobile phones. The vertical field refers to the vertical field of the industry, such as beauty makeup, mobile phones, automobiles, insurance, smart home and the like. Each vertical domain has its own unique terminology and knowledge, and a question can be identified as belonging to the vertical domain according to the terminology and knowledge. The category of the vertical domain can be set by itself or adopt the default vertical domain classification in the industry. For each specific domain, other domains and chats are non-domain related questions, for example, for question and answer in the makeup art, then 3C, car, chats are non-domain related questions.

In one example, the intention recognition binary model may be any existing intention recognition model, such as a rule-based algorithm, an intention recognition model generated based on training of a deep learning algorithm (e.g., CNN (Convolutional neural networks), LSTM (Long Short-Term Memory networks), RCNN (Regions with CNN function), C-LSTM (Convolutional LSTM, Convolutional Long Short-Term Memory networks), FastText (fast text classification algorithm), etc.).

In another example, the two-classification model for intention recognition may also be a model obtained by improving an existing intention recognition model according to actual needs, for example, a model obtained by refining based on BERT (Bidirectional Encoder replication, i.e. Encoder of Bidirectional Transformer), wherein the training process includes preprocessing and normalizing training data, including word segmentation, word de-stop, word embedding, etc., then labeling the training data based on a defined classification rule, expanding the labeled data by using a text similarity algorithm, reducing the workload of manual labeling, and forming a training data set; finally, based on BERT, finetune training model tuning is carried out, and the label means that one problem is domain-related or domain-irrelevant.

In one example, before identifying the type of the problem according to the pre-set intent identification binary model, the method further comprises:

pre-processing and standardizing the problem;

inputting the pre-processed and normalized questions into the intent recognition bi-classification model.

The preprocessing comprises the operations of removing invalid texts (such as removing url, removing @ and harshtag, removing emoticons, removing special symbols and the like), unifying case and case, unifying frequently abbreviated characters, removing blank spaces in texts, removing water and removing noise and the like for the problem.

The standardization refers to operations such as word segmentation, multiple filtering stop words, part of speech tagging, synonymy replacement and the like for the question.

When the type of the question is domain-related, executing step 202, and when the type of the question is non-domain-related, executing step 209 or 210:

step 202, identifying a vertical field to which the problem belongs and a subordinate classification to which the vertical field belongs according to an intention identification multi-classification model;

the intention recognition multi-classification model is a classification model, a problem is input, a vertical field to which the problem belongs and a subordinate classification to which the vertical field belongs are output. The multi-classification means multi-classification in which after two classifications are filtered for the first time, the two classifications enter a vertical field and need to be subjected to one-time intention, for example, in the field of mobile phones, problems related to the field of mobile phones are known, and subordinate classifications such as mobile phone price, mobile phone appearance, mobile phone configuration, charging, battery and the like need to be identified. The subordinate classifications of each vertical domain may be set by themselves or default in the industry.

In one example, the intent recognition multi-classification model may be any existing intent recognition multi-classification model, such as a rule-based algorithm, an intent recognition multi-classification model generated based on training of a deep learning algorithm (e.g., CNN (Convolutional neural networks), LSTM (Long Short-Term Memory networks), RCNN (Regions with CNN function), C-LSTM (Convolutional LSTM, Convolutional Long Short-Term Memory networks), FastText (fast text classification algorithm), etc.).

In another example, the intention recognition multi-classification model may also be a model obtained by improving an existing intention recognition model according to actual needs, for example, a model obtained by performing finetune based on BERT (Bidirectional Encoder retrieval from transformers), and the like.

Step 203, determining a corresponding word bank according to the subordinate classification of the vertical field to which the problem belongs, and extracting an entity and an entity relation corresponding to the problem by using an entity extraction algorithm based on the word bank;

the word stock comprises an industry word stock, a synonym word stock and the like. After the vertical field and the subordinate classification are determined, which word banks are needed can be accurately positioned, because different vertical fields and different subordinate classifications correspond to different word banks respectively. Even though the same word may have different meanings and synonyms in different vertical domains and in different sub-domains. For example, the boundaries of proper nouns are complex, the lengths of the proper nouns are not fixed, the parts of speech of the proper nouns are rich, and the boundary judgment is easy to be wrong; for example, the 'king fat and donkey meat fever' cannot be identified as a person name but is a part of an organization name, and if the judgment is wrong, the semantics cannot be normally understood. As another example, semantics are context dependent, such as a user input "find zhang san of business today," where "business" relates to the user's geographic location, which may refer to shanghai university of transportation or beijing university of transportation, etc.; in addition, "this year" represents the current year, relative to the current time of the system; therefore, to correctly understand the semantics, the current context must be perceived. For another example, the semantics depends on the knowledge domain, for example, if the user inputs "search for Xiaoming's Niger", where "Niger" implies gender information, the gender information does not need to be recommended when the entity recommends; if the knowledge that the user is not in a state that the user is only a woman in a pallium is not available, accurate recommendation cannot be achieved. Therefore, in order to improve the accuracy of entity extraction, it is necessary to find the correct lexicon.

In an example, the entity extraction algorithm may be any one of existing entity extraction algorithms, and processes include entity identification, relationship extraction, knowledge disambiguation, and so on, which are not described herein again in the related art.

Step 204, carrying out sentence pattern classification based on the entity and entity relation corresponding to the question, and judging whether the question is a map question-answer or an FAQ question-answer based on the sentence pattern classification;

in the field of natural language processing, an entity is simply understood as a noun, such as a name of a person, a name of an organization, a name of a place, and all other entities identified by names, and a broader entity includes numbers, dates, currencies, addresses, and the like. An entity may have multiple meanings, for example, the meaning of the same entity may be different in different contexts. For a human being, the specific meanings represented by the entities can be intuitively judged, but for a machine, the specific meanings represented by each entity can be identified by means of natural language processing technology, and different entities can be distinguished. Whereas an entity relationship describes how two or more entities are related to each other, an association can be (roughly) considered a verb, such as: a proprietary association between a company and a computer, a regulatory association between employees and departments, a performance association between actors and songs, a certification association between mathematicians and theorems, etc.

In one example, the sentence classification based on the entity and entity relationship corresponding to the question includes:

and carrying out sentence pattern classification on the problem based on a pre-trained sentence pattern classification model.

In which a sentence must be organized according to a certain pattern, called a sentence pattern. Sentence patterns can be classified into a cardinal relationship, a structure in a form, a parallel relationship, a moving guest relationship, a mediating guest relationship, etc. In an example, a sentence pattern classification model may be trained in advance for a training sample, the sentence pattern classification model takes the entity and entity relationship of the extracted question as input, the sentence pattern of the question as output, and the training sample is the entity and entity relationship corresponding to a plurality of questions marked with sentence patterns.

In one example, determining whether the question is an atlas question-answer or an FAQ question-answer based on sentence pattern classification includes:

and judging whether the question is a map question-answer or an FAQ question-answer based on a pre-trained question two-classification model.

The question recognition two-classification model can be trained in advance aiming at a training sample, the question recognition two-classification model takes the real sentence pattern of the extracted question as input, the question as an atlas question-answer or FAQ question-answer as output, and the training sample is a plurality of sentence patterns corresponding to the questions marked with the atlas question or FAQ question. For example, "how much money the Shenxian water is" is the map question and answer, and "how do I are allergic with Shenxian water" is the FAQ question and answer.

When the question is an FAQ question, step 205 is executed, and when the question is an atlas question, step 206 is executed:

step 205, performing similarity matching on the question and a preset frequently asked question answer (FAQ) by using a similarity matching algorithm, and outputting a matching result to a user when the matching result is greater than or equal to a preset similarity threshold; when the matching result is smaller than a preset similarity threshold value, the problem is manually processed;

the preset FAQ (Frequently assigned Questions) may be an existing FAQ library or system in the vertical field, or an FAQ library or system developed according to data customization of customers in different industries.

The similarity matching algorithm is any one of the existing similarity algorithms, such as a distance algorithm, and similarity is judged by calculating an average value of all word embedding in a sentence, using word embedding between two texts, and measuring the shortest distance required by a word in one text to move to another text word in a semantic space. Also for example, a BERT-based neural network optimization model, i.e., BERT + finetune's method, etc.

Step 206, checking whether the slot information preset by the question type is complete; when the slot information is incomplete, the user is asked for the slot information, and the slot information is filled according to the supplement of the user, and the rest is done until the slot information is complete; when the slot information is complete, a corresponding API is used for inquiring based on a knowledge map database of the vertical field to which the problem belongs;

for example, for a database supporting gremlin query, when slot information is complete, a gremlin query template is generated to instantiate a gremlin, and then query is performed based on a knowledge graph database in the vertical field to which the problem belongs by using a corresponding API. For another example, for a database supporting other query modes, when the slot information is complete, the corresponding query mode and the API are used for querying based on a knowledge map database in the vertical field to which the problem belongs.

Wherein, different knowledge map databases respectively provide corresponding APIs (Application programming interfaces), and the corresponding APIs can be used to access and query the corresponding knowledge map databases.

Knowledge graph is a semantic network in nature. Its nodes represent entities (entries) or concepts (concepts), and edges represent various semantic relationships between entities/concepts. The knowledge map database can be an existing knowledge map database or a knowledge map database customized according to business needs.

When the query is successful, step 207 is executed, and when the query is unsuccessful, step 208 is executed:

step 207, outputting the query result to the user;

step 208, performing similarity matching on the question and a preset FAQ, and outputting a matching result to a user when the matching result is greater than or equal to a preset similarity threshold; when the matching result is smaller than a preset similarity threshold value, the problem is manually processed;

the system can meet the requirements of multiple rounds of conversations, the multiple rounds of conversations need a filling process of slot information, the slot information needs to be specifically designed according to subordinate classifications corresponding to the multiple previous classifications, the slot information can be different according to different vertical fields, and different necessary slot filling information and selected slot filling information need to be provided for intentions of different subordinate classifications. For example, asking the price of the mobile phone, the necessary filling slots are the brand, model and configuration of the mobile phone, and the selected filling slots are provided with colors, system kernels and the like. The problem is a factual problem query based on a product attribute, and the query can be directly carried out through a knowledge graph.

The Knowledge map (Knowledge Graph) is a series of different graphs displaying the relationship between the Knowledge development process and the structure, describes Knowledge resources and carriers thereof by using a visualization technology, and excavates, analyzes, constructs, draws and displays Knowledge and the mutual relation between the Knowledge resources and the Knowledge resources. Specifically, the knowledge graph is a modern theory which achieves the purpose of multi-discipline fusion by combining theories and methods of applying subjects such as mathematics, graphics, information visualization technology, information science and the like with methods such as metrology introduction analysis, co-occurrence analysis and the like and utilizing a visualized graph to vividly display the core structure, development history, frontier field and overall knowledge framework of the subjects. The method displays the complex knowledge field through data mining, information processing, knowledge measurement and graph drawing, reveals the dynamic development rule of the knowledge field, and provides a practical and valuable reference for subject research. In this example, a knowledge graph database may be pre-constructed in advance for each vertical domain.

Step 209, when the type of the question is non-domain related and is a system chat, inputting the question into a pre-trained deep learning model, outputting an answer, and returning the answer to the user;

the deep learning model is obtained by performing chatting training according to a corpus based on a deep learning algorithm, and takes a system chatting question as input and an answer of the chatting question as output. For example, GPT2 (genetic Pre-Training) can be used for chatting Training, 100 ten thousand corpus can be prepared, and model Training and tuning can be performed.

Here, the system chat means chat in the general sense.

And step 210, when the type of the question is non-domain related and is user-defined chatting, inputting the question into a preset user-defined rule base, outputting an answer, and returning the answer to the user.

Wherein, the user-defined chat means making some customized small talk such as call and bye according to the requirement of the client, and other chat means algorithm generated answer without fixed rule and template.

Fig. 3 is a flowchart illustrating a processing method of a man-machine conversation according to another embodiment of the present invention.

As shown in fig. 3, the method includes:

firstly, receiving a user input problem, and preprocessing and standardizing the problem;

wherein, the preprocessing comprises case and case unification, simplified word propagation, water removal and noise removal and the like.

The standardization comprises word segmentation, filtering stop words, part of speech tagging and synonymy replacement.

Secondly, performing secondary classification on the processed and standardized problems;

where binary classification refers to identifying whether the type of problem is domain-related or non-domain-related.

In this embodiment, non-domain related examples are described with reference to chat sessions. The two-classification can be performed by identifying the type of the problem according to a preset intention identification two-classification model, or training a classifier to identify the type of the problem in advance.

Thirdly, distinguishing a system chatting mode or a user-defined chatting mode aiming at the chatting conversation, and outputting an answer by using a deep learning model aiming at the system chatting; outputting a response by using a custom rule base aiming at the custom chatting;

again, for domain-related, intent-multi-classification can be made;

wherein the intent multi-classification is to identify a vertical domain and a subordinate classification corresponding to the question. The intention multi-classification may be performed by identifying the vertical domain to which the problem belongs and the subordinate classification according to a preset intention recognition multi-classification model, or by training the vertical domain to which the problem belongs and the subordinate classification in advance.

Thirdly, extracting the entity and the entity relation of the problem by using an entity extraction algorithm in the vertical field so as to classify the sentence patterns;

wherein, the sentence classification refers to the sentence judgment of the question.

Thirdly, after the sentence patterns are classified, whether the question is an atlas question answer or an FAQ question answer is judged based on the sentence patterns of the question, if the question is the FAQ question answer, FAQ similarity matching is directly carried out on the question, if the matching result is larger than or equal to a preset threshold value, the matching result, namely output is output, and if the matching result is smaller than the preset threshold value, manual work is carried out;

wherein, if the manual work is changed, the problem can be added into a training expectation to optimize a training model. The training model is referred to as FAQ.

And thirdly, after the sentence pattern is classified, judging whether the question is an atlas question-answer or an FAQ question-answer based on the sentence pattern of the question, if the question is the atlas question-answer, judging whether the information is complete, if the information is incomplete, asking the user in a reverse way, supplementing a filling groove, repeating the steps until the information is complete, generating a gremlin query template for carrying out gremlin instantiation when the information is complete, querying an API (application program interface) corresponding to the vertical field in a plurality of APIs (application programming interfaces), utilizing the queried API to query in a knowledge graph database corresponding to the vertical field, if the query is successful, outputting a query result, namely output, if the query is not successful, carrying out FAQ similarity matching on the question, if the matching result is more than or equal to a preset threshold, outputting a matching result, namely output, and if the matching result is less than the preset threshold, manually switching.

In this embodiment, the knowledge map database corresponding to the vertical domain is exemplified by a cellular map database. The cellular database supports Gremlin queries. Gremlin is a graph traversal language under the Apache ThinkerPop framework, is a functional data stream language, and can enable a user to express traversal or query of a complex attribute graph in a concise mode. Each Gremlin traversal consists of a series of steps (nesting may exist), each step performing an atomic operation on the data stream (datastream). Such as how many equals "equal" queries the query price is, and how many equals "max" queries the most expensive price is.

Based on the processing method of the man-machine conversation provided by the embodiment, a detailed example is listed below, for example:

first, a question is input: there are some domestic handsets that are more expensive than the iphone's most expensive handset;

then, two classifications are carried out: the problem is domain-dependent, and the vertical domain to which the problem belongs is a mobile phone and the subordinate classification is a price; through the analysis, the problem input by the user is not simple query and needs to be compared and inferred;

then, sentence pattern judgment: "multiple entity comparison level";

then, text structuring: slot { "brand": iphone "," price ": the most expensive handset price of iphone", "comparison": more expensive than … … "};

then, Gremlin queries: g.V (). has ("date on market ', neq (' unknown ')). as (" V "). has (" brand "," iphone "). values (" price "). max (). as (" p "). select (" V "). has (" cell phone place ', "cell phone at home"). has ("cell phone price", gt (__. select ("p")))) order (). by ("date on market ', desc). limit (6). reduce (). value map (" cell phone price "," cell phone model "). toList ();

and finally, outputting a result: [ [17000, "opporeno 10 times zoom edition" ], [13000, "oppor 17pro new year special edition" ], [12999, "oppor 17pro ] ].

The query graph of this example is shown in FIG. 4.

According to the technical scheme provided by the embodiment of the invention, the natural language is converted into the query sentence of the knowledge graph, the sentence vector and the deep learning neural network model are used, the sentence pattern template is not required to be relied on, and the problem identification type is wider and more accurate.

Fig. 5 is a schematic structural diagram of an electronic device according to an embodiment of the present invention, as shown in fig. 5, including:

the query unit is used for querying a knowledge map database of a vertical field to which the question belongs when the question is a map question and answer;

In one example, the judging unit is configured to

In an example, the query unit is further configured to perform similarity matching on the question and a pre-configured frequently asked question answer (FAQ) when the query is unsuccessful, and output a matching result to the user when the matching result is greater than or equal to a preset similarity threshold; and when the matching result is smaller than a preset similarity threshold value, the problem is manually processed.

In one example, the apparatus further comprises:

the matching unit is used for matching the similarity of the question and a preset FAQ by using a similarity matching algorithm when the question is an FAQ question and answer, and outputting a matching result to a user when the matching result is greater than or equal to a preset similarity threshold; and when the matching result is smaller than a preset similarity threshold value, the problem is manually processed.

In one example, the apparatus further comprises:

and the answering unit is used for inputting the question into a pre-trained deep learning model to output an answer and returning the answer to the user when the type of the question is non-domain related and is chatty after the type of the question is identified according to the pre-set intention identification binary model.

In one example, the apparatus further comprises:

and the answer unit is used for inputting the question into a preset custom rule base to output an answer and returning the answer to the user when the type of the question is non-domain related and is custom chatting after the type of the question is identified according to the preset intention identification binary model.

An embodiment of the present invention further provides an electronic device, including: the device comprises a memory, a processor and a computer program which is stored on the memory and can run on the processor, wherein when the computer program is executed by the processor, the computer program realizes any one of the man-machine conversation processing methods.

The embodiment of the invention also provides a computer-readable storage medium, wherein an information processing program is stored on the computer-readable storage medium, and when the information processing program is executed by a processor, the information processing program realizes the man-machine conversation processing method in any one of the above items.

The man-machine conversation processing scheme provided by the embodiment of the invention can be applied to a plurality of artificial intelligence fields of intelligent customer service problems, user intention identification and the like. For example, the implementation schemes of the question-answering robot before 2014 were mainly based on traditional methods such as semantic parsing and information extraction. Since 2015, an end-to-end system based on a deep learning technology appears, and compared with the traditional method, the whole system is improved. However, although the workload consumed by a large amount of manual feature construction is reduced by pure end-to-end, the pure end-to-end deep learning model is based on learning the existing question and answer data and cannot answer new questions which do not appear; and the deep learning model has no interpretable line and can not analyze and optimize the answers given by the machine. For example, new problems that have not occurred are: in the traditional model, there is "what is the price of p 30? ", the model can only answer this if you ask" how much the Hua mate20 price "might learn to give you a P30 price point answer inside the traditional model. But if you ask a completely new question, such as: when you get on the line, the new system needs to be changed into manual work. The nature of the semantics is associative. Only the data interconnection based on the semantics can exert the nonlinear effect of data integration, and the special semantics of the big data can be obtained. The scheme of the embodiment of the invention is based on the knowledge map to complete the function of the question-answering robot, realizes the structuralization of knowledge in the field, and can realize the conversion of the natural language of the user into the structuralized associated semantic relation, thereby realizing the preparation of answering the user question by simulating the natural language by the machine.

The technical scheme provided by the embodiment of the invention can help enterprises to improve operation efficiency, save cost and provide more convenience and value-added service for customers, simultaneously easily solve various customer problems, process customer query requests and reduce the need of manual interaction. The client can easily expand the business, create the personalized experience and keep the initiative, and the enterprise can actively provide the personalized service for countless users in a humanized mode at the same time. In many scenes, the service and convenience level provided by the chat robot running based on the message transfer platform exceeds the manual level, and the customer experience is effectively improved.

It will be understood by those of ordinary skill in the art that all or some of the steps of the methods, systems, functional modules/units in the devices disclosed above may be implemented as software, firmware, hardware, and suitable combinations thereof. In a hardware implementation, the division between functional modules/units mentioned in the above description does not necessarily correspond to the division of physical components; for example, one physical component may have multiple functions, or one function or step may be performed by several physical components in cooperation. Some or all of the components may be implemented as software executed by a processor, such as a digital signal processor or microprocessor, or as hardware, or as an integrated circuit, such as an application specific integrated circuit. Such software may be distributed on computer readable media, which may include computer storage media (or non-transitory media) and communication media (or transitory media). The term computer storage media includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data, as is well known to those of ordinary skill in the art. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, Digital Versatile Disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and which can accessed by a computer. In addition, communication media typically embodies computer readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transport mechanism and includes any information delivery media as known to those skilled in the art.

Claims

1. A man-machine conversation processing method comprises the following steps:

when the query is successful, outputting a query result to the user;

2. The process of claim 1, wherein said determining whether the question is a profile question-answer or an FAQ question-answer comprises:

3. The processing method according to claim 1, wherein the querying based on the knowledge graph database of the vertical domain to which the problem belongs comprises:

4. The process of claim 1, further comprising:

5. The process of claim 1, further comprising:

6. The processing method according to claim 1, wherein after identifying the type of the problem according to the pre-set intent identification binary model, the method further comprises:

7. The processing method according to claim 1, wherein after identifying the type of the problem according to the pre-set intent identification binary model, the method further comprises:

8. An electronic device, comprising:

9. An electronic device, comprising: memory, processor and computer program stored on the memory and executable on the processor, which computer program, when executed by the processor, implements a method of processing a human-machine dialog according to any of claims 1 to 7.

10. A computer-readable storage medium, characterized in that an information processing program is stored thereon, which when executed by a processor implements a processing method of a human-machine conversation as claimed in any one of claims 1 to 7.