WO2019153613A1 - Chat response method, electronic device and storage medium - Google Patents

Abstract

Provided is a chat response method, comprising: acquiring a session question; querying a candidate question set related to the session question in a question-answer knowledge base; calculating the text similarity between the session question and each candidate question; determining whether a question similar to the session question exists; if so, searching for and outputting an answer associated with the similar question; if not, querying a candidate answer set related to the session question in the question-answer knowledge base; calculating the topic similarity between the session question and each candidate answer; determining whether an answer similar to the session question exists; if so, outputting the similar answer; if not, establishing a sequence prediction model, inputting the session question into the sequence prediction model to generate a response answer, and outputting the response answer as a target answer. The described method may provide customers with accurate and response feedback regarding the session question, thus improving the quality of service.

Description

Chat response method, electronic device and storage medium

This application claims priority to Chinese Patent Application No. 201101135747.6, entitled "Chat Response Method, Electronic Device, and Storage Media", filed on February 9, 2018, the entire contents of which are incorporated herein by reference. In the application.

Technical field

The present application relates to the field of computer technologies, and in particular, to a chat response method, an electronic device, and a storage medium.

Background technique

With the development of technology, AI (Artificial Intelligence) is gradually changing our way of life. For example, smart question and answer is one of them. When the customer consults online via text or voice, the customer can be intelligently answered by the online intelligent customer service. Intelligent Q&A can effectively alleviate the waiting situation of customer service and improve service quality, so it has a very broad prospect.

However, even in a specific service area, such as financial, banking, securities, insurance and other vertical areas, the online consultation process will contain some pure chat content. At this time, if the chat session content input by the customer cannot respond to the customer quickly, accurately, and effectively, the service quality of the intelligent customer service will be reduced, and the humanized high quality experience cannot be brought to the customer.

Summary of the invention

In view of the above reasons, it is necessary to provide a chat response method, an electronic device and a storage medium, which can provide accurate and responsive feedback to customers for conversational problems, thereby improving service quality.

To achieve the above objective, the present application provides a chat response method, the method comprising: a pre-processing step of: acquiring a session question input by a client, pre-processing the session problem, and obtaining text feature information of the session problem, the text feature The information includes part of speech, location and part-of-speech attribution information of each term in the conversation question, the word class attribution includes belonging to the keyword or the named entity; and the first calculating step: constructing an inverted index for the question-and-answer knowledge base, the question and answer The knowledge base includes a plurality of questions pre-arranged and one or more answers associated with each question, and according to the text feature information, querying candidate questions related to the conversation problem from the question-and-answer knowledge base by means of inverted index query Collecting, and separately calculating text similarity between the conversation problem and each candidate question in the candidate question set; a question retrieval step: determining whether the session exists in the candidate question set according to the preset rule and the text similarity An approximation problem of the problem, if there is an approximation of the conversation problem in the set of candidate questions And searching for the associated answer of the approximate question in the question and answer knowledge base, and outputting the associated answer as the target answer of the conversation question; and second calculating step: if the session problem does not exist in the candidate question set Approximating the problem, according to the text feature information, querying a candidate answer set related to the conversation question from the question and answer knowledge base by means of an inverted index query, and calculating each of the conversation problem and the candidate answer set separately The topic similarity of the candidate answers; the answer retrieval step: determining, according to the preset rule and the topic similarity, whether there is an approximate answer of the conversation question in the candidate answer set, if the conversation problem exists in the candidate answer set An approximate answer, the approximate answer is output as the target answer of the conversation question; the answer prediction step: if the approximate answer of the conversation problem does not exist in the candidate answer set, the seq2seq model is used in the question and answer knowledge base Each question and answer is iteratively trained in encoding and decoding to construct a sequence prediction module Type, inputting the conversation question into the sequence prediction model to generate a strain answer, and outputting the strain answer as a target answer of the conversation question.

To achieve the above object, the present application further provides an electronic device including a memory and a processor, wherein the memory includes a chat response program, and the chat response program is executed by the processor to implement the following steps: a pre-processing step Obtaining a session problem input by the client, pre-processing the session problem, and obtaining text feature information of the session problem, where the text feature information includes part of speech, location, and word class attribution information of each term in the conversation question, The term class attribution includes attribution to a keyword or a named entity; a first calculation step: constructing an inverted index for the question and answer knowledge base, the question and answer knowledge base including a plurality of questions pre-arranged and one or more answers associated with each question, Determining, according to the text feature information, a candidate question set related to the session problem from a question and answer knowledge base by means of an inverted index query, and respectively calculating the session problem and each candidate question in the candidate question set Text similarity; question retrieval step: according to preset rules and the similarity of the text, Whether there is an approximation problem of the conversation problem in the set of candidate questions, if there is an approximation problem of the conversation problem in the candidate question set, the associated answer of the approximation question is searched in the question and answer knowledge base, and the associated answer is a target answer output as the conversation problem; a second calculation step: if there is no approximation problem of the conversation problem in the candidate question set, according to the text feature information, the question and answer knowledge is obtained by means of an inverted index query Querying, in the library, a set of candidate answers related to the conversation question, and separately calculating a topic similarity between the conversation problem and each candidate answer in the candidate answer set; an answer retrieval step: according to a preset rule and the topic is similar Degree, determining whether there is an approximate answer of the conversation question in the candidate answer set, if an approximate answer of the conversation question exists in the candidate answer set, outputting the approximate answer as a target answer of the conversation question; Prediction step: if there is no approximate answer to the conversation question in the candidate answer set, An iterative training of encoding and decoding each question and answer in the Q&A knowledge base by the seq2seq model, thereby constructing a sequence prediction model, inputting the conversation problem into the sequence prediction model to generate a strain answer, and using the strain answer as The target answer output of the conversation question.

In addition, in order to achieve the above object, the present application further provides a computer readable storage medium including a chat response program, when the chat response program is executed by a processor, implementing the chat response method as described above Any step.

The chat response method, the electronic device and the storage medium proposed by the present application, after acquiring the session problem and performing pre-processing, query the candidate question set related to the conversation problem from the question-and-answer knowledge base by means of the inverted index query, and respectively Calculating a text similarity between the conversation problem and each candidate question in the candidate question set, determining whether there is an approximation problem of the conversation problem in the candidate question set, and if so, searching for an association of the approximation problem in the Q&A knowledge base The answer is that the associated answer is output as the target answer of the conversation question. If there is no approximation problem of the conversation problem in the candidate question set, the query and the query are obtained from the question and answer knowledge base by means of an inverted index query. Determining a set of candidate answers related to the conversation problem, and separately calculating a topic similarity of the conversation question and each candidate answer in the candidate answer set, and determining whether an approximate answer of the conversation question exists in the candidate answer set, and if so, Outputting the approximate answer as a target answer to the conversation question, if the candidate If there is no approximate answer to the conversation problem in the answer set, the iterative training of encoding and decoding each question and answer in the question and answer knowledge base is performed by the seq2seq model, thereby constructing a sequence prediction model, and inputting the conversation problem The sequence prediction model generates a strain answer, and outputs the strain answer as a target answer of the conversation question, and can provide accurate and responsive feedback to the client for the conversation problem, thereby improving service quality.

DRAWINGS

1 is a schematic diagram of an operating environment of a preferred embodiment of an electronic device of the present application;

2 is a schematic diagram of interaction between an electronic device and a client according to a preferred embodiment of the present application;

3 is a flow chart of a preferred embodiment of a chat response method of the present application;

4 is a program block diagram of the chat response program of FIG. 1.

The implementation, functional features and advantages of the present application will be further described with reference to the accompanying drawings.

Detailed ways

The principles and spirit of the present application are described below with reference to a number of specific embodiments. It is understood that the specific embodiments described herein are merely illustrative of the application and are not intended to be limiting.

Those skilled in the art will appreciate that embodiments of the present application can be implemented as a method, apparatus, device, system, or computer program product. Accordingly, the application can be embodied in a complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

According to an embodiment of the present application, a chat response method, an electronic device, and a storage medium are proposed.

1 is a schematic diagram of an operating environment of a preferred embodiment of an electronic device of the present application.

The electronic device 1 may be a terminal device having a storage and computing function such as a server, a portable computer, or a desktop computer.

The electronic device 1 includes a memory 11, a processor 12, a network interface 13, and a communication bus 14. The network interface 13 can optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The communication bus 14 is used to implement connection communication between the above components.

The memory 11 includes at least one type of readable storage medium. The at least one type of readable storage medium may be a non-volatile storage medium such as a flash memory, a hard disk, a multimedia card, a card type memory, or the like. In some embodiments, the readable storage medium may be an internal storage unit of the electronic device 1, such as a hard disk of the electronic device 1. In other embodiments, the readable storage medium may also be an external memory 11 of the electronic device 1, such as a plug-in hard disk equipped on the electronic device 1, a smart memory card (SMC). , Secure Digital (SD) card, Flash Card, etc.

In the present embodiment, the readable storage medium of the memory 11 is generally used to store the chat response program 10, the Q&A knowledge base 4, and the like installed in the electronic device 1. The memory 11 can also be used to temporarily store data that has been output or is about to be output.

The processor 12, in some embodiments, may be a Central Processing Unit (CPU), microprocessor or other data processing chip for running program code or processing data stored in the memory 11, such as executing a chat response program. 10 and so on.

1 shows only the electronic device 1 having the components 11-14 and the chat response program 10, but it should be understood that not all illustrated components may be implemented, and more or fewer components may be implemented instead.

Optionally, the electronic device 1 may further include a user interface, and the user interface may include an input unit such as a keyboard, a voice input device such as a microphone, a device with a voice recognition function, a voice output device such as an audio, a headphone, and the like. . Optionally, the user interface may also include a standard wired interface and a wireless interface.

Optionally, the electronic device 1 may further include a display, which may also be referred to as a display screen or a display unit. In some embodiments, it may be an LED display, a liquid crystal display, a touch liquid crystal display, and an Organic Light-Emitting Diode (OLED) display. The display is used to display information processed in the electronic device 1 and a user interface for displaying visualizations.

Optionally, the electronic device 1 further comprises a touch sensor. The area provided by the touch sensor for the user to perform a touch operation is referred to as a touch area. Further, the touch sensor described herein may be a resistive touch sensor, a capacitive touch sensor, or the like. Moreover, the touch sensor includes not only a contact type touch sensor but also a proximity type touch sensor or the like. Furthermore, the touch sensor may be a single sensor or a plurality of sensors arranged, for example, in an array. The user can initiate the chat response program 10 by touching the touch area.

In addition, the area of the display of the electronic device 1 may be the same as or different from the area of the touch sensor. Optionally, a display is stacked with the touch sensor to form a touch display. The device detects a user-triggered touch operation based on a touch screen display.

The electronic device 1 may further include a radio frequency (RF) circuit, a sensor, an audio circuit, and the like, and details are not described herein.

Referring to FIG. 2, it is a schematic diagram of interaction between the electronic device 1 and the client 2 according to a preferred embodiment of the present application. The chat response program 10 runs in the electronic device 1. In Fig. 2, the preferred embodiment of the electronic device 1 is a server. The electronic device 1 is communicatively coupled to the client 2 via a network 3. The client 2 can run in various types of terminal devices, such as smart phones, portable computers, and the like. After the user logs in to the electronic device 1 through the client 2, the session question can be input to the chat answering program 10, and the session question can be a session problem for a specific domain or a chat session content. The chat response program 10 can adopt the chat response method, determine an appropriate response content according to the session problem, and feed back the response content to the client 2.

Referring to FIG. 3, it is a flowchart of a preferred embodiment of the chat response method of the present application. The following steps of implementing the chat response method when the processor 12 of the electronic device 1 executes the chat response program 10 stored in the memory 11:

Step S1: Obtain a session question input by the client, perform pre-processing on the session problem, and obtain text feature information of the session problem, where the text feature information includes part of speech, location, and word class attribution information of each term in the conversation question. The word class attribution includes attribution to a keyword or a named entity. The session question can be, for example, a conversational question for a particular domain, such as "how long the warranty period is," or it can be a chat session content, such as "The weather is very good today." In order to facilitate subsequent processing of the session problem, step S1 may first perform some pre-processing on the session problem.

Specifically, the pre-processing performed in step S1 may include the following processing:

Performing word segmentation on the conversation problem, thereby segmenting the terms of the conversation problem. For example, the conversation question is “how long the warranty period is”, and the terms obtained after the word segmentation are “warranty period” and “yes”. , "multi", "long", the method of word segmentation includes performing a forward maximum match based on a dictionary and/or performing a reverse maximum match based on a dictionary;

Performing part-of-speech analysis on each term obtained by the word segmentation, and marking the part-of-speech of each term. For example, for the example of the above-mentioned conversation problem, the result of the part-of-speech tagging according to the preset rule is “warranty period/noun”. "Yes/verbs", "multiple/adverbs", "long/adjectives", the part of speech analysis is realized by a part-of-speech tagging model obtained through a preset large-scale corpus training;

Named entity identification is performed on the session problem, thereby identifying a named entity having a specific meaning, the named entity includes a person name, a place name, an organization, a proper noun, and the method for identifying the named entity includes a dictionary-based and rule-based method And methods based on statistical learning;

Extracting a keyword from the conversation question according to the term and the named entity, the keyword being a phrase whose number of characters is greater than a first preset threshold, or a named entity existing in a preset dictionary The preset dictionary includes a business scenario-specific dictionary.

Step S2, constructing an inverted index for the question and answer knowledge base 4, the question and answer knowledge base 4 includes a plurality of questions arranged in advance and one or more answers associated with each question, and querying through the inverted index according to the text feature information The method queries the candidate question set related to the conversation question from the question and answer knowledge base 4, and separately calculates the text similarity of the session question and each candidate question in the candidate question set.

In one embodiment, the constructing the inverted index for the Q&A knowledge base 4 includes:

Each question and answer in the Q&A Knowledge Base 4 is divided into word segmentation, part-of-speech tagging, keyword extraction, keyword occurrence location record, assignment of ID number, and assignment of each term after each word and answer segmentation. ID number;

Each question and answer in the Q&A knowledge base 4 is sorted according to the corresponding ID number, and each term obtained after each word and answer word segmentation is sorted according to the corresponding ID number, and will have the same term ID. All question IDs and answer IDs are placed in the inverted record table corresponding to the entry;

Combine all inverted record tables into the final inverted index.

At least one candidate question is included in the candidate question set, and each candidate question has a certain degree of association with the session problem due to the manner of using an inverted index query. The association between each candidate question and the conversation question may be reflected by the text similarity. If the text similarity between the conversation problem and the corresponding candidate question is higher, the conversation problem is considered to be similar to the candidate problem. .

Specifically, the method for calculating the text similarity between the session problem and each candidate question in the candidate question set in step S2 may include:

Constructing a convolutional neural network, and performing sample training on all the problem sentences in the Q&A knowledge base 4 through the convolutional neural network, and obtaining a convolutional neural network model corresponding to the problem statement in the Q&A knowledge base 4;

Entering each of the conversation problem and the candidate question set into the convolutional neural network model, respectively, and obtaining the conversation problem and the candidate by convolution kernel convolution of the convolutional neural network model a feature vector corresponding to each candidate question in the question set;

Calculating a cosine distance between the feature vector corresponding to the session problem and the feature vector corresponding to each candidate question in the candidate question set, respectively, to obtain the session problem and each candidate problem in the candidate question set Text similarity.

Step S3: determining, according to the preset rule and the text similarity, whether there is an approximation problem of the conversation problem in the candidate question set, and if there is an approximation problem of the conversation problem in the candidate question set, then the Q&A knowledge base Finding the associated answer of the approximation question, and outputting the associated answer as the target answer of the conversation question.

Specifically, the preset rule may include: determining whether there is a candidate problem that the text similarity with the conversation problem is greater than the second preset threshold, and if there is a candidate problem that the text similarity with the conversation problem is greater than the second preset threshold, Then, an approximation problem of the conversation problem exists in the candidate problem set. If there is no candidate problem that the text similarity with the conversation problem is greater than the second preset threshold, then it is determined that there is no approximation problem of the conversation problem in the candidate problem set.

If there is a candidate problem that the text similarity with the conversation problem is greater than the second preset threshold, step S3 selects a candidate corresponding to the maximum text similarity from the candidate questions with the text similarity of the conversation problem being greater than the second preset threshold. The problem is taken as the approximation question, and the associated answer of the approximation question is searched in the Q&A knowledge base 4, and the associated answer is output as the target answer of the conversation question. It should be noted that the approximation question may also have more than one associated answer in the Q&A knowledge base 4. When the approximation question has multiple associated answers in the Q&A knowledge base 4, step S3 may take the plurality of associated answers. And outputting the associated answer with the highest frequency in the preset time period (for example, the most recent week) as the target answer of the conversation question.

Step S4: If there is no approximation problem of the conversation problem in the candidate question set, query the candidate related to the conversation problem from the question and answer knowledge base 4 by using an inverted index query according to the text feature information. An answer set, and separately calculating a topic similarity of the conversation question and each candidate answer in the candidate answer set.

At least one candidate answer is included in the set of candidate answers, and each candidate answer has a certain degree of association with the session question due to the manner of using an inverted index query. The association of each candidate answer with the conversation question may be reflected by the topic similarity, and if the topic similarity between the conversation question and the corresponding candidate answer is higher, the conversation problem and the topic of the candidate answer are considered The more similar, the more likely the candidate answer is to be the answer to the conversation question.

Specifically, the method for calculating the topic similarity of each of the candidate answers in the candidate answer set in step S4 may include:

The calculating the similarity degree between the conversation problem and each candidate answer in the candidate answer set respectively includes:

Extracting the conversation problem and the topic vector of each candidate answer in the candidate answer set by using a Linear Discriminant Analysis (LDA) model;

Calculating a cosine distance between a topic vector of the conversation question and a topic vector of each candidate answer in the candidate answer set, respectively, to obtain a topic similarity between the conversation question and each candidate answer in the candidate answer set .

Step S5: determining, according to the preset rule and the topic similarity, whether an approximate answer of the conversation problem exists in the candidate answer set, and if the approximate answer of the conversation problem exists in the candidate answer set, the approximation is The answer is output as the target answer to the conversation question.

Specifically, the preset rule may include: determining whether there is a candidate answer whose topic similarity with the session problem is greater than a third preset threshold, and if there is a candidate answer with a topic similarity of the conversation problem being greater than a third preset threshold, Then, an approximate answer to the conversation question exists in the candidate answer set. If there is no candidate answer with the topic similarity of the conversation problem being greater than the third preset threshold, it is determined that there is no approximate answer of the conversation question in the candidate answer set.

If there is a candidate answer whose topic similarity to the conversation question is greater than the third preset threshold, the candidate answer is taken as an approximate answer to the conversation question, and step S5 outputs the approximate answer as the target answer of the conversation question. It is worth noting that the candidate answers with the topic similarity of the conversation problem greater than the third preset threshold may also have more than one in the Q&A knowledge base 4, when the similarity with the topic of the conversation problem is greater than the third predetermined threshold. When there are multiple in the question and answer knowledge base 4, step S5 may take the approximate answer which is the highest frequency of the conversation question in the preset time period (for example, the most recent week).

Step S6, if there is no approximate answer of the conversation question in the candidate answer set, the iterative training of encoding and decoding each question and answer in the Q&A knowledge base 4 by using the seq2seq model, thereby constructing a sequence prediction model, The conversation question is input to the sequence prediction model to generate a strain answer, and the strain answer is output as a target answer of the conversation question. The seq2seq model consists of a forward long and short memory network LSTM model and a backward LSTM model for performing the encoding and decoding iterative training, and an attention mechanism for calculating hidden layer information weights for each encoding and decoding.

According to the chat response method provided in this embodiment, after the session problem is acquired and pre-processed, the candidate question set related to the session problem is queried from the question-and-answer knowledge base 4 by means of an inverted index query, and the said a textual similarity between the conversation problem and each candidate question in the candidate question set, determining whether there is an approximation problem of the conversation problem in the candidate question set, and if so, searching for the associated answer of the approximation question in the Q&A knowledge base 4, Outputting the associated answer as a target answer of the conversation question. If there is no approximation problem of the conversation problem in the candidate question set, querying and speaking from the Q&A knowledge base 4 by means of an inverted index query a set of candidate answers related to the conversation problem, and respectively calculating a topic similarity of the conversation question and each candidate answer in the candidate answer set, and determining whether there is an approximate answer of the conversation question in the candidate answer set, and if so, The approximate answer is output as the target answer of the conversation question, if not in the candidate answer set An approximate answer to the conversation problem exists, and iterative training of encoding and decoding each question and answer in the question and answer knowledge base by the seq2seq model, thereby constructing a sequence prediction model, and inputting the conversation problem into the sequence prediction model A strain answer is generated, and the strain answer is output as a target answer to the conversation question. The chat response method provided by the embodiment can provide accurate and responsive feedback for the client for the session problem, thereby improving the service quality.

Referring to FIG. 4, it is a program module diagram of the chat response program 10 in FIG. In the present embodiment, the chat response program 10 is divided into a plurality of modules that are stored in the memory 11 and executed by the processor 12 to complete the present application. A module as referred to in this application refers to a series of computer program instructions that are capable of performing a particular function.

The chat response program 10 can be divided into: a pre-processing module 110, a first calculation module 120, a question retrieval module 130, a second calculation module 140, an answer retrieval module 150, and an answer prediction module 160.

The pre-processing module 110 is configured to obtain a session problem input by the client, and perform pre-processing on the session problem to obtain text feature information of the session problem, where the text feature information includes the part of speech and location of each term in the conversation problem. And word class attribution information, the word class attribution includes attribution to a keyword or a named entity.

Specifically, the pre-processing module 110 is configured to perform the following pre-processing on the session problem:

Performing word segmentation on the conversation problem, thereby segmenting the terms of the conversation problem, the method of word segmentation includes performing forward maximum matching based on the dictionary and/or performing reverse maximum matching based on the dictionary;

Performing part-of-speech analysis on each term obtained by the word segmentation process, and labeling the part-of-speech of each term, and the part-of-speech analysis is realized by a part-of-speech tagging model obtained through a preset large-scale corpus training;

Named entity identification is performed on the session problem, thereby identifying a named entity having a specific meaning, the named entity includes a person name, a place name, an organization, a proper noun, and the method for identifying the named entity includes a dictionary-based and rule-based method And methods based on statistical learning;

Extracting a keyword from the conversation question according to the term and the named entity, the keyword being a phrase whose number of characters is greater than a first preset threshold, or a named entity existing in a preset dictionary The preset dictionary includes a business scenario-specific dictionary.

a first calculation module 120, configured to build an inverted index for the Q&A knowledge base 4, the Q&A knowledge base includes a plurality of questions pre-arranged and one or more answers associated with each question, according to the text feature information, The manner of inverting the index query queries the candidate question set related to the session question from the question and answer knowledge base 4, and separately calculates the text similarity of the session question and each candidate question in the candidate question set.

Specifically, the first calculating module 120 is configured to construct an inverted index for the question and answer knowledge base 4 by:

Each question and answer in the Q&A Knowledge Base 4 is divided into word segmentation, part-of-speech tagging, keyword extraction, keyword occurrence location record, assignment of ID number, and assignment of each term after each word and answer segmentation. ID number;

Each question and answer in the Q&A knowledge base 4 is sorted according to the corresponding ID number, and each term obtained after each word and answer word segmentation is sorted according to the corresponding ID number, and will have the same term ID. All question IDs and answer IDs are placed in the inverted record table corresponding to the entry;

Combine all inverted record tables into the final inverted index.

The first calculation module 120 calculates a text similarity between the conversation problem and each candidate question in the candidate question set, including:

Constructing a convolutional neural network, and performing sample training on all the problem sentences in the Q&A knowledge base 4 through the convolutional neural network, and obtaining a convolutional neural network model corresponding to the problem statement in the Q&A knowledge base 4;

Entering each of the conversation problem and the candidate question set into the convolutional neural network model, respectively, and obtaining the conversation problem and the candidate by convolution kernel convolution of the convolutional neural network model a feature vector corresponding to each candidate question in the question set;

Calculating a cosine distance between the feature vector corresponding to the session problem and the feature vector corresponding to each candidate question in the candidate question set, respectively, to obtain the session problem and each candidate problem in the candidate question set Text similarity.

The problem retrieving module 130 is configured to determine, according to the preset rule and the text similarity, whether there is an approximation problem of the session problem in the candidate question set, and if there is an approximation problem of the session problem in the candidate question set, Finding the associated answer of the approximate question in the question and answer knowledge base, and outputting the associated answer as the target answer of the conversation question.

Specifically, the problem retrieval module 130 determines whether there is a candidate problem that the text similarity with the conversation problem is greater than the second preset threshold, and if so, from the candidate problem that the text similarity with the conversation problem is greater than the second preset threshold Selecting a candidate question corresponding to the maximum text similarity as the approximation question; if there is no candidate problem that the text similarity with the session problem is greater than the second preset threshold, determining that the session problem does not exist in the candidate question set Approximate problem.

The second calculating module 140 is configured to: if the approximation problem of the session problem does not exist in the candidate question set, query and query from the Q&A knowledge base 4 by using an inverted index query according to the text feature information A set of candidate answers related to the conversation question, and calculating a topic similarity of the conversation question and each candidate answer in the candidate answer set, respectively.

The second calculation module 140 calculates a topic similarity between the conversation question and each candidate answer in the candidate answer set, including:

Extracting the conversation problem and the topic vector of each candidate answer in the candidate answer set by using a linear discriminant analysis model;

Calculating a cosine distance between a topic vector of the conversation question and a topic vector of each candidate answer in the candidate answer set, respectively, to obtain a topic similarity between the conversation question and each candidate answer in the candidate answer set .

The answer retrieval module 150 is configured to determine, according to the preset rule and the topic similarity, whether an approximate answer of the conversation problem exists in the candidate answer set, if an approximate answer of the conversation problem exists in the candidate answer set, The approximate answer is output as a target answer to the conversation question.

Specifically, the answer retrieval module 150 determines whether there is a candidate answer whose topic similarity to the conversation problem is greater than the third preset threshold, and if so, from the candidate answers that the topic similarity with the conversation question is greater than the third preset threshold Selecting a candidate answer corresponding to the maximum topic similarity as the approximate answer; if there is no candidate answer with a topic similarity of the conversation problem greater than the third preset threshold, determining that the session problem does not exist in the candidate answer set Approximate answer.

The answer prediction module 160 is configured to: if the approximate answer of the conversation problem does not exist in the candidate answer set, perform iterative training on encoding and decoding each question and answer in the Q&A knowledge base 4 by using the seq2seq model, thereby constructing a sequence And predicting a model, inputting the conversation question into the sequence prediction model to generate a strain answer, and outputting the strain answer as a target answer of the conversation question. The answer prediction module 160 describes the seq2seq model by a forward-long memory network LSTM model and a backward LSTM model for performing the encoding and decoding iterative training, and attention for calculating hidden layer information weights for each encoding and decoding. Mechanism composition.

In the operating environment diagram of the preferred embodiment of the electronic device 1 shown in FIG. 1, the memory 11 including the readable storage medium may include an operating system, a chat response program 10, and a question and answer knowledge base 4. When the processor 12 executes the chat response program 10 stored in the memory 11, the following steps are implemented:

a pre-processing step: obtaining a session question input by the client, pre-processing the session problem, and obtaining text feature information of the session problem, where the text feature information includes part of speech, location, and part-of-speech attribution of each term in the conversation problem Information, the word class attribution includes belonging to a keyword or a named entity;

a first calculating step: constructing an inverted index for the question and answer knowledge base, the question and answer knowledge base including a plurality of questions arranged in advance and one or more answers associated with each question, and querying through the inverted index according to the text feature information Means querying a candidate question set related to the conversation problem from a question and answer knowledge base, and separately calculating text similarity between the conversation problem and each candidate question in the candidate question set;

a problem retrieval step: determining, according to a preset rule and the text similarity, whether there is an approximation problem of the conversation problem in the candidate question set, and if there is an approximation problem of the conversation problem in the candidate question set, then the question and answer knowledge Finding an associated answer of the approximate problem in the library, and outputting the associated answer as a target answer of the conversation question;

a second calculating step: if there is no approximation problem of the session problem in the candidate question set, querying, according to the text feature information, a query related to the session problem from the Q&A knowledge base by means of an inverted index query a set of candidate answers, and separately calculating a topic similarity of the conversation question and each candidate answer in the candidate answer set;

An answer retrieval step: determining, according to a preset rule and the topic similarity, whether an approximate answer of the conversation question exists in the candidate answer set, and if an approximate answer of the conversation problem exists in the candidate answer set, The approximate answer is output as the target answer to the conversation question;

An answer prediction step: if an approximate answer of the conversation problem does not exist in the candidate answer set, iteratively trains and decodes each question and answer in the question and answer knowledge base through the seq2seq model, thereby constructing a sequence prediction model, The conversation question is input to the sequence prediction model to generate a strain answer, and the strain answer is output as a target answer of the conversation question.

The pre-processing of the session problem includes:

Performing word segmentation on the conversation problem, thereby segmenting the terms of the conversation problem, the method of word segmentation includes performing forward maximum matching based on the dictionary and/or performing reverse maximum matching based on the dictionary;

Performing part-of-speech analysis on each term obtained by the word segmentation process, and labeling the part-of-speech of each term, and the part-of-speech analysis is realized by a part-of-speech tagging model obtained through a preset large-scale corpus training;

Named entity identification is performed on the session problem, thereby identifying a named entity having a specific meaning, the named entity includes a person name, a place name, an organization, a proper noun, and the method for identifying the named entity includes a dictionary-based and rule-based method And methods based on statistical learning;

Extracting a keyword from the conversation question according to the term and the named entity, the keyword being a phrase whose number of characters is greater than a first preset threshold, or a named entity existing in a preset dictionary The preset dictionary includes a business scenario-specific dictionary.

And calculating, respectively, the text similarity between the conversation problem and each candidate question in the candidate question set includes:

Constructing a convolutional neural network, and performing sample training on all problem sentences in the question and answer knowledge base through the convolutional neural network, and obtaining a convolutional neural network model corresponding to the problem statement in the question and answer knowledge base;

Entering each of the conversation problem and the candidate question set into the convolutional neural network model, respectively, and obtaining the conversation problem and the candidate by convolution kernel convolution of the convolutional neural network model a feature vector corresponding to each candidate question in the question set;

Calculating a cosine distance between the feature vector corresponding to the session problem and the feature vector corresponding to each candidate question in the candidate question set, respectively, to obtain the session problem and each candidate problem in the candidate question set Text similarity

The calculating the similarity degree between the conversation problem and each candidate answer in the candidate answer set respectively includes:

Extracting the conversation problem and the topic vector of each candidate answer in the candidate answer set by using a linear discriminant analysis model;

Calculating a cosine distance between a topic vector of the conversation question and a topic vector of each candidate answer in the candidate answer set, respectively, to obtain a topic similarity between the conversation question and each candidate answer in the candidate answer set .

The approximating the problem of determining whether the session problem exists in the candidate question set according to the preset rule and the problem similarity includes:

Determining whether there is a candidate problem that the text similarity with the conversation problem is greater than the second preset threshold, and if yes, selecting the maximum text similarity corresponding to the candidate problem that the text similarity with the conversation problem is greater than the second preset threshold Candidate questions as the approximation problem;

If there is no candidate problem that the text similarity with the conversation problem is greater than the second preset threshold, determining that there is no approximation problem of the conversation problem in the candidate problem set;

The determining, according to the preset rule and the topic similarity, determining whether the candidate answer exists in the candidate answer set includes:

Determining whether there is a candidate answer whose topic similarity to the conversation problem is greater than a third preset threshold, and if yes, selecting a maximum topic similarity corresponding to the candidate answers having the topic similarity of the conversation problem being greater than the third preset threshold a candidate answer as the approximate answer;

If there is no candidate answer with the topic similarity of the conversation problem being greater than the third preset threshold, it is determined that the approximate answer of the conversation question does not exist in the candidate answer set.

The constructing the inverted index for the Q&A knowledge base includes:

Each question and answer in the Q&A knowledge base is divided into word segmentation, part-of-speech tagging, keyword extraction, keyword location location record, assignment ID number, and ID assigned to each term after each word and answer segmentation. number;

Each question and answer in the Q&A knowledge base is sorted according to the corresponding ID number, and each term obtained after each word and answer word segmentation is sorted according to the corresponding ID number, and all the items with the same item ID are The question ID and the answer ID are placed in the inverted record table corresponding to the entry;

Combine all inverted record tables into the final inverted index.

The seq2seq model consists of a forward long and short memory network LSTM model and a backward LSTM model for performing the encoding and decoding iterative training, and an attention mechanism for calculating hidden layer information weights for each encoding and decoding.

For specific principles, please refer to the program module diagram of the chat response program 10 in FIG. 4 and the flowchart of the preferred embodiment of the chat response method in FIG.

In addition, the embodiment of the present application further provides a computer readable storage medium, which may be a hard disk, a multimedia card, an SD card, a flash memory card, an SMC, a read only memory (ROM), and an erasable programmable Any combination or combination of any one or more of read only memory (EPROM), portable compact disk read only memory (CD-ROM), USB memory, and the like. The computer readable storage medium includes a Q&A knowledge base 4, a chat response program 10, and the like. When the chat response program 10 is executed by the processor 12, the following operations are implemented:

a pre-processing step: obtaining a session question input by the client, pre-processing the session problem, and obtaining text feature information of the session problem, where the text feature information includes part of speech, location, and part-of-speech attribution of each term in the conversation problem Information, the word class attribution includes belonging to a keyword or a named entity;

a first calculating step: constructing an inverted index for the question and answer knowledge base, the question and answer knowledge base including a plurality of questions arranged in advance and one or more answers associated with each question, and querying through the inverted index according to the text feature information Means querying a candidate question set related to the conversation problem from a question and answer knowledge base, and separately calculating text similarity between the conversation problem and each candidate question in the candidate question set;

a problem retrieval step: determining, according to a preset rule and the text similarity, whether there is an approximation problem of the conversation problem in the candidate question set, and if there is an approximation problem of the conversation problem in the candidate question set, then the question and answer knowledge Finding an associated answer of the approximate problem in the library, and outputting the associated answer as a target answer of the conversation question;

a second calculating step: if there is no approximation problem of the session problem in the candidate question set, querying, according to the text feature information, a query related to the session problem from the Q&A knowledge base by means of an inverted index query a set of candidate answers, and separately calculating a topic similarity of the conversation question and each candidate answer in the candidate answer set;

An answer retrieval step: determining, according to a preset rule and the topic similarity, whether an approximate answer of the conversation question exists in the candidate answer set, and if an approximate answer of the conversation problem exists in the candidate answer set, The approximate answer is output as the target answer to the conversation question;

An answer prediction step: if an approximate answer of the conversation problem does not exist in the candidate answer set, iteratively trains and decodes each question and answer in the question and answer knowledge base through the seq2seq model, thereby constructing a sequence prediction model, The conversation question is input to the sequence prediction model to generate a strain answer, and the strain answer is output as a target answer of the conversation question.

The pre-processing of the session problem includes:

Performing word segmentation on the conversation problem, thereby segmenting the terms of the conversation problem, the method of word segmentation includes performing forward maximum matching based on the dictionary and/or performing reverse maximum matching based on the dictionary;

Performing part-of-speech analysis on each term obtained by the word segmentation process, and labeling the part-of-speech of each term, and the part-of-speech analysis is realized by a part-of-speech tagging model obtained through a preset large-scale corpus training;

Named entity identification is performed on the session problem, thereby identifying a named entity having a specific meaning, the named entity includes a person name, a place name, an organization, a proper noun, and the method for identifying the named entity includes a dictionary-based and rule-based method And methods based on statistical learning;

Extracting a keyword from the conversation question according to the term and the named entity, the keyword being a phrase whose number of characters is greater than a first preset threshold, or a named entity existing in a preset dictionary The preset dictionary includes a business scenario-specific dictionary.

And calculating, respectively, the text similarity between the conversation problem and each candidate question in the candidate question set includes:

Constructing a convolutional neural network, and performing sample training on all problem sentences in the question and answer knowledge base through the convolutional neural network, and obtaining a convolutional neural network model corresponding to the problem statement in the question and answer knowledge base;

Entering each of the conversation problem and the candidate question set into the convolutional neural network model, respectively, and obtaining the conversation problem and the candidate by convolution kernel convolution of the convolutional neural network model a feature vector corresponding to each candidate question in the question set;

Calculating a cosine distance between the feature vector corresponding to the session problem and the feature vector corresponding to each candidate question in the candidate question set, respectively, to obtain the session problem and each candidate problem in the candidate question set Text similarity

The calculating the similarity degree between the conversation problem and each candidate answer in the candidate answer set respectively includes:

Extracting the conversation problem and the topic vector of each candidate answer in the candidate answer set by using a linear discriminant analysis model;

Calculating a cosine distance between a topic vector of the conversation question and a topic vector of each candidate answer in the candidate answer set, respectively, to obtain a topic similarity between the conversation question and each candidate answer in the candidate answer set .

The approximating the problem of determining whether the session problem exists in the candidate question set according to the preset rule and the problem similarity includes:

Determining whether there is a candidate problem that the text similarity with the conversation problem is greater than the second preset threshold, and if yes, selecting the maximum text similarity corresponding to the candidate problem that the text similarity with the conversation problem is greater than the second preset threshold Candidate questions as the approximation problem;

If there is no candidate problem that the text similarity with the conversation problem is greater than the second preset threshold, determining that there is no approximation problem of the conversation problem in the candidate problem set;

The determining, according to the preset rule and the topic similarity, determining whether the candidate answer exists in the candidate answer set includes:

Determining whether there is a candidate answer whose topic similarity to the conversation problem is greater than a third preset threshold, and if yes, selecting a maximum topic similarity corresponding to the candidate answers having the topic similarity of the conversation problem being greater than the third preset threshold a candidate answer as the approximate answer;

If there is no candidate answer with the topic similarity of the conversation problem being greater than the third preset threshold, it is determined that the approximate answer of the conversation question does not exist in the candidate answer set.

The constructing the inverted index for the Q&A knowledge base includes:

Each question and answer in the Q&A knowledge base is divided into word segmentation, part-of-speech tagging, keyword extraction, keyword location location record, assignment ID number, and ID assigned to each term after each word and answer segmentation. number;

Each question and answer in the Q&A knowledge base is sorted according to the corresponding ID number, and each term obtained after each word and answer word segmentation is sorted according to the corresponding ID number, and all the items with the same item ID are The question ID and the answer ID are placed in the inverted record table corresponding to the entry;

Combine all inverted record tables into the final inverted index.

The seq2seq model consists of a forward long and short memory network LSTM model and a backward LSTM model for performing the encoding and decoding iterative training, and an attention mechanism for calculating hidden layer information weights for each encoding and decoding.

The specific implementation of the computer readable storage medium of the present application is substantially the same as the above-described chat response method and the specific embodiment of the electronic device 1, and details are not described herein again.

It is to be understood that the term "comprises", "comprising", or any other variants thereof, is intended to encompass a non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a series of elements includes those elements. It also includes other elements not explicitly listed, or elements that are inherent to such a process, device, item, or method. An element that is defined by the phrase "comprising a ..." does not exclude the presence of additional equivalent elements in the process, the device, the item, or the method that comprises the element.

Through the description of the above embodiments, those skilled in the art can clearly understand that the foregoing embodiment method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be through hardware, but in many cases, the former is better. Implementation. Based on such understanding, portions of the technical solution of the present application that contribute substantially or to the prior art may be embodied in the form of a software product stored in a storage medium as described above, including a number of instructions. To enable a terminal device (which may be a mobile phone, a computer, a server, or a network device, etc.) to perform the methods described in the various embodiments of the present application.

The above is only a preferred embodiment of the present application, and is not intended to limit the scope of the patent application, and the equivalent structure or equivalent process transformations made by the specification and the drawings of the present application, or directly or indirectly applied to other related technical fields. The same is included in the scope of patent protection of this application.

Claims (20)

A chat response method, the method comprising: a pre-processing step: obtaining a session question input by the client, pre-processing the session problem, and obtaining text feature information of the session problem, where the text feature information includes part of speech, location, and part-of-speech attribution of each term in the conversation problem Information, the word class attribution includes belonging to a keyword or a named entity; a first calculating step: constructing an inverted index for the question and answer knowledge base, the question and answer knowledge base including a plurality of questions arranged in advance and one or more answers associated with each question, and querying through the inverted index according to the text feature information Means querying a candidate question set related to the conversation problem from a question and answer knowledge base, and separately calculating text similarity between the conversation problem and each candidate question in the candidate question set; a problem retrieval step: determining, according to a preset rule and the text similarity, whether there is an approximation problem of the conversation problem in the candidate question set, and if there is an approximation problem of the conversation problem in the candidate question set, then the question and answer knowledge Finding an associated answer of the approximate problem in the library, and outputting the associated answer as a target answer of the conversation question; a second calculating step: if there is no approximation problem of the session problem in the candidate question set, querying, according to the text feature information, a query related to the session problem from the Q&A knowledge base by means of an inverted index query a set of candidate answers, and separately calculating a topic similarity of the conversation question and each candidate answer in the candidate answer set; An answer retrieval step: determining, according to a preset rule and the topic similarity, whether an approximate answer of the conversation question exists in the candidate answer set, and if an approximate answer of the conversation problem exists in the candidate answer set, The approximate answer is output as the target answer to the conversation question; An answer prediction step: if an approximate answer of the conversation problem does not exist in the candidate answer set, iteratively trains and decodes each question and answer in the question and answer knowledge base through the seq2seq model, thereby constructing a sequence prediction model, The conversation question is input to the sequence prediction model to generate a strain answer, and the strain answer is output as a target answer of the conversation question. The chat response method according to claim 1, wherein the preprocessing the session problem comprises: Performing word segmentation on the conversation problem, thereby segmenting the terms of the conversation problem, the method of word segmentation includes performing forward maximum matching based on the dictionary and/or performing reverse maximum matching based on the dictionary; Performing part-of-speech analysis on each term obtained by the word segmentation process, and labeling the part-of-speech of each term, and the part-of-speech analysis is realized by a part-of-speech tagging model obtained through a preset large-scale corpus training; Named entity identification is performed on the session problem, thereby identifying a named entity having a specific meaning, the named entity includes a person name, a place name, an organization, a proper noun, and the method for identifying the named entity includes a dictionary-based and rule-based method And methods based on statistical learning; Extracting a keyword from the conversation question according to the term and the named entity, the keyword being a phrase whose number of characters is greater than a first preset threshold, or a named entity existing in a preset dictionary The preset dictionary includes a business scenario-specific dictionary. The chat response method according to claim 1, wherein the calculating the text similarity of each of the session problem and each candidate question in the candidate question set separately comprises: Constructing a convolutional neural network, and performing sample training on all problem sentences in the question and answer knowledge base through the convolutional neural network, and obtaining a convolutional neural network model corresponding to the problem statement in the question and answer knowledge base; Entering each of the conversation problem and the candidate question set into the convolutional neural network model, respectively, and obtaining the conversation problem and the candidate by convolution kernel convolution of the convolutional neural network model a feature vector corresponding to each candidate question in the question set; Calculating a cosine distance between the feature vector corresponding to the session problem and the feature vector corresponding to each candidate question in the candidate question set, respectively, to obtain the session problem and each candidate problem in the candidate question set Text similarity The calculating the similarity degree between the conversation problem and each candidate answer in the candidate answer set respectively includes: Extracting the conversation problem and the topic vector of each candidate answer in the candidate answer set by using a linear discriminant analysis model; Calculating a cosine distance between a topic vector of the conversation question and a topic vector of each candidate answer in the candidate answer set, respectively, to obtain a topic similarity between the conversation question and each candidate answer in the candidate answer set . The chat response method according to claim 1, wherein the approximating the problem of whether the session problem exists in the candidate question set according to the preset rule and the problem similarity includes: Determining whether there is a candidate problem that the text similarity with the conversation problem is greater than the second preset threshold, and if yes, selecting the maximum text similarity corresponding to the candidate problem that the text similarity with the conversation problem is greater than the second preset threshold Candidate questions as the approximation problem; If there is no candidate problem that the text similarity with the conversation problem is greater than the second preset threshold, determining that there is no approximation problem of the conversation problem in the candidate problem set; The determining, according to the preset rule and the topic similarity, determining whether the candidate answer exists in the candidate answer set includes: Determining whether there is a candidate answer whose topic similarity to the conversation problem is greater than a third preset threshold, and if yes, selecting a maximum topic similarity corresponding to the candidate answers having the topic similarity of the conversation problem being greater than the third preset threshold a candidate answer as the approximate answer; If there is no candidate answer with the topic similarity of the conversation problem being greater than the third preset threshold, it is determined that the approximate answer of the conversation question does not exist in the candidate answer set. The chat response method according to claim 1, wherein the constructing the inverted index for the question and answer knowledge base comprises: Each question and answer in the Q&A knowledge base is divided into word segmentation, part-of-speech tagging, keyword extraction, keyword location location record, assignment ID number, and ID assigned to each term after each word and answer segmentation. number; Each question and answer in the Q&A knowledge base is sorted according to the corresponding ID number, and each term obtained after each word and answer word segmentation is sorted according to the corresponding ID number, and all the items with the same item ID are The question ID and the answer ID are placed in the inverted record table corresponding to the entry; Combine all inverted record tables into the final inverted index. The chat response method according to claim 1, wherein said seq2seq model is used by a forward-long memory network LSTM model and a backward LSTM model for performing said encoding and decoding iterative training, and for calculating each time The attention and mechanism of the hidden layer information weighting of encoding and decoding. A chat response method according to any one of claims 2 to 5, wherein said seq2seq model is composed of a forward long and short memory network LSTM model and a backward LSTM model for performing said coding and decoding iterative training, and A attention mechanism for calculating the weight of hidden layer information for each encoding and decoding. An electronic device includes a memory and a processor, wherein the memory includes a chat response program, and the chat response program is executed by the processor to implement the following steps: a pre-processing step: obtaining a session question input by the client, pre-processing the session problem, and obtaining text feature information of the session problem, where the text feature information includes part of speech, location, and part-of-speech attribution of each term in the conversation problem Information, the word class attribution includes belonging to a keyword or a named entity; a first calculating step: constructing an inverted index for the question and answer knowledge base, the question and answer knowledge base including a plurality of questions arranged in advance and one or more answers associated with each question, and querying through the inverted index according to the text feature information Means querying a candidate question set related to the conversation problem from a question and answer knowledge base, and separately calculating text similarity between the conversation problem and each candidate question in the candidate question set; a problem retrieval step: determining, according to a preset rule and the text similarity, whether there is an approximation problem of the conversation problem in the candidate question set, and if there is an approximation problem of the conversation problem in the candidate question set, then the question and answer knowledge Finding an associated answer of the approximate problem in the library, and outputting the associated answer as a target answer of the conversation question; a second calculating step: if there is no approximation problem of the session problem in the candidate question set, querying, according to the text feature information, a query related to the session problem from the Q&A knowledge base by means of an inverted index query a set of candidate answers, and separately calculating a topic similarity of the conversation question and each candidate answer in the candidate answer set; An answer retrieval step: determining, according to a preset rule and the topic similarity, whether an approximate answer of the conversation question exists in the candidate answer set, and if an approximate answer of the conversation problem exists in the candidate answer set, The approximate answer is output as the target answer to the conversation question; An answer prediction step: if an approximate answer of the conversation problem does not exist in the candidate answer set, iteratively trains and decodes each question and answer in the question and answer knowledge base through the seq2seq model, thereby constructing a sequence prediction model, The conversation question is input to the sequence prediction model to generate a strain answer, and the strain answer is output as a target answer of the conversation question. The electronic device of claim 8 wherein said pre-processing said session problem comprises: Performing word segmentation on the conversation problem, thereby segmenting the terms of the conversation problem, the method of word segmentation includes performing forward maximum matching based on the dictionary and/or performing reverse maximum matching based on the dictionary; Performing part-of-speech analysis on each term obtained by the word segmentation process, and labeling the part-of-speech of each term, and the part-of-speech analysis is realized by a part-of-speech tagging model obtained through a preset large-scale corpus training; Named entity identification is performed on the session problem, thereby identifying a named entity having a specific meaning, the named entity includes a person name, a place name, an organization, a proper noun, and the method for identifying the named entity includes a dictionary-based and rule-based method And methods based on statistical learning; Extracting a keyword from the conversation question according to the term and the named entity, the keyword being a phrase whose number of characters is greater than a first preset threshold, or a named entity existing in a preset dictionary The preset dictionary includes a business scenario-specific dictionary. The electronic device according to claim 9, wherein the calculating the text similarity of each of the conversation problem and each candidate question in the candidate question set respectively comprises: Constructing a convolutional neural network, and performing sample training on all problem sentences in the question and answer knowledge base through the convolutional neural network, and obtaining a convolutional neural network model corresponding to the problem statement in the question and answer knowledge base; Entering each of the conversation problem and the candidate question set into the convolutional neural network model, respectively, and obtaining the conversation problem and the candidate by convolution kernel convolution of the convolutional neural network model a feature vector corresponding to each candidate question in the question set; Calculating a cosine distance between the feature vector corresponding to the session problem and the feature vector corresponding to each candidate question in the candidate question set, respectively, to obtain the session problem and each candidate problem in the candidate question set Text similarity The calculating the similarity degree between the conversation problem and each candidate answer in the candidate answer set respectively includes: Extracting the conversation problem and the topic vector of each candidate answer in the candidate answer set by using a linear discriminant analysis model; Calculating a cosine distance between a topic vector of the conversation question and a topic vector of each candidate answer in the candidate answer set, respectively, to obtain a topic similarity between the conversation question and each candidate answer in the candidate answer set . The electronic device according to claim 8, wherein the approximating the problem of whether the session problem exists in the candidate question set according to the preset rule and the problem similarity includes: Determining whether there is a candidate problem that the text similarity with the conversation problem is greater than the second preset threshold, and if yes, selecting the maximum text similarity corresponding to the candidate problem that the text similarity with the conversation problem is greater than the second preset threshold Candidate questions as the approximation problem; If there is no candidate problem that the text similarity with the conversation problem is greater than the second preset threshold, determining that there is no approximation problem of the conversation problem in the candidate problem set; According to the preset rule and the topic similarity, determining whether the candidate answer has an approximate answer of the conversation problem includes: Determining whether there is a candidate answer whose topic similarity to the conversation problem is greater than a third preset threshold, and if yes, selecting a maximum topic similarity corresponding to the candidate answers having the topic similarity of the conversation problem being greater than the third preset threshold a candidate answer as the approximate answer; If there is no candidate answer with the topic similarity of the conversation problem being greater than the third preset threshold, it is determined that the approximate answer of the conversation question does not exist in the candidate answer set. The electronic device according to claim 8, wherein the constructing the inverted index for the Q&A knowledge base comprises: Each question and answer in the Q&A knowledge base is divided into word segmentation, part-of-speech tagging, keyword extraction, keyword location location record, assignment ID number, and ID assigned to each term after each word and answer segmentation. number; Each question and answer in the Q&A knowledge base is sorted according to the corresponding ID number, and each term obtained after each word and answer word segmentation is sorted according to the corresponding ID number, and all the items with the same item ID are The question ID and the answer ID are placed in the inverted record table corresponding to the entry; Combine all inverted record tables into the final inverted index. The electronic device of claim 8 wherein said seq2seq model is comprised of a forward long and short memory network LSTM model and a backward LSTM model for performing said encoding and decoding iterative training, and for calculating each encoding And the attention mechanism of the decoded hidden layer information weights. The electronic device according to any one of claims 9 to 12, wherein the seq2seq model is used by a forward-long memory network LSTM model and a backward LSTM model for performing the encoding and decoding iterative training, and The attention mechanism is constructed to calculate the hidden layer information weights for each encoding and decoding. A computer readable storage medium, comprising: a chat response program, wherein when the chat response program is executed by a processor, the chat response program is executed by the processor to implement the following steps : a pre-processing step: obtaining a session question input by the client, pre-processing the session problem, and obtaining text feature information of the session problem, where the text feature information includes part of speech, location, and part-of-speech attribution of each term in the conversation problem Information, the word class attribution includes belonging to a keyword or a named entity; a first calculating step: constructing an inverted index for the question and answer knowledge base, the question and answer knowledge base including a plurality of questions arranged in advance and one or more answers associated with each question, and querying through the inverted index according to the text feature information Means querying a candidate question set related to the conversation problem from a question and answer knowledge base, and separately calculating text similarity between the conversation problem and each candidate question in the candidate question set; a problem retrieval step: determining, according to a preset rule and the text similarity, whether there is an approximation problem of the conversation problem in the candidate question set, and if there is an approximation problem of the conversation problem in the candidate question set, then the question and answer knowledge Finding an associated answer of the approximate problem in the library, and outputting the associated answer as a target answer of the conversation question; a second calculating step: if there is no approximation problem of the session problem in the candidate question set, querying, according to the text feature information, a query related to the session problem from the Q&A knowledge base by means of an inverted index query a set of candidate answers, and separately calculating a topic similarity of the conversation question and each candidate answer in the candidate answer set; An answer retrieval step: determining, according to a preset rule and the topic similarity, whether an approximate answer of the conversation question exists in the candidate answer set, and if an approximate answer of the conversation problem exists in the candidate answer set, The approximate answer is output as the target answer to the conversation question; An answer prediction step: if an approximate answer of the conversation problem does not exist in the candidate answer set, iteratively trains and decodes each question and answer in the question and answer knowledge base through the seq2seq model, thereby constructing a sequence prediction model, The conversation question is input to the sequence prediction model to generate a strain answer, and the strain answer is output as a target answer of the conversation question. The computer readable storage medium of claim 15 wherein said preprocessing said session problem comprises: Performing word segmentation on the conversation problem, thereby segmenting the terms of the conversation problem, the method of word segmentation includes performing forward maximum matching based on the dictionary and/or performing reverse maximum matching based on the dictionary; Performing part-of-speech analysis on each term obtained by the word segmentation process, and labeling the part-of-speech of each term, and the part-of-speech analysis is realized by a part-of-speech tagging model obtained through a preset large-scale corpus training; Named entity identification is performed on the session problem, thereby identifying a named entity having a specific meaning, the named entity includes a person name, a place name, an organization, a proper noun, and the method for identifying the named entity includes a dictionary-based and rule-based method And methods based on statistical learning; Extracting a keyword from the conversation question according to the term and the named entity, the keyword being a phrase whose number of characters is greater than a first preset threshold, or a named entity existing in a preset dictionary The preset dictionary includes a business scenario-specific dictionary. The computer readable storage medium of claim 16, wherein the calculating the text similarity of the session question and each candidate question in the candidate question set respectively comprises: Constructing a convolutional neural network, and performing sample training on all problem sentences in the question and answer knowledge base through the convolutional neural network, and obtaining a convolutional neural network model corresponding to the problem statement in the question and answer knowledge base; Entering each of the conversation problem and the candidate question set into the convolutional neural network model, respectively, and obtaining the conversation problem and the candidate by convolution kernel convolution of the convolutional neural network model a feature vector corresponding to each candidate question in the question set; Calculating a cosine distance between the feature vector corresponding to the session problem and the feature vector corresponding to each candidate question in the candidate question set, respectively, to obtain the session problem and each candidate problem in the candidate question set Text similarity The calculating the similarity degree between the conversation problem and each candidate answer in the candidate answer set respectively includes: Extracting the conversation problem and the topic vector of each candidate answer in the candidate answer set by using a linear discriminant analysis model; Calculating a cosine distance between a topic vector of the conversation question and a topic vector of each candidate answer in the candidate answer set, respectively, to obtain a topic similarity between the conversation question and each candidate answer in the candidate answer set . The computer readable storage medium according to claim 15, wherein the approximating the problem of whether the session problem exists in the candidate question set according to the preset rule and the problem similarity includes: Determining whether there is a candidate problem that the text similarity with the conversation problem is greater than the second preset threshold, and if yes, selecting the maximum text similarity corresponding to the candidate problem that the text similarity with the conversation problem is greater than the second preset threshold Candidate questions as the approximation problem; If there is no candidate problem that the text similarity with the conversation problem is greater than the second preset threshold, determining that there is no approximation problem of the conversation problem in the candidate problem set; The determining, according to the preset rule and the topic similarity, determining whether the candidate answer exists in the candidate answer set includes: Determining whether there is a candidate answer whose topic similarity to the conversation problem is greater than a third preset threshold, and if yes, selecting a maximum topic similarity corresponding to the candidate answers having the topic similarity of the conversation problem being greater than the third preset threshold a candidate answer as the approximate answer; If there is no candidate answer with the topic similarity of the conversation problem being greater than the third preset threshold, it is determined that the approximate answer of the conversation question does not exist in the candidate answer set. The computer readable storage medium of claim 15, wherein the constructing the inverted index for the question and answer knowledge base comprises: Each question and answer in the Q&A knowledge base is divided into word segmentation, part-of-speech tagging, keyword extraction, keyword location location record, assignment ID number, and ID assigned to each term after each word and answer segmentation. number; Each question and answer in the Q&A knowledge base is sorted according to the corresponding ID number, and each term obtained after each word and answer word segmentation is sorted according to the corresponding ID number, and all the items with the same item ID are The question ID and the answer ID are placed in the inverted record table corresponding to the entry; Combine all inverted record tables into the final inverted index. The computer readable storage medium of claim 15 wherein said seq2seq model is comprised of a forward long and short memory network LSTM model and a backward LSTM model for performing said encoding and decoding iterative training, and for computing Each time the encoding and decoding of the hidden layer information weights attention mechanism is constructed.

Images