CN113157885B - An efficient intelligent question answering system for artificial intelligence domain knowledge - Google Patents

Abstract

The invention relates to a high-efficiency intelligent question-answering system oriented to knowledge in the field of artificial intelligence, which comprises a preparation module and a question-answering module; the preparation module comprises a data collection module, a model training module and a knowledge structure construction module of a question-answering system; the question-answering module comprises an input preprocessing module, a question-answering module based on a knowledge base, a question-answering module based on a text base and a question recommending module based on the knowledge base. According to the invention, through the preparation module and the question-answering module, the word segmentation accuracy of the user questions, the knowledge base questions and the text base questions is greatly enhanced, and the overall accuracy of the full question-answering system is greatly improved, so that the user experience is greatly improved, and the knowledge question-answering service with low cost, high efficiency and high user experience is realized.

Description

An efficient intelligent question answering system for artificial intelligence domain knowledge

technical field

The invention relates to the technical fields of artificial intelligence and natural language processing, in particular to an efficient intelligent question answering system oriented to knowledge in the field of artificial intelligence.

Background technique

In recent years, artificial intelligence technology has developed rapidly and has broad application prospects in education, medical care, agriculture, transportation and other fields. However, obtaining knowledge in the field of artificial intelligence requires a certain professional foundation. Practitioners in all walks of life lack a convenient and accurate way to obtain artificial intelligence knowledge, which makes it difficult to popularize artificial intelligence technology in many fields and virtually hinders the society. development of productivity. Unstructured text in the field of artificial intelligence carries a large amount of knowledge in this field. If a knowledge question answering system based on text understanding in this field can be completed, it can provide people with an efficient and convenient way to acquire knowledge and promote the further development of artificial intelligence technology.

The existing knowledge question answering system has the following problems: First, the information extraction model lacks the support of entity names and entity aliases. The former causes related technical terms to be incorrectly segmented, which in turn affects the performance of search engines. The latter lacks an understanding of synonyms. Make subsequent search results one-sided. Both of these can adversely affect the overall performance of the question answering system. Secondly, as a complex natural language processing task, machine reading comprehension has problems such as high complexity and large amount of calculation, and the construction of the knowledge base relies on unstructured text, which is time-consuming and laborious if constructed manually, and difficult to form. A knowledge base of sufficient scale, both of which restrict the practical deployment of question answering systems. Finally, the existing question answering systems still lack the ability to efficiently obtain accurate and comprehensive answers from different types of texts across paragraphs, documents, and forms, and lack the ability to guide users to further explore relevant knowledge in the field.

Contents of the invention

In order to solve the technical problems existing in the prior art, the present invention provides an efficient intelligent question answering system for knowledge in the field of artificial intelligence. Through the preparation module and the question answering module, the accuracy of word segmentation for user questions, knowledge base questions, and text library questions can be improved. Greatly enhanced, and then greatly improved the overall accuracy of the full question answering system, thereby greatly improving user experience, and realizing knowledge question answering services with low cost, high efficiency and high user experience.

The present invention adopts the following technical solutions to achieve: an efficient intelligent question answering system for knowledge in the field of artificial intelligence, including: a preparation module and a question answering module; wherein, the preparation module includes a data collection module, a model training module and a question answering system knowledge structure building module; The question answering module includes input preprocessing module, question answering module based on knowledge base, question answering module based on text base and question recommendation module based on knowledge base;

The preparation module marks the collected unstructured knowledge text paragraphs in the field of artificial intelligence through the data collection module, and trains the information extraction module and machine reading comprehension module of the model training module, and at the same time collects or defines synonymous and different The short text matching model is trained by using meaningful questions, and the knowledge structure building module of the question answering system is used to extract knowledge triples from the trained information extraction model and form question and answer pairs. Provide semantics for search engines by improving the construction method of knowledge base and text base inverted index, and build a knowledge base keyword index;

The question answering module preprocesses the questions input by the user through the input preprocessing module, and uses the question answering module based on the knowledge base to search for answers. The question answering module of the library looks for and prepares to return answers, and uses the question recommendation module based on the knowledge base to recommend questions to users, and finally returns the answers and recommended questions to users.

Compared with the prior art, the present invention has the following advantages and beneficial effects:

1. The present invention supplements the jieba word segmentation dictionary through the entity name and its alias extracted by the information extraction model, so that the accuracy of word segmentation for user questions, knowledge base questions, and text library questions is greatly enhanced, thereby greatly improving the overall question-and-answer system accuracy, thereby greatly improving the user experience.

2. The present invention extracts the corresponding relationship between entity names and aliases through the information extraction model, as well as the synonyms dictionary obtained from the Internet, and uses the improved BM25 knowledge base coarse recall module, so that single retrieval hardly increases the reasoning time and sorts at the same time All content with different keywords is synonymous, and the impact of the document paragraph on the frequency of the subject word and the difference in the length of the document is more robust, so that the retrieval effect is improved.

3. The present invention uses information extraction technology to build a question-and-answer knowledge base based on unstructured text paragraphs, and preferably uses other available semi-structured and structured relevant text paragraphs as supplements, making knowledge acquisition channels more diverse and flexible , when the question expressed by the user through natural language is consistent with the existing knowledge semantics in the knowledge base, the matching can be completed, and the richness of the answer is enhanced at the same time.

4. The question answering system of the present invention can recommend relevant questions for users, guide users to further explore the knowledge system and inspire users to ask questions, which has high social value and strong practical significance.

5. The present invention has little demand and consumption of computing resources.

Description of drawings

Fig. 1 is a system structure diagram of the present invention;

Fig. 2 is the data collection module diagram in the preparation module of the present invention;

Fig. 3 is the model training module diagram in the preparation module of the present invention;

Fig. 4 is the HBT model training flowchart in the model training in the preparation module of the present invention;

Fig. 5 is the ESIM model training flowchart in the model training in the preparation module of the present invention;

Fig. 6 is the RoBERTa-QA model figure of the present invention;

Fig. 7 is a building block diagram of the knowledge structure of the question answering system in the preparation module of the present invention;

Fig. 8 is a building block diagram of the knowledge base inverted index in the question answering system knowledge structure building block in the preparation block of the present invention;

Fig. 9 is a building block diagram of the knowledge base keyword index in the question answering system knowledge structure building block in the preparation block of the present invention;

Fig. 10 is a building block diagram of the inverted index of the text library in the question answering system knowledge structure building block in the preparation block of the present invention;

Fig. 11 is the overall flowchart of the question answering module of the present invention;

Fig. 12 is a flow chart of a reference resolution method in the preprocessing module in the question answering module of the present invention;

Fig. 13 is a flow chart of the rough recall module in the question and answer module based on the knowledge base in the question and answer module of the present invention;

Fig. 14 is a flow chart of a method for judging the synonymity of questions by using ESIM in the question-and-answer module based on the knowledge base in the question-and-answer module of the present invention;

Fig. 15 is a flow chart of the coarse recall module in the question and answer module based on the text library in the question and answer module of the present invention;

Fig. 16 is a diagram of the question recommendation module based on the knowledge base in the question answering module of the present invention.

Detailed ways

The present invention will be further described in detail below in conjunction with the embodiments and the accompanying drawings, but the embodiments of the present invention are not limited thereto.

Example

As shown in Figure 1, the present embodiment is an efficient intelligent question answering system for knowledge in the field of artificial intelligence, including a preparation module and a question answering module; wherein, the preparation module includes a data collection module, a model training module and a question answering system knowledge structure building module; question answering The modules include input preprocessing module, question answering module based on knowledge base, question answering module based on text base and question recommendation module based on knowledge base;

The preparation module marks the collected unstructured knowledge text paragraphs in the field of artificial intelligence through the data collection module, and trains the information extraction module and machine reading comprehension module of the model training module, and at the same time collects or defines synonymous and different The short text matching model can be trained with meaningful questions, and the knowledge structure building module of the question answering system can be used to extract knowledge triples from the trained information extraction model and form question and answer pairs. Search, and then provide semantics for the search engine by improving the construction method of the knowledge base and the inverted index of the text base, and build a knowledge base keyword index to help the implementation of question recommendation;

The question answering module preprocesses the questions input by the user through the input preprocessing module, and uses the question answering module based on the knowledge base to search for answers. The question answering module of the library looks for and prepares to return answers, and uses the question recommendation module based on the knowledge base to recommend questions to users, and finally returns the answers and recommended questions to users.

As shown in Figure 2, in the present embodiment, the implementation process of the data collection module is as follows:

S21. Collect unstructured knowledge text paragraphs from sources such as scientific publications, literature, and network popular science knowledge related to the field of artificial intelligence. For text paragraphs that are too long, split them according to periods. Specifically, limit the length of each text paragraph. more than 480 characters;

S22. Define the key information triplet types extracted in the information extraction model, and define the following triplet types using the common relational definition method: entity-description-content, entity-proposer-content, entity-include-content, Entity-application-content, entity-alias-content, and mark the triplet types that have been defined, and collectively call the triplets subject-predicate-object. First, use the brat text labeling tool in the collected artificial intelligence In the field unstructured knowledge text paragraph, mark all the subjects by checking the subject, then mark all the objects by checking the object, and finally mark all the correspondence between the subject and the object in the text paragraph by connecting lines;

S23. Using the machine reading comprehension model, after collecting unstructured knowledge text paragraphs from sources such as scientific publications, literature, and popular science knowledge in the field of artificial intelligence in step S21, if there are corresponding questions, then in the text paragraphs Mark the starting and ending positions of the relevant content of the answer corresponding to the question, or directly simulate the real user scenario to define a variety of questions about scientific knowledge in the field of artificial intelligence based on part of the content in the text paragraph, and mark the answer corresponding to the question in the text paragraph start and end positions.

S24. Use the short text matching model to directly collect some synonymous questions and non-synonymous questions within 30 characters in this field. There are two pairs of synonymous questions, and the corresponding label is 1; two pairs of non-synonymous questions are paired, and the corresponding labels are 0.

As shown in Figure 3, in this embodiment, the implementation process of the model training module is as follows:

S31. Use the HBT model to build an information extraction model, then use the RoBERTa pre-training model to initialize model parameters, and use the marked triplet data for training. Specifically, the HBT model training process is shown in Figure 4. First, add the special symbol [CLS] at the beginning of the input text paragraph, add the special symbol [SEP] at the end of the text paragraph, and supplement it with [PAD] to 512 characters, after the vocabulary of the RoBERTa pre-trained model becomes the corresponding number , which is mapped to a vector of 756-dimensional feature dimensions through the Embedding layer of the RoBERTa pre-training model, and the output feature vector is obtained after the vector passes through all layers of the Transformer of the RoBERTa pre-training model. On the one hand, the output feature vector passes through the fully connected layer and the Sigmoid activation changes into two channels to obtain the starting point and end point prediction probability results of the subject; The output results of the fully connected layer feature fuser are added and input to the fully connected layer to change into a feature vector of 10 channels. The 10 channels correspond to the start and end points of the five predicate types. Finally, the feature vector is activated by Sigmoid to obtain the predicate-object The predicted probability of the starting point and ending point of . The predicted probabilities of the start and end points of the subject and predicate-object are calculated using the binary cross-entropy loss and the labeled data to calculate the error, and then the error is backpropagated to adjust the model parameters, and iterated in turn.

S32. Build an ESIM model to train a short text matching model. As shown in FIG. 5, first use the full corpus of Chinese Wikipedia to train Chinese character vectors using the word2vec method, and then use the Chinese translation results of the Quora Question Pairs dataset to pre-train the ESIM model. On the basis of the pre-trained model, use the large-scale short text matching data set LCMQC open in Chinese and the collected short text matching data set in the field of artificial intelligence to fine-tune the parameters of the ESIM model. The data set is resampled to a data size similar to that of the LCMQC data set, and the training process uses the binary cross-entropy loss to calculate the cross-entropy loss between the ESIM model output and the label result of 0 or 1.

S33. Build the RoBERTa-QA model to train the machine reading comprehension model. First, build the RoBERTa-QA model. The RoBERTa-QA model is shown in Figure 6. The input of this model is the splicing of text paragraphs and questions, and after splicing, it must follow the RoBERTa vocabulary The encoding method encodes the characters. Before encoding, add the special character [CLS] at the beginning position, add the special character [SEP] between the text paragraph and the question, add [SEP] at the end of the question, and use [PAD ] added to 512 characters. Based on RoBERTa, the RoBERTa-QA model adds a fully connected layer with an input channel number of 756 and an output channel number of 2 and a Softmax layer connected to it after the last layer of Transformer to predict the probability of the answer at the beginning and end of the text. , and then use the large-scale open data set DuReader to conduct further pre-training on the basis of the Chinese pre-training model RoBERTa, and use the collected machine reading comprehension annotation data in the field of artificial intelligence to perform parameter fine-tuning training, and use the binary cross-entropy during training loss to calculate the error.

As shown in Figure 7, in this embodiment, the implementation process of the knowledge structure building module of the question answering system is as follows:

S41. Collect a large number of unstructured knowledge texts in the field of artificial intelligence; specifically, collect unstructured text paragraphs from sources such as scientific publications, literature, and popular science knowledge in the field of artificial intelligence. For paragraphs that are too long, break them down according to the period points; specifically, limit the length of each paragraph to no more than 480 characters, and each sentence in the paragraph is complete.

S42. Use the information extraction model trained in step S31 to extract triples in a large number of unstructured knowledge text paragraphs in the field of artificial intelligence. The input text paragraphs add a special symbol [CLS] at the beginning of the text paragraph, and add at the end of the text. The special symbol [SEP] is supplemented with [PAD] to 512 characters. After the vocabulary of RoBERTa is changed to the corresponding number, it is mapped to a vector of 756-dimensional feature dimensions through the Embedding layer of RoBERTa, and the vector passes through all layers of Transformer of RoBERTa. Finally, the output feature vector is obtained, and the output feature vector passes through the fully connected layer and the Sigmoid activation changes to 2 channels to obtain the starting point and ending point prediction probability results of the subject. Take the starting point and ending point probability greater than 0.5 as the starting point and ending point, and the distance between the starting point and its back The closest endpoints are combined and paired into subject locations. The starting and ending points of each pair of subjects take out the corresponding output eigenvectors, copy them, add the output results of the fully connected layer feature fusion device, and then input them to the fully connected layer object predictor to change them into 10-channel feature vectors, which correspond to 5 types. The start point and end point of the predicate type. Finally, the feature vector is activated by Sigmoid to obtain the predicted probability of the start point and end point of the predicate-object. The start point and end point probability are greater than 0.5 as the start point and end point. . In the text paragraph, find the subject content and the corresponding predicate-object content according to the start and end positions, and combine them into several triples in the form of subject-predicate-object.

S43. Separately propose the entity name and entity alias content in the triplet to supplement the dictionary of the jieba word segmentation tool to help greatly improve the accuracy of word segmentation; on the other hand, extract the triplet results according to certain rules Form question answer key-value pairs, such as: entity-description-content formation {entity} what is: {content}; entity-application-content formation {entity} what application: {content}; entity-inclusion-content formation {entity } What are included: {content}; Entity-proposer-content formation {entity} Who proposed: {content}; Crawl the triple knowledge in Wikipedia and form a question and answer pair to supplement the knowledge base according to the format of {entry name}'s {attribute}:{content}. If you can find high-quality question-answer pairs that already exist in this field, put them into the knowledge base to supplement them. The question-answer pairs in the knowledge base are stored in the form of key-value pairs with the question as the key and the answer as the value.

S44. Establish an inverted index and a keyword index for all questions in the knowledge base, as shown in FIG. 8 and FIG. 9 respectively. When building an inverted index, use the jieba word segmentation tool to segment and remove stop words for each question text paragraph in each question-answer pair in the knowledge base to obtain a set of words, and then count all the words in the knowledge base after such processing The set of , that is, all the words in the knowledge base. First, build an index with all the words in the knowledge base as the key and the frequency of each word in each segment as the value; then adjust the index according to the dictionary of synonyms, the specific method is: if word A and Word B is a synonym for each other, so the corresponding value of the word A key in the adjusted new index is: find each segment where word A or word B appears in the index before adjustment, if word A and word B only have If one appears in this paragraph, then the value is the word frequency of the word that appears in this paragraph; if word A and word B appear in this paragraph at the same time, then the value is the sum of the word frequencies of word A and word B in this paragraph, That is, the word frequency fi of a certain word ci in a certain paragraph in the text database is recorded as:

fi=freq(ci)+∑freq(pi)

Among them, freq is to count the word frequency of this word in a certain paragraph in the text library, and pi is a synonym for ci;

When constructing the keyword index of the knowledge base, traverse each question in the knowledge base, and judge whether the words in it overlap with the set of entity names and aliases extracted by the information extraction model. The value corresponding to the entity name or alias key plus this question.

S45. Directly save a large number of unstructured knowledge texts in the field of artificial intelligence as a text library and build an inverted index for them, as shown in FIG. 10 . Use the jieba word segmentation tool to perform word segmentation and remove stop words for each text paragraph in the text library to obtain a set of words; then count the collection of words obtained after all text paragraphs in the text library have been processed in this way, that is, in the text library all words. First, build an index that uses all the words in the text library as keys, and the frequency of each word in each segment as the value; then adjust the index according to the dictionary of synonyms, the specific method is: if word A and Word B is a synonym for each other, so the corresponding value of the word A key in the adjusted new index is: find each segment where word A or word B appears in the index before adjustment, if word A and word B only have If one appears in this paragraph, then the value is the word frequency of the word that appears in this paragraph; if word A and word B appear in this paragraph at the same time, then the value is the sum of the word frequencies of word A and word B in this paragraph, That is, the word frequency fi of a certain word ci in a certain paragraph in the text database is recorded as:

fi=freq(ci)+∑freq(pi)

Among them, freq is to count the word frequency of this word in a certain paragraph in the text library, and pi is a synonym for ci.

As shown in Figure 11, in this embodiment, the specific processing process of the question and answer module is as follows:

S51. Use the input preprocessing module to remove the stop words in the user input question, and use the new words obtained by the information extraction model and the jieba word segmentation tool to perform word segmentation and part-of-speech tagging, and perform instruction based on the results of word segmentation and part-of-speech tagging and the LTP language model of Harbin Institute of Technology Generation digestion. In the resolution of anaphora, as shown in Figure 12, the part-of-speech tagging in the jieba word segmentation and the syntactic analysis of the LTP model are used to obtain the word segmentation result, part-of-speech and dependency arc relationship. First, judge whether there is a subject, if not, locate the part of speech of the pronoun in the part of speech array, if its dependent arc relationship is not a verb-object relationship (VOB) or a fixed-center relationship (ATT), replace the pronoun with the subject recorded last time; If it does not exist, locate the pronoun and judge whether its dependency arc relationship is a parallel relationship (COO). If so, replace the pronoun with the input subject. If there is no subject and pronoun, record the input subject for later reference resolution.

S52. Using the question answering module based on the knowledge base, after receiving the result obtained by the preprocessing module, first, quickly find whether there is a question exactly the same as the question in the knowledge base, and if there is, directly get the corresponding answer, which is in the user recognition system The recommended questions will be used later and can greatly save system response time and occupancy of computing resources. When no exact match is found, the preprocessed user questions will be firstly retrieved by the improved BM25 knowledge base coarse recall module in all question texts in the knowledge base, and then the preprocessed user questions will be analyzed using the ESIM model Perform fine semantic matching with the coarse search results of the question to find the result with the closest semantics. If the matching score of the result exceeds the set threshold, the answer corresponding to the result will be used as the answer to be returned to the user. Among them, the ESIM model judges whether two questions are synonymous as shown in Figure 14. The input is a batch of question pairs, and the output is the synonymous probability of each pair of questions in this batch of question pairs, that is, the matching score. If neither the rough recall module of the BM25 knowledge base nor the ESIM model finds any matching result with a matching score exceeding the threshold in this step, the preprocessed user questions are input into the question answering module based on the text library.

Specifically, the improved BM25 knowledge base coarse recall module, as shown in Figure 13, first judges whether each word in the preprocessed user input question exists in the set of all words in the knowledge base, and if so, then Reserve it directly, if it does not exist and the word has a synonym, then judge whether any synonym of the word exists in the set of all words in the knowledge base, if it exists, replace the word with the synonym and keep it, if it does not exist or the word If there is no synonym, the word is removed to obtain the user question after further processing. Finally, the correlation score of each word and each question in the knowledge base retained after the above processing is calculated, and the correlation scores of each word are summed to obtain the correlation score of the query and each question in the knowledge base.

The correlation score between the further processed user input question Q and the jth question d _j in the knowledge base is:

Among them, q _i is the i-th word in the user input question after further processing; W _ij is the weight of the j-th question in the knowledge base corresponding to the i-th word in the user input question; IDF(q _i ) is the further processed The inverse text frequency index of the i-th word in the user input question; n is the number of words in the user input question; i refers to the i-th word in the user input question;

specifically,

Among them; N is the total number of question-answer pairs in the knowledge base; n(q _i ) is the number of question-answer pairs in the knowledge base that contains the i-th vocabulary in the user input question after further processing;

specifically,

Among them, k ₁ is the term frequency saturation, which is 1.5 here; b is the paragraph length constraint, which is 0.75 here; dl _j is the string length of the jth question in the knowledge base; the average of all questions in the avgdl knowledge base String length.

S53. If the question answering module based on the knowledge base does not find the answer to the user's input question, the question answering module based on the text base is used, and the user's question after further processing is combined with the knowledge in the text base through the improved BM25 knowledge base rough recall module Text paragraphs are retrieved and rough recalled; then the rough recall results and preprocessed user questions are input into the RoBERTa-QA model, which predicts in parallel the starting and ending points of the answers to user questions in the results of each paragraph of rough recall locations and corresponding probabilities. The product of the probability values corresponding to the position with the highest probability of the starting point and the ending point is the probability that the text paragraph has the answer. text passages as the answer returned to the user.

As shown in Figure 15, in this embodiment, the improved BM25 text library rough recall module calculates the relevance scores of each word and all text paragraphs of the text library after calculating the further processed user input question, and then calculates the correlation score of each word The sum of the scores of the further processed user question and the correlation score of all text paragraphs in the text database, that is, the correlation score between the further processed user input question Q and the jth paragraph d _j in the text database is:

Among them, q _i is the i-th word in the user input question after further processing; W _ij is the weight of the j-th question in the knowledge base corresponding to the i-th word in the user input question; IDF(q _i ) is the further processed The inverse text frequency index of the i-th word in the user input question; n is the number of words in the user input question; i is the i-th word in the user input question;

specifically,

Among them; N is the total number of question-answer pairs in the knowledge base; n(q _i ) is the number of question-answer pairs in the knowledge base that contains the i-th vocabulary in the user input question after further processing;

specifically,

In W _ij , k ₁ is the term frequency saturation, which is 1.5 here; b is the text length constraint, which is 0.75 here; dl _j is the string length of the jth question in the knowledge base; all questions in the avgdl knowledge base The average string length of .

S54, using the question-and-answer recommendation module based on the knowledge base, as shown in Figure 16, this module first utilizes the knowledge base keyword index established in step S44 to search for entity names and aliases coincident with the knowledge base and preprocessed user input questions If the question can be found, randomly select a question from it and recommend it to the user, otherwise directly select a question from all the knowledge bases, check whether the question is the same as the preprocessed user question or the answer is the same as the preprocessed user question and will return The answer to the user is the same, if so, repeat the previous steps until a question that satisfies the inconsistent condition is obtained, and the obtained answer to the user's question and the question recommended to the user are returned and displayed to the user in turn.

The above-mentioned embodiment is a preferred embodiment of the present invention, but the embodiment of the present invention is not limited by the above-mentioned embodiment, and any other changes, modifications, substitutions, combinations, Simplifications should be equivalent replacement methods, and all are included in the protection scope of the present invention.

Claims (6)

1. An efficient intelligent question answering system for knowledge in the field of artificial intelligence, characterized in that it includes a preparation module and a question answering module; wherein the preparation module includes a data collection module, a model training module and a question answering system knowledge structure building module; the question answering module includes input Preprocessing module, question answering module based on knowledge base, question answering module based on text base and question recommendation module based on knowledge base; The preparation module marks the collected unstructured knowledge text paragraphs in the field of artificial intelligence through the data collection module, and trains the information extraction module and machine reading comprehension module of the model training module, and at the same time collects or defines synonymous and different The short text matching model is trained by using meaningful questions, and the knowledge structure building module of the question answering system is used to extract knowledge triples from the trained information extraction model and form question and answer pairs. Provide semantics for search engines by improving the construction method of knowledge base and text base inverted index, and build a knowledge base keyword index; The question answering module preprocesses the questions input by the user through the input preprocessing module, and uses the question answering module based on the knowledge base to search for answers. The question answering module of the library looks for and prepares to return the answer, and uses the question recommendation module based on the knowledge base to recommend questions to the user, and finally returns the answer and the recommended question to the user; The implementation process of the model training module is as follows: S31. Use the HBT model to build an information extraction model, then use the RoBERTa pre-training model to initialize model parameters, and use the marked triplet data for training; S32. Build an ESIM model to train a short text matching model, use the full corpus of Chinese Wikipedia to train Chinese character vectors using word2vec, and then use the Chinese translation results of the Quora Question Pairs dataset to pre-train the ESIM model, based on the pre-trained model Use the large-scale short text matching data set LCMQC open in Chinese and the collected short text matching data set in the field of artificial intelligence to fine-tune the parameters of the ESIM model; S33. Build the RoBERTa-QA model to train the machine reading comprehension model, then use the open data set DuReader to conduct further pre-training on the basis of the Chinese pre-training model RoBERTa, and use the collected machine reading comprehension annotation data in the field of artificial intelligence to set parameters fine-tuning training; The realization process of the question answering system knowledge structure building block is as follows: S41. Collect unstructured knowledge text paragraphs in the field of artificial intelligence; S42. Using the information extraction model trained in step S31 to extract triples in unstructured knowledge text paragraphs in the field of artificial intelligence; S43. Separately propose the entity name and entity alias content in the triple group, form the key-value pair of the question answer from the extracted triple group result, and then crawl the triple group knowledge in Baidu Encyclopedia and Wikipedia Form a question-answer pair and put it into the knowledge base. The question-answer pair in the knowledge base is stored in the form of a key-value pair with the question as the key and the answer as the value; S44. Establish inverted index and keyword index for all questions in the knowledge base. When establishing the inverted index, use the jieba word segmentation tool to perform word segmentation and remove stop words on the question text paragraphs in each question-answer pair in the knowledge base to obtain a set of words. Group words, and then count the collection of words obtained after all the questions in the knowledge base have been processed in this way, that is, all the words in the knowledge base; build a keyword index in the knowledge base, traverse each question in the knowledge base, if the words and information in it If there is an intersection between the entity name and alias set extracted by the extraction model, then this question is added to the value corresponding to the entity name or alias key in the keyword index of the knowledge base; S45. Directly save the unstructured knowledge text paragraphs in the field of artificial intelligence as a text library and establish an inverted index for it, use the jieba word segmentation tool to perform word segmentation and remove stop words for each text paragraph in the text library, and obtain a set of Words; and count all the text paragraphs in the text database after the collection of words obtained after such processing, that is, all the words in the text database. 2. a kind of high-efficiency intelligent question answering system facing artificial intelligence field knowledge according to claim 1, is characterized in that, the realization process of data collection module is as follows: S21. Collect unstructured knowledge text paragraphs of scientific publications, literature, and network popular science knowledge in the field of artificial intelligence, and the text paragraphs are split according to periods; S22. Define the key information triplet types extracted in the information extraction model, and use the common relational definition method to define the triplet types as follows: entity-description-content, entity-proposer-content, entity-include-content, Entity-Application-Content, Entity-Alias-Content, and mark the defined triplet types, use the brat text markup tool to mark by checking the subject in the collected unstructured knowledge text paragraphs in the field of artificial intelligence Display all the subjects, and then mark all the objects by checking the object, and mark all the correspondence between the subject and the object in the text paragraph through the connection line; S23. Using the machine reading comprehension model, after collecting unstructured knowledge text paragraphs of scientific publications, literature, and network popular science knowledge in the field of artificial intelligence in step S21, if there are corresponding questions, mark the questions in the text paragraphs The start and end positions of the corresponding answer-related content, otherwise, directly simulate real user scenarios to define diverse scientific knowledge questions in the field of artificial intelligence based on part of the content in the text paragraph, and mark the start and end positions of the answer corresponding to the question in the text paragraph; S24. Use the short text matching model to directly collect synonymous questions and non-synonymous questions. Two pairs of synonymous questions have a corresponding label of 1; two pairs of non-synonymous questions have a corresponding label of 0. 3. a kind of high-efficiency intelligent question answering system facing artificial intelligence domain knowledge according to claim 1, it is characterized in that, the implementation process of input preprocessing module is: utilize input preprocessing module to remove stop words in user input question, and Use the new words obtained by the information extraction model and the jieba word segmentation tool to perform word segmentation and part-of-speech tagging, and perform reference resolution based on the results of word segmentation and part-of-speech tagging and the LTP language model of Harbin Institute of Technology. 4. A kind of high-efficiency intelligent question answering system facing artificial intelligence domain knowledge according to claim 1, it is characterized in that, the realization process of the question answering module based on knowledge base is: after receiving the result that the input preprocessing module obtains, by searching If the exact same question as this question is found in the knowledge base, the corresponding answer will be obtained directly; if no exact match is found, the preprocessed user question will also be stored in the knowledge base by the improved BM25 knowledge base rough recall module Perform coarse retrieval in all question texts, and use the ESIM model to perform fine semantic matching on the preprocessed user questions and the rough retrieval results of the questions, and find the result. If the matching score of the result exceeds the set threshold, then The answer corresponding to the result is taken as the answer to be returned to the user. If neither the BM25 knowledge base rough recall module nor the ESIM model finds any matching result with a matching score exceeding the threshold in this step, the preprocessed user question is input into the In the question answering module of the text library. 5. a kind of high-efficiency intelligent question answering system facing artificial intelligence field knowledge according to claim 1, it is characterized in that, the realization process of the question answering module based on text base is as follows: first by the rough recall module of improved BM25 knowledge base will further process The final user question and the knowledge text paragraph in the text library are retrieved and rough recalled; then the rough recall result and the preprocessed user question are input into the RoBERTa-QA model, and the model predicts the result of each rough recalled paragraph in parallel. The starting point and ending position of the answer to the user's question and the corresponding probability, the product of the probability value corresponding to the position with the highest probability of the starting point and the ending point is the probability that the text paragraph has the answer, if any paragraph has the probability of having the answer exceeds the set Threshold, then the text paragraph between the starting point and the position with the highest probability of the end point in the text paragraph will be used as the answer that will be returned to the user. 6. A kind of high-efficiency intelligent question answering system facing artificial intelligence domain knowledge according to claim 1, it is characterized in that, the realization process of the question and answer recommendation module based on knowledge base is as follows: utilize the knowledge base keyword index that step S44 sets up to search and If there are overlapping entity names and aliases in the knowledge base and preprocessed user input questions, if they can be found, randomly select questions from them and recommend them to the user; otherwise, randomly select questions from all knowledge bases and check after the selection is complete. Is the question the same as the preprocessed user question or the answer is the same as the answer that will be returned to the user? If so, repeat the previous steps until a question that meets the inconsistent conditions is obtained, and the obtained answer and recommendation about the user question Questions given to the user are in turn returned for display to the user.