Intelligent question answering system oriented to bid invitation document question answering scene

Abstract

The invention discloses an intelligent question-answering system oriented to a bid invitation file question-answering scene, which is based on artificial intelligence and a large model technology, performs knowledge extraction on two different forms of data of a text and a table of a bid invitation file to form training data, and further generates questions and corresponding answers by utilizing the large model technology. The accuracy of question answering is improved through self-learning, the matching rate of consultation questions is improved, the accuracy of intelligent question answering is improved, and the bidding and tendering efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and more specifically to an intelligent question-answering system for a bidding document question-answering scenario. Background Art

[0002] With the rapid advancement of digital transformation, the way of answering questions in tender announcements has been continuously upgraded. Traditional Q&A requires manual answers one by one, but with the introduction of natural language processing (NLP) technology, intelligent Q&A has become more important. This method can automatically answer based on the knowledge base, thereby providing a faster response speed and reducing the pressure of manual Q&A work. However, the current intelligent question system is mainly based on matching technology, which automatically identifies questions for matching and returns set answers, but cannot answer when the consulting question is out of range. In addition, different descriptions of the same question will reduce the matching rate with the knowledge base, and the research on the semantic understanding of the question is relatively insufficient. After the emergence of large language models, the level of understanding of long articles and complex contexts has been significantly improved, and the ability to automatically summarize and summarize knowledge has been comprehensively strengthened, making the intelligent question-answering library based on language models a feasible and more effective choice. To solve this problem, this patent combines a large model to propose an intelligent question-answering system for tender document question-answering scenarios, improve the matching rate of consulting questions, improve the accuracy of intelligent questions and answers, and improve the efficiency of bidding. Summary of the invention

[0003] In order to solve the above problems, the present invention provides an intelligent question and answer system for the bidding document question and answer scenario, which can extract knowledge from the content of the bidding document, automatically generate questions and answers in combination with a large model, and use self-learning technology to improve the accuracy of answers to questions.

[0004] The present invention discloses an intelligent question-answering system for a bidding document question-answering scenario, comprising:

[0005] Bidding document knowledge extraction module: knowledge extraction is performed on the text and table data of the bidding document to form training data. For text information, the present invention first uses the G l ove model based on binary loss to generate contextual embedding representations and graph embedding representations of named entities, then constructs a BERT encoder and a graph convolutional encoder, dynamically initializes the graph attention mechanism of the graph convolutional encoder using the transformer self-attention weight in the BERT encoder, and inputs the word embedding representation and the graph embedding representation into the BERT encoder and the initialized graph convolutional encoder one by one, uses the BERT encoder to perform feature extraction on the word embedding representation to obtain the local contextual feature representation of the named entity, uses the graph convolutional encoder to perform feature extraction on the graph embedding representation to obtain the dynamic global feature representation of the named entity; finally, the contextual feature representation and the global feature representation are feature fused to obtain the enhanced feature representation, and the enhanced features are sequence labeled based on the conditional random field model to obtain the key knowledge of the text data. For the tabular information, the present invention firstly models the basic bidding knowledge into ontology, then uses the pdfplumber library in Python to implement the layout analysis of the bidding documents, identifies and reconstructs the table structure, parses the data structure of the complex table, and extracts the feature entities of these data through the random forest algorithm, finally realizing the structured information extraction of the bidding document table.

[0006] The automatic Q&A generation module based on the bidding document first generates a question summary, uses BERT as a pre-trained large language model, converts the question and dependency into a vector representation for dependency weight calculation and summary generation, uses the A-GCN dependency weighted network to perform weighted calculation on the input dependency matrix, obtains the question vector representation with weight information, then uses the feature A-GCN network to calculate the weight, and performs maximum pooling on the feature vector to highlight the key features of the question vector representation. Next, T5-Pegasus is used as a pre-trained model to generate a summary of the optimized weighted question. ; Then generate answers to questions. First, perform similarity matching on the input bidding questions and the questions in the initial question-answer library one by one to obtain similarity scores, and continuously refresh the highest score and corresponding answers through a loop mechanism. Secondly, after the matching is completed, judge the highest score obtained, set a threshold, and return answers that exceed the threshold to the user. For questions that do not exceed the threshold, use the big model to generate answers. After entity extraction, map them according to the defined question template, extract relevant content in the knowledge graph as domain knowledge, and then send the domain knowledge as context information together with the question into the generative big model for answer generation.

[0007] Consulting Q&A self-learning module: In order to improve the accuracy of the intelligent question-answering system, this patent uses the self-learning method to update the knowledge base and reduce the errors that are easy to occur in the reasoning process of large models. First, a timestamp is assigned when reading the knowledge base. After the user asks a question, the question and answer pair is stored in the temporary library and also assigned a timestamp. When the question-answering framework actively asks the user for feedback on whether the answer is accurate, it is converted into a binary classification question, and the user's feedback result label 1 (satisfied) and label 0 (unsatisfied) are used as storage logic to control the update of the knowledge base. When the label is 0, the question and answer pair is deleted from the temporary library. When the label is 1, the knowledge base is incrementally updated according to the difference in timestamps between the temporary library and the knowledge base, and the current timestamp of the knowledge base is updated to improve the accuracy of the answer.

[0008] Compared with the prior art, the present invention has the following beneficial effects:

[0009] The present invention is based on artificial intelligence and big model technology, and extracts knowledge from two different forms of data, namely text and table in bidding documents, to form training data, and then uses big model technology to generate questions and corresponding answers, and self-learns to improve the accuracy of answering questions, improve the matching rate of consulting questions, improve the accuracy of intelligent question and answer, and improve the efficiency of bidding. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is an implementation flow chart of the present invention. DETAILED DESCRIPTION

[0011] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0012] The present invention provides an intelligent question-answering system for bidding document question-answering scenarios, including:

[0013] Bidding document knowledge extraction module: knowledge extraction is performed on the text and table data of the bidding document to form training data. For text information, the present invention first uses the G l ove model based on binary loss to generate contextual embedding representations and graph embedding representations of named entities, and then constructs a BERT encoder and a graph convolutional encoder, and uses the pre-trained transformer self-attention weights in the BERT encoder to dynamically initialize the graph attention mechanism of the graph convolutional encoder, and inputs the word embedding representation and the graph embedding representation into the BERT encoder and the initialized graph convolutional encoder one by one, and uses the BERT encoder to extract features from the word embedding representation to obtain the local contextual feature representation of the named entity, and uses the graph convolutional encoder to extract features from the graph embedding representation to obtain the dynamic global feature representation of the named entity; finally, the contextual feature representation and the global feature representation are feature fused to obtain the enhanced feature representation, and the enhanced features are sequence labeled based on the conditional random field model to obtain the key knowledge of the text data. For the tabular information, the present invention firstly models the basic knowledge of bidding ontologies, and then uses the pdfp l umber library in python to realize the layout analysis of bidding documents, identify and reconstruct the table structure, parse the data structure of complex tables, and extract feature entities from these data through the random forest algorithm. The random forest parameters are set to 20 decision trees, and finally the structured information extraction of bidding document tables is realized.

[0014] The automatic Q&A generation module based on the bidding document first generates a question summary, uses BERT as a pre-trained large language model, converts the question and dependency into a vector representation for dependency weight calculation and summary generation, uses the A-GCN dependency weighted network to perform weighted calculation on the input dependency matrix, obtains the question vector representation with weight information, then uses the feature A-GCN network to calculate the weight, and performs maximum pooling on the feature vector to highlight the key features of the question vector representation, then uses T5-Pegasus as a pre-trained model to generate a summary of the optimized weighted question; then To generate answers to questions, firstly, the similarity of the input bidding questions is matched with the questions in the initial question-answer library one by one to obtain the similarity score, and the highest score and the corresponding answer are continuously refreshed through a loop mechanism. Secondly, after the matching is completed, the highest score obtained is judged, and the threshold is set to 0.9. Answers that exceed the threshold are returned to the user. For questions that do not exceed the threshold, the big model is used to generate answers. After entity extraction, they are mapped according to the defined question template, and the relevant content in the knowledge graph is extracted as domain knowledge. Then the domain knowledge is sent to the generative big model together with the question as context information for answer generation.

[0015] Consulting Q&A self-learning module: In order to improve the accuracy of the intelligent question-answering system, this patent uses the self-learning method to update the knowledge base and reduce the errors that are easy to occur in the reasoning process of large models. First, a timestamp is assigned when reading the knowledge base. After the user asks a question, the question and answer pair is stored in the temporary library and also assigned a timestamp. When the question-answering framework actively asks the user for feedback on whether the answer is accurate, it is converted into a binary classification question, and the user's feedback result label 1 (satisfied) and label 0 (unsatisfied) are used as storage logic to control the update of the knowledge base. When the label is 0, the question and answer pair is deleted from the temporary library. When the label is 1, the knowledge base is incrementally updated according to the difference in timestamps between the temporary library and the knowledge base, and the current timestamp of the knowledge base is updated to improve the accuracy of the answer.

[0016] The above are only some embodiments of the present invention. It should be pointed out that for ordinary technicians in this field, other modifications and improvements can be made without departing from the creative concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. An intelligent question-answering system for bidding document question-answering scenarios, characterized in that: include: The bidding document knowledge extraction module can extract knowledge from two different forms of data, text and table, in the bidding document to form training data; The automatic Q&A generation module based on the bidding documents can generate questions and corresponding answers using big model technology; The consulting question and answer self-learning module can self-learn to improve the accuracy of answering questions, increase the matching rate of consulting questions, improve the accuracy of intelligent questions and answers, and improve bidding efficiency.

2. The system according to claim 1, characterized in that The specific processing method of the bidding document knowledge extraction module is as follows: for text information, first use the Glove model based on binary loss to generate context embedding representation and graph embedding representation of named entities, then construct a BERT encoder and a graph convolution encoder, use the transformer self-attention weight in the BERT encoder to dynamically initialize the graph attention mechanism of the graph convolution encoder, and input the word embedding representation and the graph embedding representation into the BERT encoder and the initialized graph convolution encoder one by one, use the BERT encoder to extract features from the word embedding representation to obtain the local context feature representation of the named entity, use the graph convolution encoder to extract features from the graph embedding representation to obtain the dynamic global feature representation of the named entity; finally, perform feature fusion on the context feature representation and the global feature representation to obtain the enhanced feature representation, and perform sequence annotation on the enhanced features based on the conditional random field model to obtain the key knowledge of text data; for table information, the present invention first performs ontology modeling on the basic bidding knowledge, and then uses the pdfplumber library in python to realize the layout analysis of the bidding document, identify and reconstruct the table structure, parse the data structure of the complex table, and extract feature entities from these data through the random forest algorithm, and finally realize the structured information extraction of the bidding document table.

3. The system according to claim 1, characterized in that The specific processing method of the Q&A automatic generation module based on the bidding document is as follows: first generate a question summary, use BERT as a pre-trained large language model, convert the question and dependency relationship into a vector representation for dependency weight calculation and summary generation, use the A-GCN dependency weighted network to perform weighted calculation on the input dependency matrix, obtain the question vector representation with weight information, then use the feature A-GCN network to calculate the weight, and perform maximum pooling on the feature vector to highlight the key features of the question vector representation, then use T5-Pegasus as a pre-trained model to perform weighted summarization on the optimized question with weights. Summary generation; then generate the answer to the question. First, the input bidding question is matched with the questions in the initial question and answer library one by one to obtain the similarity score, and the highest score and the corresponding answer are continuously refreshed through a loop mechanism. Secondly, after the matching is completed, the highest score obtained is judged, a threshold is set, and the answer that exceeds the threshold is returned to the user. For questions that do not exceed the threshold, the big model is used to generate answers. After entity extraction, it is mapped according to the defined question template, and the relevant content in the knowledge graph is extracted as domain knowledge. Then the domain knowledge is sent as context information together with the question into the generative big model for answer generation.

4. The system according to claim 1, characterized in that The specific processing method of the question and answer self-learning module is: to update the knowledge base by borrowing the self-learning method, to reduce the errors that are easy to occur in the large model reasoning process, first assign a timestamp when reading the knowledge base, and after the user asks a question, store the question and answer pair in the temporary library and also assign a timestamp, when the question and answer framework actively asks the user for feedback on whether the answer is accurate, it is converted into a binary classification problem, and the user's feedback result label 1 and label 0 are used as storage logic to control the update of the knowledge base, label 1 is satisfied, label 0 is dissatisfied, when the label is 0, the question and answer pair is deleted from the temporary library, when the label is 1, the knowledge base is incrementally updated according to the timestamp difference between the temporary library and the knowledge base, and the current timestamp of the knowledge base is updated to improve the accuracy of the answer.

Images