Self-adaptive fuzzy query intention disassembling and dynamic optimization method

Abstract

The invention discloses a self-adaptive fuzzy query intention disassembling and dynamic optimization method, which comprises the following steps of: obtaining a query problem input by a user, and understanding and analyzing the input of the user by using a natural language processing technology; generating a hypothesis document for answering the question input by the user based on the user fuzzy question and the question keyword by using a large language model; comparing the hypothesis document with documents in a preset domain knowledge base so as to find domain information related to the user question; an auto-reflection mechanism is introduced, a large language model or a document similarity comparison model is used, correlation between documents in a knowledge base and user fuzzy problems is compared, intention splitting is conducted again based on an evaluation result, and an original problem is decomposed into a plurality of smaller and more specific sub-problems; and presenting the disassembled sub-questions to a user, and performing man-machine interaction to improve the disassembling result. According to the method, the query intention of the user can be accurately captured, and the accuracy and efficiency of information retrieval are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent information retrieval, and more specifically, to a method for disassembling and dynamically optimizing adaptive fuzzy query intentions. Background Art

[0002] With the rapid development of information technology, users' demand for information retrieval is also growing. In this context, the application of intelligent assistants, chatbots and various intelligent systems is also growing. However, when asking questions, users often cannot accurately express their query intentions, resulting in the inability of search engines to accurately return relevant information. This is particularly evident when users ask ambiguous questions, because the uncertainty of ambiguous questions makes it difficult for search engines to capture the user's true intentions. User intention decomposition, as an important branch of natural language processing, aims to enable computers to accurately understand and process users' query intentions in order to accurately answer users' questions, which is crucial to improving the naturalness and efficiency of human-computer interaction. Traditional intent recognition methods, such as rule-based pattern matching or some relatively basic machine learning models, have played an important role in the field of natural language processing. These methods can indeed show certain effects when processing simple and direct instructions. They identify user intentions by setting clear rules or patterns learned from large amounts of data, providing machines with the ability to initially understand human language. However, when faced with more complex language structures, such as long sentences, nested clauses, or expressions containing metaphors and hints, the recognition performance of these traditional methods is stretched. Especially when it is necessary to identify implicit intent or deal with situations that are closely related to the context, they often have difficulty accurately capturing the user's true intent, thus limiting the fluency and accuracy of human-computer interaction. Therefore, how to more accurately capture the user's query intent and improve the accuracy and efficiency of information retrieval is of great significance. Summary of the invention

[0003] The present invention provides a method for disassembling and dynamically optimizing adaptive fuzzy query intentions, which solves the problem that the user's query intentions cannot be accurately understood and processed during existing information retrieval, can accurately capture the user's query intentions, and improve the accuracy and efficiency of information retrieval.

[0004] To achieve the above objectives, the present invention provides the following technical solutions:

[0005] A method for disassembling and dynamically optimizing adaptive fuzzy query intentions, comprising:

[0006] Step 1: Obtain the query question input by the user and use natural language processing technology to understand and parse the user input, including semantic analysis and sentiment analysis, to accurately capture the core of the question;

[0007] Step 2: Using a large language model, based on the user's fuzzy question and question keywords, generate a hypothetical document for answering the user's input question, wherein the hypothetical document contains possible query intent and related information;

[0008] Step 3: Compare the hypothetical document with the documents in the preset domain knowledge base to find domain information related to the user's question;

[0009] Step 4: Introduce a self-reflection mechanism, use a large language model or a document similarity comparison model to compare the relevance of documents in the knowledge base with the user's fuzzy questions, and re-execute steps 2 and 3 based on the evaluation results;

[0010] Step 5: Submit domain-related knowledge documents and user fuzzy questions to a large language model, set intention decomposition prompt words, and decompose the original question into multiple smaller and more specific sub-questions;

[0011] Step 6: Present the decomposed sub-problems to the user for human-computer interaction to improve the decomposition results.

[0012] Preferably, the steps to capture the core of the problem include:

[0013] Pre-process the query questions entered by users to correct spelling errors and grammatical errors to improve the accuracy of subsequent processing;

[0014] Segment the query and remove stop words;

[0015] Using a semantic analysis model to identify keywords in the query question, the keywords represent the main intent and focus of the user's query;

[0016] Combined with sentiment analysis models, the sentiment tendencies in user queries are analyzed to more comprehensively understand the user's query intentions.

[0017] Preferably, the step of generating a hypothesis document includes:

[0018] Taking user fuzzy queries and keywords as input, a large language model is used to answer user questions and generate a series of hypothetical documents as an example of the query to better understand the user's query intent;

[0019] The text vectorization model is used to encode the hypothetical document, and the corresponding vector is calculated for each user query question.

[0020] Preferably, comparing the hypothetical document with documents in a preset domain knowledge base to find domain information related to the user's question includes:

[0021] For different user objects, collect and organize relevant domain documents and materials to build a relevant domain knowledge base;

[0022] A text vectorization model is used to convert documents in the knowledge base into numerical vectors and store them in a vector database to facilitate subsequent query and comparison;

[0023] Using the semantic similarity model, domain knowledge related to the user's query question is found from the domain knowledge base and used as candidate documents.

[0024] Preferably, the use of a large language model or a document similarity comparison model to compare the relevance between the documents in the knowledge base and the user's fuzzy questions includes:

[0025] For the retrieval results, a large language model is used to judge whether the retrieved candidate documents are relevant to the user's query question and to score the retrieval results.

[0026] Preferably, re-executing step 2 and step 3 based on the evaluation result includes:

[0027] If the retrieved candidate documents are irrelevant to the user's query, the user's fuzzy query is rewritten using a large language model, and steps 2 and 3 are re-executed;

[0028] If the retrieved candidate documents are relevant to the user's query, the subsequent intent decomposition steps are performed.

[0029] Preferably, presenting the decomposed sub-problems to the user for human-computer interaction to improve the decomposition results includes:

[0030] Present the broken-down sub-questions to the user, and proactively ask the user whether the sub-questions accurately reflect the user's concerns or whether there are other specific aspects required;

[0031] According to the modification suggestions given by the user on the sub-problems, the disassembly results are refined or adjusted.

[0032] Preferably, presenting the decomposed sub-problems to the user for human-computer interaction to improve the decomposition results further includes:

[0033] For the modification suggestion Q given by the user, construct a prompt text: Prompt={Inst,E,Q}, where Inst is the explanation text consisting of description, usage and interactive format, E represents a set of examples used to guide the model to answer, and two complete examples constructed in interactive format for each type of question;

[0034] In each round of interaction, let the large language model generate an action a based on the Prompt and the interaction history H. T , then according to action aT Make corresponding adjustments and modifications to the sub-questions.

[0035] Preferably, the large language model generates an action a according to the Prompt and the interaction history H. T , including the following processes:

[0036] a T =LLM({Prompt,H});

[0037] H={c0,a0,o0,…,c (T-1) ,a (T-1) ,o (T-1)};

[0038] Among them, c represents the intermediate thinking process, action a∈{modify, delete, add}, and o is the last output result.

[0039] Preferably, it also includes: presenting all sub-questions to the user again, including modified and unmodified sub-questions, resubmitting user feedback to the large language model, and further improving the decomposition results to obtain more accurate intent decomposition results.

[0040] The present invention provides a method for disassembling and dynamically optimizing adaptive fuzzy query intentions, which accurately captures and analyzes the user's fuzzy query intentions by combining natural language processing technology, large language models and domain knowledge bases, and disassembles them into multiple specific sub-questions, interacts with the user, and makes real-time adjustments based on the user's feedback, so that the final information retrieval results are more in line with the user's actual needs. The problem that the user's query intentions cannot be accurately understood and processed during existing information retrieval is solved, and the user's query intentions can be accurately captured, thereby improving the accuracy and efficiency of information retrieval. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the specific embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below.

[0042] Figure 1 It is a schematic diagram of a method for disassembling and dynamically optimizing an adaptive fuzzy query intention provided by the present invention. DETAILED DESCRIPTION

[0043] In order to enable persons skilled in the art to better understand the solutions of the embodiments of the present invention, the embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings and implementation modes.

[0044] In view of the problem that the user's query intention cannot be accurately understood and processed during current information retrieval, the present invention provides a method for disassembling and dynamically optimizing adaptive fuzzy query intention, which solves the problem that the user's query intention cannot be accurately understood and processed during existing information retrieval, can accurately capture the user's query intention, and improve the accuracy and efficiency of information retrieval.

[0045] like Figure 1 As shown, a method for disassembling and dynamically optimizing adaptive fuzzy query intent includes:

[0046] Step 1: Obtain the query question input by the user and use natural language processing technology to understand and parse the user input, including semantic analysis and sentiment analysis, to accurately capture the core of the question;

[0047] Step 2: Using a large language model, based on the user's fuzzy question and question keywords, generate a hypothetical document for answering the user's input question, wherein the hypothetical document contains possible query intent and related information;

[0048] Step 3: Compare the hypothetical document with the documents in the preset domain knowledge base to find domain information related to the user's question;

[0049] Step 4: Introduce a self-reflection mechanism, use a large language model or a document similarity comparison model to compare the relevance of documents in the knowledge base with the user's fuzzy questions, and re-execute steps 2 and 3 based on the evaluation results;

[0050] Step 5: Submit domain-related knowledge documents and user fuzzy questions to a large language model, set intention decomposition prompt words, and decompose the original question into multiple smaller and more specific sub-questions;

[0051] Step 6: Present the decomposed sub-problems to the user for human-computer interaction to improve the decomposition results.

[0052] Specifically, natural language processing technology is used to perform multi-dimensional analysis of fuzzy queries input by users, including semantic analysis and sentiment analysis, to ensure that the user's query intention can be fully and accurately understood. At the same time, deep learning models are used to extract keywords from queries. These keywords not only represent the main content of the user's query, but also reflect the user's focus and interest. Combined with the generation capability of large language models and relying on domain-related knowledge, the accuracy of intent capture is improved. The system can adaptively adjust the query intent decomposition strategy and retrieval parameters according to changes in different users, different queries, and different fields, ensuring that high retrieval accuracy can be maintained in different scenarios.

[0053] Through a self-reflection mechanism, the retrieval results are evaluated in real time, and the query intention decomposition strategy and retrieval parameters are dynamically adjusted according to the evaluation results to continuously optimize the retrieval process. At the same time, in the form of human-computer interaction, using the chain of thought method, interact with users to obtain user feedback and dynamically adjust the intention decomposition results.

[0054] By presenting the decomposed sub-questions to the user in a clear and understandable way, the user can intuitively understand the query intention decomposition process, enhancing the user's trust in the system. Provide a human-computer interaction interface that allows users to give feedback on the sub-questions, and the system further adjusts the decomposition results according to the feedback to ensure that the final information retrieval results meet the actual needs of the user.

[0055] In one embodiment, an adaptive fuzzy query intention decomposition and dynamic optimization method using the HyDE algorithm includes the following steps:

[0056] S1. Preprocess the fuzzy query input by the user and extract the problem core.

[0057] S1.1. Based on a text correction model, correct spelling and grammar errors in the fuzzy query.

[0058] S1.2. Perform basic operations such as word segmentation and stop word removal on the query problem. Word segmentation is to split the sentence into words, and stop word removal is to remove common words such as "de" (of) and "shi" (is).

[0059] S1.3. Use a semantic analysis model to identify the keywords in the query problem, and at the same time generate a series of implicit related keywords, which represent the main intentions and concerns of the user's query.

[0060] S1.4. Use a sentiment analysis model to analyze the sentiment tendency in the user's query to more comprehensively understand the user's query intention.

[0061] In practical applications, HyDE, short for Hypothetical Document Embeddings, is a method for improving information retrieval. It aligns the semantic spaces of queries and documents by generating hypothetical documents that can be used to answer the questions input by users. These documents are generated by large language models based on the knowledge they have learned themselves, providing a new approach for information retrieval. The semantic analysis model can use BERT, RoBERTa, etc. to identify the keywords in the query problem. The sentiment analysis model can use VADER, TextBlob, etc. to analyze the sentiment tendency in the user's query.

[0062] S2. Generate hypothetical documents;

[0063] S2.1. Taking user fuzzy queries and keywords as input, a large language model is used to answer user questions and generate a series of hypothetical documents as an example of the query to help better understand the user's query intent.

[0064] S2.2. Use an unsupervised contrastive learning encoder, such as Contriever, to encode the hypothetical document. For each user query question q, calculate its corresponding d-dimensional vector v q , that is, v q =f(q,D f ), where f is an unsupervised contrastive learning encoder, i.e., a text vectorization model, and D f is the hypothetical document corresponding to the query question q.

[0065] S3, domain-related knowledge retrieval;

[0066] S3.1. For different user objects, collect and organize relevant domain documents and materials to build a domain-related knowledge base.

[0067] S3.2. Use vectorization models (such as TF-IDF, Word2Vec, BERT, etc.) to convert documents in the knowledge base into numerical vectors and store them in vector databases (such as Elasticsearch, Milvus, etc.) to facilitate subsequent queries and comparisons.

[0068] S3.3. Use the inner product to calculate v q The similarity between d1 and all document vectors is used to find the most relevant document. n ], define a matrix V, where the i-th row corresponds to the vector of the i-th document. Then, for a query question q, and its corresponding vector v q , use a similarity function (such as cosine similarity) to calculate similar documents, sort the documents according to the similarity, and get the most relevant document list.

[0069] S3.4. Use a multimodal large language model, set image parsing prompt words, parse the illustrations in the document, and associate the illustration title and context based on the coordinate information identified by the layout.

[0070] S4, self-reflection on search results;

[0071] S4.1. For the search results, use a large language model to judge whether the candidate documents are relevant to the user's query and score the search results.

[0072] S4.2. Determine whether to re-execute step 2 and step 3 based on the evaluation results.

[0073] S4.2.1. If the retrieved candidate documents are irrelevant to the user's query, use a large language model to rewrite the user's fuzzy query and re-execute steps 2 and 3.

[0074] S4.2.1. If the retrieved candidate document is relevant to the user's query, perform the subsequent intent decomposition steps.

[0075] S5. Decompose user fuzzy questions. Submit domain-related knowledge documents, user fuzzy questions and keywords to a large language model to decompose the original question into multiple smaller and more specific sub-questions.

[0076] S6, Human-computer interaction;

[0077] S6.1. Present the decomposed sub-questions to the user and actively ask the user whether these sub-questions accurately reflect their concerns or whether there are other specific aspects that need further exploration by the model. The model further adopts the thinking chain method to interact with the user.

[0078] S6.2. After reviewing these sub-questions, users can provide feedback to the model, indicating which sub-questions meet their actual needs and which need further refinement or adjustment.

[0079] S6.3. For the modification suggestion Q given by the user, construct a prompt text: Prompt={Inst,E,Q}, where Inst is the description text consisting of description, usage and interaction format, which is database-specific and pre-written. E represents a set of examples used to guide the model to answer. For each type of question, two complete examples are constructed in an interactive format. In each round of interaction, let the large language model generate an action based on Prompt and the interaction history H. Specifically, this process is:

[0080] a T =LLM({Prompt,H});

[0081] H={c0,a0,o0,…,c (T-1) ,a (T-1) ,o (T-1)};

[0082] Among them, c represents the intermediate thinking process, action a∈{modify, delete, add}, and o is the last output result. Except for the first round, in each subsequent round of interaction, the large language model needs to generate a thought c to clearly explain its reasoning process based on the last output result and generate the action a of the current round. T , and then make corresponding adjustments and modifications to the sub-problems based on the actions.

[0083] S6.4. Present all sub-questions to the user again, including those that have been modified and those that have not been modified. Submit the user feedback back to the large language model to further improve the decomposition results to obtain more accurate intent decomposition results.

[0084] It can be seen that the present invention provides a method for disassembling and dynamically optimizing adaptive fuzzy query intentions, which accurately captures and analyzes the user's fuzzy query intentions by combining natural language processing technology, large language models and domain knowledge bases, and disassembles them into multiple specific sub-questions, interacts with the user, and makes real-time adjustments based on the user's feedback, so that the final information retrieval results are more in line with the user's actual needs. It solves the problem that the user's query intentions cannot be accurately understood and processed during existing information retrieval, can accurately capture the user's query intentions, and improve the accuracy and efficiency of information retrieval.

[0085] The above describes in detail the structure, features and effects of the present invention based on the embodiments shown in the drawings. The above is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the drawings. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments with equivalent changes, which still do not exceed the spirit covered by the description and drawings, should be within the protection scope of the present invention.

Claims

1. A method for disassembling and dynamically optimizing adaptive fuzzy query intent, characterized in that: include: Step 1: Obtain the query question input by the user and use natural language processing technology to understand and parse the user input, including semantic analysis and sentiment analysis, to accurately capture the core of the question; Step 2: Using a large language model, based on the user's fuzzy question and question keywords, generate a hypothetical document for answering the user's input question, wherein the hypothetical document contains possible query intent and related information; Step 3: Compare the hypothetical document with the documents in the preset domain knowledge base to find domain information related to the user's question; Step 4: Introduce a self-reflection mechanism, use a large language model or a document similarity comparison model to compare the relevance of documents in the knowledge base with the user's fuzzy questions, and re-execute steps 2 and 3 based on the evaluation results; Step 5: Submit domain-related knowledge documents and user fuzzy questions to a large language model, set intention decomposition prompt words, and decompose the original question into multiple smaller and more specific sub-questions; Step 6: Present the decomposed sub-problems to the user for human-computer interaction to improve the decomposition results.

2. The method for disassembling and dynamically optimizing adaptive fuzzy query intention according to claim 1 is characterized in that: Steps to getting to the heart of the problem include: Pre-process the query questions entered by users to correct spelling errors and grammatical errors to improve the accuracy of subsequent processing; Segment the query and remove stop words; Using a semantic analysis model to identify keywords in the query question, the keywords represent the main intent and focus of the user's query; Combined with sentiment analysis models, the sentiment tendencies in user queries are analyzed to more comprehensively understand the user's query intentions.

3. The method for disassembling and dynamically optimizing adaptive fuzzy query intention according to claim 2 is characterized in that: The steps to generate an assumptions document include: Taking user fuzzy queries and keywords as input, a large language model is used to answer user questions and generate a series of hypothetical documents as an example of the query to better understand the user's query intent; The text vectorization model is used to encode the hypothetical document, and the corresponding vector is calculated for each user query question.

4. The method for disassembling and dynamically optimizing the adaptive fuzzy query intention according to claim 3 is characterized in that: The step of comparing the hypothetical document with documents in a preset domain knowledge base to find domain information related to the user's question includes: For different user objects, collect and organize relevant domain documents and materials to build a relevant domain knowledge base; A text vectorization model is used to convert documents in the knowledge base into numerical vectors and store them in a vector database to facilitate subsequent query and comparison; Using the semantic similarity model, we can find domain knowledge related to the user's query from the domain knowledge base and use it as candidate documents.

5. The method for disassembling and dynamically optimizing the adaptive fuzzy query intention according to claim 4 is characterized in that: The use of a large language model or a document similarity comparison model to compare the relevance between documents in the knowledge base and the user's fuzzy questions includes: For the retrieval results, a large language model is used to judge whether the retrieved candidate documents are relevant to the user's query question and to score the retrieval results.

6. The method for disassembling and dynamically optimizing the adaptive fuzzy query intention according to claim 5 is characterized in that: The re-execution of step 2 and step 3 based on the evaluation result includes: If the retrieved candidate documents are irrelevant to the user's query, the user's fuzzy query is rewritten using a large language model, and steps 2 and 3 are re-executed; If the retrieved candidate documents are relevant to the user's query, the subsequent intent decomposition steps are performed.

7. The method for disassembling and dynamically optimizing the adaptive fuzzy query intention according to claim 6 is characterized in that: The disassembled sub-problems are presented to the user for human-computer interaction to improve the disassembly results, including: Present the broken-down sub-questions to the user, and proactively ask the user whether the sub-questions accurately reflect the user's concerns or whether there are other specific aspects required; According to the modification suggestions given by the user to the sub-problems, the disassembly results are refined or adjusted.

8. The method for disassembling and dynamically optimizing the adaptive fuzzy query intention according to claim 7 is characterized in that: The presenting the decomposed sub-problems to the user for human-computer interaction to improve the decomposition results also includes: For the modification suggestion Q given by the user, construct a prompt text: Prompt={Inst,E,Q}, where Inst is the explanation text consisting of description, usage and interactive format, E represents a set of examples used to guide the model to answer, and two complete examples constructed in interactive format for each type of question; In each round of interaction, let the large language model generate an action a based on the Prompt and the interaction history H. T , then according to action a T Make corresponding adjustments and modifications to the sub-questions.

9. The method for disassembling and dynamically optimizing the adaptive fuzzy query intention according to claim 8 is characterized in that: The large language model generates an action a based on Prompt and interaction history H. T , including the following processes: a T =LLM({Prompt,H}); H={c0,a0,o0,…,c (T-1) ,a (T-1) ,o (T-1) }; Among them, c represents the intermediate thinking process, action a∈{modify, delete, add}, and o is the last output result.

10. The method for disassembling and dynamically optimizing the adaptive fuzzy query intention according to claim 9 is characterized in that: Also includes: All sub-questions are presented to the user again, including modified and unmodified sub-questions, and the user feedback is resubmitted to the large language model to further improve the decomposition results to obtain more accurate intent decomposition results.

Images