Knowledge retrieval-oriented government affair document enhanced generation method, electronic equipment and medium

Abstract

The invention belongs to the technical field of natural language processing, and discloses a knowledge retrieval-oriented government affair document enhanced generation method, electronic equipment and a medium, and the knowledge retrieval-oriented government affair document enhanced generation method comprises the following steps: identifying a title of a government affair document; segmenting the text according to titles to obtain a plurality of text blocks with complete semantics; semantic segmentation is carried out on the text blocks with complete semantics to obtain a plurality of medium-granularity text blocks, and the medium-granularity text blocks are text blocks conforming to a preset length; and establishing an index according to the plurality of medium-granularity text blocks. By identifying the title and segmenting the text according to the title structure, the text can be effectively subjected to first-level semantic segmentation according to the chapter structure, the internal semantic relevance and integrity are ensured, and the retrieval accuracy is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of natural language processing, and specifically relates to a government document enhancement generation method, electronic equipment and medium for knowledge retrieval. Background Art

[0002] Retrieval-Augmented Generation (RAG) technology is a method that combines information retrieval and generation technologies. It uses a large model with strong semantic understanding and few-sample learning capabilities to identify, summarize and condense the retrieved information related to the question, and then respond to the question, thereby improving the performance of the natural language generation system.

[0003] In related technologies, the text slicing method mainly divides the text into semantic segments to maximize the semantic relevance within the block. In fact, the semantic segmentation-based method may still have cross-structured text, resulting in poor semantic segmentation effect. In the scenario of large-model retrieval and enhanced question-answering, there are problems such as incoherent document slicing semantics and inaccurate retrieval. Summary of the invention

[0004] In view of this, the present invention provides a government document enhancement generation method, electronic device and storage medium for knowledge retrieval to solve the problems of incoherent document slice semantics and inaccurate retrieval in large-model retrieval enhanced question-answering scenarios in the prior art.

[0005] In a first aspect, the present invention provides a method for enhancing the generation of government documents for knowledge retrieval, comprising: identifying the title of the government document; segmenting the text according to the title to obtain a plurality of semantically complete text blocks; semantically segmenting the semantically complete text blocks to obtain a plurality of medium-granularity text blocks, wherein the medium-granularity text blocks are text blocks that meet a preset length; and establishing an index based on the plurality of medium-granularity text blocks.

[0006] In some embodiments, the text is segmented according to the title to obtain multiple semantically complete text blocks, specifically including: if the title has hierarchical titles, the title information of each level of the title is identified separately; and the title information of multiple levels is merged to obtain the multiple semantically complete text blocks.

[0007] In some embodiments, the semantically complete text block is semantically segmented to obtain multiple medium-granularity text blocks, specifically including: if the length of the semantically complete text block is greater than the preset length, segmenting it according to punctuation to obtain single sentences; and semantically merging adjacent single sentences to obtain the medium-granularity text block.

[0008] In some embodiments, adjacent single sentences are semantically merged to obtain the medium-granularity text block, specifically including: using the first sentence after the end of the previous merged block to merge forward semantically; overlapping blocks semantically; wherein the maximum length of the merged block does not exceed the preset length.

[0009] In some implementations, establishing an index based on the multiple medium-granularity text blocks specifically includes: hierarchically summarizing the multiple medium-granularity text blocks to obtain multiple upper-layer blocks; simulating user abstract questions based on the medium-granularity text blocks and the upper-layer blocks; and establishing an index based on the user abstract questions.

[0010] In some implementations, any of the methods for enhancing the generation of government documents for knowledge retrieval further includes: acquiring propositions based on the multiple medium-granularity text blocks to obtain multiple propositions; and establishing indexes based on the propositions.

[0011] In some implementations, indexing specifically includes extracting topics and keywords to form vocabulary-level information and semantic-level information, and vectorizing them together.

[0012] Any of the methods for enhancing the generation of government documents for knowledge retrieval described in any one of the above items also includes: extracting the title; if the title does not exist, directly obtaining multiple semantically complete text blocks.

[0013] In a second aspect, the present invention also provides an electronic device, comprising: at least one processor; a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can: identify the title of a government document; segment the text according to the title to obtain a plurality of semantically complete text blocks; semantically segment the semantically complete text blocks to obtain a plurality of medium-granularity text blocks, wherein the medium-granularity text blocks are text blocks that meet a preset length; and establish an index based on the plurality of medium-granularity text blocks.

[0014] In a third aspect, the present invention provides a storage medium storing a computer program, which, when executed by a processor, implements any one of the methods for enhancing the generation of government documents for knowledge retrieval.

[0015] The beneficial effects of the present invention are:

[0016] The present invention can effectively perform first-level semantic segmentation of the text according to the chapter structure by identifying the title and segmenting the text according to the title structure, thereby ensuring the internal semantic relevance and integrity and improving the accuracy of retrieval. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 A flowchart of a method for enhancing generation of government documents oriented to knowledge retrieval according to an embodiment of the present invention;

[0019] Figure 2 A flowchart of a method for enhancing generation of government documents oriented to knowledge retrieval according to another embodiment of the present invention;

[0020] Figure 3 The original text of an example of a method for enhancing and generating government documents for knowledge retrieval according to an embodiment of the present invention;

[0021] Figure 4 for Figure 3 The first text block under the title “(I) Living allowances for disabled persons in difficulty” in the original text;

[0022] Figure 5 for Figure 3 The second text block under the title “(I) Living allowance for disabled persons in difficulty” in the original text;

[0023] Figure 6 for Figure 3 The third text block under the title “(I) Living allowance for disabled persons in difficulty” in the original text;

[0024] Figure 7 The tree-like knowledge structure obtained by text slicing in the government document enhancement generation method for knowledge retrieval according to an embodiment of the present invention;

[0025] Figure 8 The figure is a block diagram of an electronic device according to an embodiment of the present invention.

[0026] Description of reference numerals:

[0027] 110. Processor;

[0028] 120. Memory. DETAILED DESCRIPTION

[0029] In related technologies, RAG can effectively reduce the occurrence of "hallucination" phenomena by providing necessary knowledge details for LLM generation. Document RAG is currently the most widely used RAG type with relatively controllable results. The retrieval information comes from the uploaded documents. A more fine-grained re-ranking model is used to reduce retrieval noise and improve the accuracy and robustness of downstream question-answering tasks. Due to the limited length of the context window of the large model, the original document needs to be sliced ​​to adapt to its input length when performing large model retrieval enhancement generation. However, if the slice is too short, the semantic information may be incomplete and affect the generation effect; if the slice is too long, it may exceed the context window of the model after recalling multiple slices, resulting in truncation and information loss. In addition, the redundant information contained in the longer slice is also likely to affect the retrieval hit effect.

[0030] Some studies have also modified the original text information, for example, semantically clustering and summarizing the text to enhance the question-answering ability of coarse-grained information; and extracting propositions from text information, which is actually rewriting the information to obtain multiple concise sentences that independently describe a certain feature of an object.

[0031] Existing methods lack the analysis of document title structure and ignore the semantic information that is actually the most meaningful for segmentation. Methods based on semantic segmentation may still have cross-structured text, resulting in poor semantic segmentation results; some methods ignore finer-grained details and also face the problem of low matching hit rate between shorter user questions and relatively long text blocks; others only focus on the finest-grained proposition information and ignore text information with a higher level of abstraction and a larger receptive field.

[0032] All of the above methods directly regard all texts as information of the same level, ignoring the rich structural information of the document itself, and are actually unable to truly preserve the semantic coherence and integrity within the block; and the above methods do not consider the short-long matching problem that may arise during the retrieval process. In fact, most user questions are short in content, and there is a difference in the amount of information with relatively long text blocks, which may lead to a low retrieval hit rate.

[0033] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0034] Combine the following Figures 1 to 8 , describing an embodiment of the present invention.

[0035] like Figure 1and Figure 2 As shown, according to an embodiment of the present invention, on the one hand, a method for enhancing generation of government documents for knowledge retrieval is provided, comprising the following steps:

[0036] Step S101: Identify the title of the government document.

[0037] Step S103: segment the text according to the title to obtain a plurality of semantically complete text blocks.

[0038] Step S105: semantically segment the semantically complete text block to obtain a plurality of medium-granularity text blocks, where the medium-granularity text block is a text block that meets a preset length.

[0039] Step S107: Create an index based on multiple medium-granularity text blocks.

[0040] In this embodiment, the structures such as pictures and tables in the document are extracted, and the corresponding semantic information is generated through the model to retrieve questions and answers about the information in the pictures and tables. If there is a title in the document, the title is identified and the title information in the document is extracted. The possible title text is obtained by using various methods such as the document's own structural information and regular expression recognition, and the final title is obtained through a screening algorithm. The text corresponding to the title is then segmented into an initial text block to obtain multiple semantically complete text blocks, so that the text blocks generated by identifying the document structure and semantically segmenting the document structure are semantically complete and coherent.

[0041] Semantically segment the semantically complete text blocks to obtain multiple medium-granularity text blocks, and then create indexes based on the multiple medium-granularity text blocks to improve the accuracy of retrieval. For example, in the face of large-model retrieval and enhanced question-and-answer scenarios in the government affairs field, retrieval can be more accurate.

[0042] Further, based on step S103, the text is segmented according to the title to obtain a plurality of semantically complete text blocks, specifically including:

[0043] Step S1031: If there are hierarchical titles, the title information of each level of title is identified respectively.

[0044] Step S1033: Merge multi-level title information to obtain multiple semantically complete text blocks.

[0045] A title segmentation module can be set to effectively extract titles and preliminarily segment the text by title. Considering that a document may have multiple titles and each title has a hierarchy, in the process of identifying titles, the form of the title can be identified first. For example, Figure 3As shown, first identify the title formats "1." and "(1)", and extract the title information based on this, namely "1. Clarify the target standards" and "(1) Living subsidies for disabled people in difficulty." Then cut the text according to the title to obtain the text belonging to the title "1. Clarify the target standards" "The recipients of living subsidies for disabled people in difficulty and the recipients of nursing subsidies for severely disabled people should have local household registration in our city, hold a valid "Disability Certificate of the People's Republic of China", and meet the relevant requirements for policy coordination." and the subsequent text belonging to the title "(1) Living subsidies for disabled people in difficulty."

[0046] A multi-level title merging module can be set up to identify the title level by delimiter judgment, sequence type judgment, text position judgment, etc., and merge multi-level titles to obtain broader complete semantic information. Among them, multi-level titles are identified and the corresponding information of lower-level titles is merged upward. Specifically, after obtaining the above-mentioned title segmentation results, for example, for Figure 2 Title recognition obtains the first-level title "I. Clarify the standards for recipients" and the second-level title "(I) Living allowance for disabled people in need.". Since "(I) Living allowance for disabled people in need." belongs to "I. Clarify the standards for recipients", their information is merged to obtain the text information corresponding to "I. Clarify the standards for recipients": Figure 3 To merge multi-level title information, the titles are first identified hierarchically, and the information is continuously merged with the upper-level titles according to the title levels, so as to construct all the text information corresponding to the upper-level titles.

[0047] Further, based on step S105, semantic segmentation is performed on the semantically complete text block to obtain a plurality of medium-granularity text blocks, specifically including:

[0048] Step S1051: If the length of the semantically complete text block is greater than a preset length, it is segmented according to punctuation marks to obtain single sentences.

[0049] Step S1053: merge adjacent single sentences according to semantics to obtain medium-granularity text blocks.

[0050] In this embodiment, the semantic segmentation stage includes a sentence segmentation module and a semantic merging module. The sentence segmentation module can segment all text blocks that do not meet the preset length obtained in the above-mentioned title segmentation stage into the finest-grained sentences. Specifically, for the smallest text unit obtained by the above-mentioned title segmentation, if its length exceeds the set text length, such as the text corresponding to the title "(i) Living allowance for disabled people in difficulty." The length is too long. First, the above-mentioned text is segmented by sentences to obtain single sentences such as "1. Severely disabled people in low-income families will be given a living allowance of 38% of the minimum living allowance standard of our city;"... "The income is verified by the verification mechanism of the economic status of residents' families." Then, the semantic relevance is judged between each adjacent sentence, and the semantically related ones are merged, and the maximum length does not exceed the set length. The text blocks obtained by punctuation segmentation and semantic merging are as follows: Figure 4 , Figure 5 , Figure 6 As shown. If the minimum text block obtained by segmenting the above title still exceeds the set text length, it is segmented according to sentence punctuation to obtain a series of sentences.

[0051] Merge the above sentences semantically, ensure that the length does not exceed the preset length, and overlap the blocks semantically. In this stage, the above title can be segmented into longer texts semantically, and the semantic associations in the segmented blocks can be maintained.

[0052] Furthermore, in step S1053, adjacent single sentences are merged according to semantics to obtain medium-granularity text blocks, which specifically include:

[0053] S10531: Use the first sentence after the end of the previous merged block to merge forward semantically.

[0054] S10533: Perform block overlap based on semantics.

[0055] The maximum length of the merged block does not exceed the preset length. To ensure the semantic coherence between sentences, the sentences are merged according to the semantics. There is a strong intrinsic semantic association between the sentences that can be merged, which ensures that the semantics in the same block are coherent and related.

[0056] Furthermore, step S107, establishing an index according to the multiple medium-granularity text blocks, specifically includes:

[0057] Step S1071: hierarchically summarize the multiple medium-granularity text blocks to obtain multiple upper-layer blocks.

[0058] Step S1073: simulating user abstract questions based on the medium-granularity text blocks and the upper-layer blocks;

[0059] Step S1075: Create an index based on the user's abstract question.

[0060] A hierarchical summary module can be set up to obtain text information with a larger receptive field within a more reasonable length. Specifically, after the above processing, a text block that meets the length requirements and is semantically coherent is obtained, that is, a medium-grained text block. The medium-grained text block is hierarchically summarized to obtain a summary, which is the upper block. The upper block is a coarse-grained text block with a large receptive field and contains the semantic information of multiple text blocks. For example, first Figure 4 , Figure 5 , Figure 6 The first level is to summarize the text in the first level, then summarize the content corresponding to the second level title "(I) Living allowance for disabled people in difficulty." Then, the first level title "I. Clarify the standards for the recipients" is to summarize the multi-level text information obtained by the title segmentation to obtain the key information.

[0061] A simulation questioning module can be set up to simulate user questions for all the obtained medium-granularity text blocks and upper-level block information to adapt to the user's questions. Simulation questioning is a method of clue enhancement. For example, for the above Figure 4 and Figure 5 The question in the middle part of the content is "There is a conscripted soldier in the family. The income of other family members is less than twice the city's minimum living standard. Should subsidies be issued?" The actual question of the user may be "My child is in the army, but my income is very low. Can I get subsidies?" It is difficult to match the user's question with complete relevant information in the original text, but the information difference between the constructed question and the actual question is smaller, and it is easier to match. By asking abstract questions to the medium-granularity text blocks and hierarchical summary content obtained above, we can obtain expressions in a broader, more colloquial, and more global form of questions, which can align with the questions that users may ask.

[0062] By simulating questions for regular text blocks and summary blocks obtained from hierarchical summaries, we can imitate users' relatively short and abstract questions. By matching questions with simulated questions, we can alleviate the problem of low hit rate of matching questions with text blocks.

[0063] Furthermore, any of the government document enhancement generation methods for knowledge retrieval also includes the following steps:

[0064] Step S109: Acquire propositions according to multiple medium-granularity text blocks to obtain multiple propositions.

[0065] Step S1011: Create an index based on the proposition.

[0066] A proposition extraction module can also be set up to extract a complete sentence that independently describes an object or a certain characteristic of an object from these text blocks that meet the requirements. The conventional text blocks obtained above are subdivided and rewritten through the model to obtain sentences that independently describe a certain characteristic of an object, and do not contain nouns with unclear references. Specifically, for Figure 5 The content in the block is extracted to extract propositions such as "household per capita income refers to the per capita income of family members living together", "the applicant is a family member living together", "the applicant's spouse is a family member living together", etc. The model is used to obtain a complete sentence in the text block that independently describes an object / a certain characteristic of an object. The pronouns in the sentence are replaced with the original object to obtain the most fine-grained information, that is, the most fine-grained text block.

[0067] Furthermore, indexing is established, specifically including: extracting topics and keywords to form vocabulary-level information and semantic-level information, and vectorizing them together.

[0068] During vectorization, in addition to semantics, topics and keywords are extracted to obtain information at the vocabulary level to prevent semantic or individual non-keyword bias in retrieval, thereby improving the retrieval hit rate. Among them, for propositions, the information itself is used as an index. For both regular text blocks and upper-level blocks, their own information and questions are used to establish multiple indexes corresponding to the same block of information.

[0069] The entire construction process can be completed by indexing the constructed questions and the blocks themselves and storing them in the database. The final knowledge structure is as follows: Figure 7 As shown, it is a tree structure.

[0070] Furthermore, any of the government document enhancement generation methods for knowledge retrieval also includes the following steps:

[0071] Step S1013: Extract the title.

[0072] Step S1015: If no title exists, multiple semantically complete text blocks are directly obtained.

[0073] like Figure 8 As shown, according to an embodiment of the present invention, on the other hand, an electronic device is provided, including: at least one processor 110; a memory 120 communicatively connected to the at least one processor, wherein the memory 120 stores instructions executable by the at least one processor 110, and the instructions are executed by the at least one processor 110, so that the at least one processor 110 can perform the following steps:

[0074] Step S101: Identify the title of the government document.

[0075] Step S103: segment the text according to the title to obtain a plurality of semantically complete text blocks.

[0076] Step S105: semantically segment the semantically complete text block to obtain a plurality of medium-granularity text blocks, where the medium-granularity text block is a text block that meets a preset length.

[0077] Step S107: Create an index based on multiple medium-granularity text blocks.

[0078] The above-mentioned electronic device can extract document titles, split and merge them by titles, so as to effectively utilize the document structure to maintain the semantic association of blocks. By simulating questions, it can bridge the gap between text blocks and user questions, improve the accuracy of text retrieval, and better adapt to large-model retrieval and enhanced question-and-answer scenarios in the government field.

[0079] According to an embodiment of the present invention, on the other hand, a storage medium is provided, storing a computer program, which, when executed by a processor, implements any one of the methods for enhancing the generation of government documents for knowledge retrieval.

[0080] The computer executes the instructions to perform the following steps:

[0081] Step S101: Identify the title of the government document.

[0082] Step S103: segment the text according to the title to obtain a plurality of semantically complete text blocks.

[0083] Step S105: semantically segment the semantically complete text block to obtain a plurality of medium-granularity text blocks, where the medium-granularity text block is a text block that meets a preset length.

[0084] Step S107: Create an index based on multiple medium-granularity text blocks.

[0085] The present invention proposes a method for enhancing the generation of government documents for knowledge retrieval. By segmenting the text by title structure and merging information of multi-level titles, the text can be effectively segmented into the first level of semantics according to the chapter structure, thereby ensuring the internal semantic relevance and integrity. Simulated questions are generated by imitating relatively short and abstract questions asked by users, and the questions are matched with the simulated questions during retrieval, thus solving the problem of low hit rate of short-to-long matching and improving the accuracy of abstract question retrieval. By introducing information at the vocabulary level on the basis of the semantic level and fusing the subject and keyword information, semantic or vocabulary bias during retrieval can be prevented, thereby improving the accuracy of knowledge retrieval.

[0086] The present invention identifies the document structure to ensure the semantic integrity within the block and the semantic distinction between blocks, then uses hierarchical summary to obtain coarse-grained text information and ensure the text length, uses proposition extraction to obtain fine-grained, concise and independent text information, uses coarse-grained, medium-grained and fine-grained information of different granularities to meet the different levels of user questioning needs, and then establishes an index through simulated questioning, which solves the problem of low short-to-long matching hit rate between shorter user questions and relatively longer text blocks. Finally, on the basis of semantics, the topic and keyword information are introduced to improve the hit rate of related slices, thereby improving the overall accuracy of retrieval.

[0087] For the government affairs field, the present invention can improve the accuracy of document segmentation in the government affairs field through title structure segmentation and semantic segmentation methods, enhance the semantic association within the text block, and ensure the semantic integrity within the block as much as possible, and obtain multi-granular text information to respond to users' different levels of questions from concrete to abstract. In addition, it also improves the hit rate of matching user questions to correct related texts through simulated questioning and vectorization methods that incorporate vocabulary information, effectively improving the accuracy of knowledge retrieval and improving the enhanced question-answering effect of large-model retrieval in the government affairs field.

[0088] Although the embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention, and such modifications and variations are all within the scope defined by the appended claims.

Claims

1. A method for enhancing the generation of government documents for knowledge retrieval, characterized in that: include: Identify the title of government documents; Segment the text according to the title to obtain multiple semantically complete text blocks; Semantically segmenting the semantically complete text block to obtain a plurality of medium-granularity text blocks, wherein the medium-granularity text blocks are text blocks that meet a preset length; An index is established based on the multiple medium-granularity text blocks.

2. The method for enhancing the generation of government documents for knowledge retrieval according to claim 1 is characterized in that: The text is segmented according to the title to obtain multiple semantically complete text blocks, including: If the title has a hierarchical title, then identify the title information of each level of title respectively; The title information at multiple levels is merged to obtain the multiple semantically complete text blocks.

3. The method for enhancing the generation of government documents for knowledge retrieval according to claim 1 is characterized in that: The semantically complete text block is semantically segmented to obtain multiple medium-granularity text blocks, specifically including: If the length of the semantically complete text block is greater than the preset length, segmenting the text block according to punctuation marks to obtain a single sentence; The adjacent single sentences are merged according to semantics to obtain the medium-granularity text block.

4. The method for enhancing the generation of government documents for knowledge retrieval according to claim 3 is characterized in that: Merging the adjacent single sentences according to semantics to obtain the medium-granularity text block specifically includes: Use the first sentence after the end of the previous merge block to merge forward semantically; Inter-block overlap by semantics; The maximum length of the merged block does not exceed the preset length.

5. The method for enhancing generation of government documents for knowledge retrieval according to any one of claims 1 to 4, characterized in that: Establishing an index according to the multiple medium-granularity text blocks specifically includes: Summarizing the multiple medium-granularity text blocks in layers to obtain multiple upper-layer blocks; Simulating user abstract questions according to the medium-granularity text blocks and the upper-level blocks; An index is established according to the user abstract question.

6. The method for enhancing generation of government documents for knowledge retrieval according to any one of claims 1 to 4, characterized in that: Also includes: Acquire propositions according to the multiple medium-granularity text blocks to obtain multiple propositions; An index is created based on the proposition.

7. The method for enhancing generation of government documents for knowledge retrieval according to any one of claims 1 to 4, characterized in that: Establishing an index includes: extracting topics and keywords to form vocabulary-level information and semantic-level information, and vectorizing them together.

8. The method for enhancing generation of government documents for knowledge retrieval according to any one of claims 1 to 4, characterized in that: Also includes: Extract title; If no title exists, multiple semantically complete text blocks are obtained directly.

9. An electronic device, characterized in that: include: at least one processor (110); a memory (120) communicatively coupled to the at least one processor (110); The memory (120) stores instructions executable by the at least one processor (110), and the instructions are executed by the at least one processor (110) to enable the at least one processor (110) to: Identify the title of government documents; Segment the text according to the title to obtain multiple semantically complete text blocks; Semantically segmenting the semantically complete text block to obtain a plurality of medium-granularity text blocks, wherein the medium-granularity text blocks are text blocks that meet a preset length; An index is established based on the multiple medium-granularity text blocks.

10. A storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a processor, a government document enhancement generation method for knowledge retrieval as described in any one of claims 1 to 8 is implemented.

Images