Intelligent tendering and bidding question-answering method based on document segmentation

Abstract

The invention discloses an intelligent tendering and bidding question-answering method based on document segmentation, and relates to the technical field of document segmentation, and the method comprises the steps: carrying out the layout analysis of a tendering and bidding document, so as to extract a structural unit set, construct a structural hierarchical relation, and carrying out the sentence processing of each structural unit; constructing a semantic window for the structural unit and generating a mixed semantic window vector, and performing primary segmentation on the structural unit by calculating a semantic gradient to generate a semantic continuous block set; second-level segmentation is carried out on the basis of the length constraint of the token, a text block set is generated, cross-level semantic coding is carried out on text blocks, and block-level vector codes are generated; outputting a local abstract for each text block according to a cross-level coding result, constructing a local abstract set, performing compression to generate a global abstract, and constructing a bidirectional mapping system of clause numbers and the text blocks; and efficient and traceable bidding and tendering document analysis and intelligent question and answer are provided for the user.

Description

Technical Field

[0001] This invention relates to the field of document segmentation technology, and in particular to an intelligent question-answering method for bidding based on document segmentation. Background Technology

[0002] With the development of natural language processing and large-scale language models, the demand for intelligent solutions in engineering management, legal text parsing, and public procurement activities continues to grow. Particularly in the bidding and tendering field, related documents are typically lengthy, complex in structure, and contain numerous cross-chapter related obligatory clauses, technical specifications, business conditions, and legal procedural requirements. The hierarchical structure and semantic dependencies within the text exhibit a high degree of nonlinearity. Traditional text retrieval tools, keyword matching methods, and retrieval-based question-answering methods that rely solely on vector semantics struggle to effectively identify the semantic continuity and logical constraints between clauses in long texts. Furthermore, while generative models based on Transformers possess some language understanding capabilities in recent years, their context windows are limited. When processing long bidding and tendering documents containing tens of thousands to hundreds of thousands of tokens, problems such as context loss, inability to maintain cross-segment logic, and deviation of generated content from the original text often arise. Therefore, how to achieve accurate extraction of structured information, semantically preserved segmented processing, and traceable question-answering reasoning in large-scale bidding and tendering documents has become an important research direction for intelligent bidding and tendering technology.

[0003] Existing technologies generally suffer from three key shortcomings: First, traditional document segmentation is mostly based on fixed token lengths or simple inter-sentence similarity, which fails to reflect the hierarchical structure features within bidding documents, leading to a break in semantic continuity after segmentation. Second, existing long text compression and summary generation methods are mostly based on single-layer summarization, which cannot simultaneously guarantee the preservation of clause-level key points and the complete presentation of the global logical chain. Third, existing question-answering systems rely solely on vector similarity to select evidence fragments, lacking clause logic extraction and clause number mapping mechanisms, which often result in system-generated answers lacking verifiability and traceability. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a document segmentation-based intelligent question-answering method for bidding, which solves the problems of lack of semantic continuity, logical consistency and auditability in existing technologies.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a document segmentation-based intelligent question-answering method for bidding and tendering, which includes, The layout of the bidding documents is analyzed to extract the set of structural units, construct the hierarchical relationship of the structure, and process each structural unit into sentences; A semantic window is constructed for the structural unit and a hybrid semantic window vector is generated. By calculating the semantic gradient, the structural unit is segmented at the first level to generate a set of semantically continuous blocks. Based on the length constraint of the token, a two-level segmentation is performed to generate a set of text blocks. Cross-level semantic encoding is then performed on the text blocks to generate block-level vector encoding. For each text block, a local summary is output based on the cross-level encoding results. A set of local summaries is constructed, compressed to generate a global summary, and a bidirectional mapping system between clause numbers and text blocks is constructed. The questions are encoded into vectors for semantic similarity calculation to obtain a set of evidence, and clause logic is extracted to form a logical chain. Based on user questions, evidence sets, logical chains, and global summaries, two rounds of intelligent question-answering generation are performed to output structured question-answering results.

[0007] As a preferred embodiment of the document segmentation-based intelligent question-answering method for bidding and tendering as described in this invention, the step of parsing the layout of the bidding and tendering document to extract a set of structural units, constructing a hierarchical relationship, and performing sentence-based processing on each structural unit includes: The input bidding documents are analyzed for layout, headers and footers are removed, structural units are extracted, and a set of structural units is formed. Perform OCR and text cleaning operations on the text of each structural unit; Each structural unit is segmented into a sentence sequence using a sentence segmenter with legal and contractual text optimization rules.

[0008] As a preferred embodiment of the document segmentation-based intelligent question-answering method for bidding and tendering described in this invention, the method involves: constructing semantic windows for structural units and generating hybrid semantic window vectors; calculating semantic gradients to perform first-level segmentation of structural units; and generating a set of semantically continuous blocks, including... For each structural unit Constructing a semantic window ; A weighted fusion of LSA and word2vec is used to generate a hybrid semantic window vector for semantic windows; The semantic gradient is obtained by comparing adjacent window vectors to assess the degree of semantic change. ; For all Sort and set percentile thresholds ,like Greater than or equal to Then in the structural unit and Mark the first-level segmentation points between the structural units, and use these first-level segmentation points as boundaries to divide the set of structural units into M semantically continuous blocks, forming a set of semantically continuous blocks. .

[0009] As a preferred embodiment of the intelligent question-answering method for bidding based on document segmentation described in this invention, wherein: the secondary segmentation based on the length constraint of the token generates a set of text blocks, including, For each semantically contiguous block Use GPT's tokenizer to calculate the number of tokens. ; Set the maximum token length based on the token length constraint. ; If a semantically continuous block If the number of tokens is greater than the maximum token length, then start accumulating tokens from the first sentence of the block, find the maximum total sentence length p, and make p the maximum total sentence length p. and At sentence p, Divided into two new blocks and ,like Still exceeding This process is repeated until a set of text blocks that meet the token length limit is obtained. .

[0010] As a preferred embodiment of the document segmentation-based intelligent question-answering method for bidding and tendering described in this invention, the step of performing cross-level semantic encoding on text blocks to generate block-level vector encoding includes: Each sentence in the text block is word2vec encoded to generate a sequence of word vectors, and then sentence-level vectors are generated using word-level convolutional attention. ; Sentence-level vectors are aggregated using sentence-level attention to generate block-level vectors.

[0011] As a preferred embodiment of the document segmentation-based intelligent question-answering method for bidding and tendering described in this invention, the method involves: outputting a local summary for each text block based on cross-level encoding results, constructing a set of local summaries, and compressing it to generate a global summary, including... Input the sentence sequence and corresponding sentence-level attention weights of each text block into the large model, extract the obligation, condition, exception, and consequence information, and output a locally compressed summary. Length is limited to ; All The data are concatenated into a set of locally compressed summaries, and then globally compressed. The large model generates a global summary based on logical relationships.

[0012] As a preferred embodiment of the document segmentation-based intelligent question-answering method for bidding and tendering described in this invention, the construction of a bidirectional mapping system between clause numbers and text blocks includes, Perform regular expression matching on each text block to identify all clause numbers, and record the text block and position of each clause number to build a clause index table; Source clause tags are automatically inserted into the generated global summary to form a reverse path from the summary to the original text.

[0013] As a preferred embodiment of the document segmentation-based intelligent question-answering method for bidding and tendering described in this invention, wherein: the step of encoding the question into a vector and calculating semantic similarity to obtain an evidence set includes, The user-input question is encoded into a question vector, and the semantic similarity between the question vector and each block-level vector is calculated using dot product similarity. The text blocks are sorted in descending order of semantic similarity, and the top n text blocks in the list are selected as the evidence set.

[0014] As a preferred embodiment of the intelligent question-answering method for bidding based on document segmentation described in this invention, the step of extracting clause logic to form a logical chain includes: The obligations, conditions, additional examples, and consequences of the local compressed summary of each text block in the evidence set are logically extracted to obtain a logical set. All logical sets are then merged into a logical chain, and conflicts are checked.

[0015] As a preferred embodiment of the document segmentation-based intelligent question-answering method for bidding and tendering described in this invention, the method involves: generating two rounds of intelligent question-answering based on user questions, evidence sets, logical chains, and global summaries, and outputting structured question-answering results, including: Perform the first round of reasoning on each text block in the evidence set to obtain a local answer for each text block. ; The local answer is combined with the original question, logical chain, and global summary. Global reasoning is performed using a large model, and the clause numbers pointing to the original text are cited from the clause index table to generate the final answer text. The final output structured question-and-answer structure includes the final answer text, a list of cited bidding terms, evidence text block numbers, and logical chain reference points.

[0016] The beneficial effects of this invention are as follows: by constructing a set of structural units, generating hybrid semantic window vectors, performing first-level semantic gradient segmentation, second-level token constraint segmentation, cross-level encoding, summary compression, clause mapping, selecting a set of evidence, constructing logical chains, generating two rounds of intelligent question answering, and outputting structured results, this invention improves the efficiency of bidding document processing while achieving accurate extraction of structured information, semantically continuous segmented processing, and traceable question answering reasoning, thus providing users with efficient and traceable bidding document analysis and question answering. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the intelligent question-answering method for bidding based on document segmentation in Example 1.

[0019] Figure 2 This is a flowchart of the semantic window construction and segmentation process in Example 1. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0023] Example 1, referring to Figure 1 and Figure 2 This is the first embodiment of the present invention, which provides a document segmentation-based intelligent question-answering method for bidding, including the following steps: S1. Analyze the layout of the bidding documents to extract the set of structural units, construct the hierarchical relationship of the structure, and process each structural unit into sentences; Specifically, the input bidding documents are analyzed for layout, headers and footers are removed, structural units are extracted, and a set of structural units is formed, represented as follows: ; in, This represents the i-th structural unit, which includes chapter titles, clause numbers, body text, list and table content; Perform OCR and text cleaning operations on the text of each structural unit; Each structural unit is segmented into a sentence sequence using a sentence segmenter with legal and contractual text optimization rules.

[0024] By parsing the layout of bidding documents, extracting structural unit sets, constructing hierarchical relationships, and combining legal / contract optimization rules for sentence-based processing, this method transforms the originally complex bidding documents, which contain mixed information such as headers, footers, tables, and lists, into a clean text representation at two levels: "structural unit - sentence." On the one hand, it eliminates noise information irrelevant to the contract semantics, reducing the risk of misjudgment by subsequent models. On the other hand, by incorporating chapter titles, clause numbers, body text, lists, and table content into a unified structural unit system, each piece of content has a clear structural positioning and legal context. This facilitates the subsequent establishment of semantic windows and clause indexes at the clause level, providing a standardized input format for subsequent steps such as semantic segmentation, cross-level coding, and clause number mapping. Ultimately, this significantly improves the robustness of subsequent algorithms and provides a high-quality structured data foundation for the entire intelligent question-answering process.

[0025] S2. Construct semantic windows for structural units and generate hybrid semantic window vectors. Calculate semantic gradients to perform first-level segmentation of structural units and generate a set of semantically continuous blocks. Specifically, for each structural unit Constructing a semantic window , is represented as: ; A weighted fusion of LSA and word2vec is used to generate a hybrid semantic window vector, represented as follows: ; in, Represents a mixed semantic window vector. The weighting coefficients are determined through cross-validation. Represents the topic semantic vector. This represents the average of word vectors; The semantic gradient is obtained by comparing adjacent window vectors to assess the degree of semantic change. , is represented as: ; For all Sort and set percentile thresholds ,like Greater than or equal to Then in the structural unit and Mark the first-level segmentation points between the structural units, and use these first-level segmentation points as boundaries to divide the set of structural units into M semantically continuous blocks, forming a set of semantically continuous blocks. .

[0026] By constructing a semantic window for each structural unit and generating a hybrid semantic window vector through weighted fusion of LSA and word2vec, and then calculating the semantic gradient by the difference between adjacent window vectors, and determining whether to set a first-level segmentation point based on a percentile threshold, a first-level segmentation of structural units based on the "degree of semantic change" is achieved, resulting in a set of semantically continuous blocks. This not only considers the current structural unit but also captures semantic transition features through windows formed by preceding and following structural units, avoiding missegmentation caused by the semantic instability of a single structural unit. At the same time, using a statistical threshold to control the number of segmentation points allows for more precise control over block size and segmentation sensitivity. The generated semantically continuous blocks are semantically coherent and have clear boundaries, providing a high-quality "semantic foundation" for subsequent token-constrained segmentation and cross-level semantic encoding. This approach is more suitable for subsequent intelligent question-answering tasks than traditional methods of segmentation by page, chapter, or fixed length.

[0027] S3. Perform secondary segmentation based on the length constraint of the token to generate a set of text blocks. Perform cross-level semantic encoding on the text blocks to generate block-level vector encoding. Specifically, for each semantically contiguous block Use GPT's tokenizer to calculate the number of tokens. ; Set the maximum token length based on the token length constraint. ; If a semantically continuous block If the number of tokens is greater than the maximum token length, then start accumulating tokens from the first sentence of the block, find the maximum total sentence length p, and make p the maximum total sentence length p. and At sentence p, Divided into two new blocks and ,like Still exceeding This process is repeated until a set of text blocks that meet the token length limit is obtained. ; Furthermore, each sentence in the text block is encoded using word2vec to generate a sequence of word vectors, and then sentence-level vectors are generated using word-level convolutional attention. , is represented as: ; in, Represents sentence-level vectors. This represents a convolution operation used to extract n-gram features. Representing word vectors, Indicates a word index; Sentence-level vectors are aggregated using sentence-level attention to generate block-level vectors, represented as follows: ; ; in, Represents a text block block-level vectors, Represents sentence-level attention weights. This represents a trainable matrix.

[0028] By using token constraints, it is ensured that each text block can be processed completely by the large model in a single pass. This achieves the subdivision of all semantically continuous blocks into a set of text blocks with strictly controlled lengths that do not cross model context limitations. This guarantees that all subsequent large model calls (summarizing, question answering) can see the complete content of a text block in a single round of inference, avoiding the problem of "input being truncated and the answer missing context". By capturing local patterns through convolution and word vectors, sentence-level semantic representations are generated, making it easier for subsequent attention and question answering to "grasp the key sentences", thereby improving the model's ability to handle complex contractual sentence structures (long sentences, nested sentences). This approach leverages the understanding capabilities of clauses and conditional sentences to generate block-level vectors through sentence-level attention aggregation. This unifies fine-grained semantics at the word level, importance differences at the sentence level, and overall semantics at the block level into a single vector space. This provides a computationally achievable and indexable semantic feature foundation for subsequent vector retrieval, semantic matching, and summary compression. Simultaneously, it ensures stable operation under limited computing power and context windows during engineering deployment. Ultimately, it achieves the beneficial effect of compressing massive amounts of bidding texts to a granularity that can be processed by large models without sacrificing semantic integrity, and constructing a high-quality semantic representation that supports multiple tasks.

[0029] S4. For each text block, output a local summary based on the cross-level encoding results, construct a set of local summaries, compress them to generate a global summary, and construct a bidirectional mapping system between clause numbers and text blocks; Specifically, the large model is fed with the sentence sequence of each text block and the corresponding sentence-level attention weights to extract information on obligations, conditions, exceptions, and consequences, and outputs a locally compressed summary. Length is limited to ; All The local compressed summary set is concatenated, and then the local compressed summary set is compressed globally. The large model generates a global summary based on logical relationships. Furthermore, regular expression matching is performed on each text block to identify all clause numbers, and the text block and position of each clause number are recorded to construct a clause index table; Source clause tags are automatically inserted into the generated global summary to form a reverse path from the summary to the original text.

[0030] By inputting the sentence sequence and sentence-level attention weights of each text block into a large model based on cross-level encoding results, local compressed summaries are extracted according to four categories of information: obligations, conditions, exceptions, and consequences, thus achieving the structured extraction of key points for each text block. Subsequently, all local summaries are concatenated into a set of local summaries and subjected to a second global compression to generate a global summary that maintains the main logical thread of the contract. At the same time, regular expression matching is performed on each text block to identify the clause number and record its text block and position, constructing a clause index table, and inserting source clause tags into the global summary to form a reverse path from the summary to the original text. This allows subsequent Q&A to cite specific clause numbers, meeting the rigid requirement of "answers must be based on clauses" in scenarios such as engineering legal affairs, auditing, and complaint handling. Ultimately, this achieves the beneficial effect of significantly compressing the amount of information while maintaining legal traceability and accurate clause-level mapping.

[0031] S5. Encode the questions into vectors, calculate semantic similarity, obtain the evidence set, extract the clause logic, and form a logical chain; Specifically, the user-input question is encoded into a question vector, and the semantic similarity between the question vector and each block-level vector is calculated using dot product similarity, expressed as: ; in, Let represent the semantic similarity between the j-th text block and the question q. Represents the problem vector; The text blocks are sorted in descending order of semantic similarity, and the top n text blocks in the list are selected as the evidence set.

[0032] Furthermore, the obligations, conditions, additional examples, and consequences of the local compressed summaries of each text block in the evidence set are logically extracted to obtain a logical set. All logical sets are then merged into a logical chain, and conflicts are checked.

[0033] By encoding user questions into question vectors and using dot product similarity to calculate the semantic relevance between questions and text blocks, and sorting text blocks in descending order of similarity and selecting the top n blocks as the evidence set, this method achieves precise localization of the contract content most relevant to the question within a three-layer encoding space of block-sentence-word. Through the construction of logical chains, contractual obligations, conditions, and consequences scattered across multiple blocks are linked into a reasonable rule structure, which can identify logical conflicts or complementary relationships between multiple clauses under the same question. This prepares a "filtered and structured evidence space" for the model before intelligent question answering, ensuring that subsequent answers are reasoned within the scope of clauses that fit the semantics of the question, rather than blindly generated across the entire contract, thus providing a solid foundation for generating "reasonable and well-founded" explanatory answers.

[0034] S6. Based on user questions, evidence sets, logical chains, and global summaries, perform two rounds of intelligent question answering generation and output structured question answering results; Specifically, the first round of reasoning is performed on each text block in the evidence set to obtain a local answer for each text block. , is represented as: ; The local answer is combined with the original question, logical chain, and global summary. Global reasoning is performed using a large model, and the clause numbers pointing to the original text are cited from the clause index table to generate the final answer text. The final output structured question-and-answer structure includes the final answer text, a list of cited bidding terms, evidence text block numbers, and logical chain reference points.

[0035] The first round of local reasoning addresses each evidence block individually, avoiding "attention dilution" caused by inputting all blocks together. The second round of global reasoning unifies and coordinates the local answers, logical chains, and global summaries, achieving a secondary judgment from a "local clause perspective" to a "holistic contract perspective." This allows the large model to resolve conflicts between different clauses and explain the priority of exception clauses. Simultaneously, by leveraging the global summary, the overall principles of the contract are considered, ensuring that the final answer achieves a balance between clause applicability, logical consistency, and the spirit of the overall contract. Through structured output, the answer can be processed by machines and verified by users, thereby improving the interpretability, reusability, and compliance implementation capabilities of intelligent question answering results.

[0036] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A document segmentation-based intelligent question-answering method for bidding and tendering, characterized by: include, The layout of the bidding documents is analyzed to extract the set of structural units, construct the hierarchical relationship of the structure, and process each structural unit into sentences; A semantic window is constructed for the structural unit and a hybrid semantic window vector is generated. By calculating the semantic gradient, the structural unit is segmented at the first level to generate a set of semantically continuous blocks. Based on the length constraint of the token, a two-level segmentation is performed to generate a set of text blocks. Cross-level semantic encoding is then performed on the text blocks to generate block-level vector encoding. For each text block, a local summary is output based on the cross-level encoding results. A set of local summaries is constructed, compressed to generate a global summary, and a bidirectional mapping system between clause numbers and text blocks is constructed. The questions are encoded into vectors for semantic similarity calculation to obtain a set of evidence, and clause logic is extracted to form a logical chain. Based on user questions, evidence sets, logical chains, and global summaries, two rounds of intelligent question-answering generation are performed to output structured question-answering results.

2. The intelligent question-answering method for bidding based on document segmentation as described in claim 1, characterized in that: The process involves parsing the layout of the bidding documents to extract a set of structural units, constructing a hierarchical structure, and processing each structural unit into sentences, including... The input bidding documents are analyzed for layout, headers and footers are removed, structural units are extracted, and a set of structural units is formed. Perform OCR and text cleaning operations on the text of each structural unit; Each structural unit is segmented into a sentence sequence using a sentence segmenter with legal and contractual text optimization rules.

3. The intelligent question-answering method for bidding based on document segmentation as described in claim 2, characterized in that: The process involves constructing semantic windows for structural units and generating hybrid semantic window vectors. By calculating semantic gradients, the structural units are segmented at the first level to generate a set of semantically continuous blocks, including... For each structural unit Constructing a semantic window ; A weighted fusion of LSA and word2vec is used to generate a hybrid semantic window vector for semantic windows; The semantic gradient is obtained by comparing adjacent window vectors to assess the degree of semantic change. ; For all Sort and set percentile thresholds ,like Greater than or equal to Then in structural unit and Mark the first-level segmentation points between the structural units, and use these first-level segmentation points as boundaries to divide the set of structural units into M semantically continuous blocks, forming a set of semantically continuous blocks. .

4. The intelligent question-answering method for bidding based on document segmentation as described in claim 3, characterized in that: The second-level segmentation based on the token length constraint generates a set of text blocks, including... For each semantically contiguous block Use GPT's tokenizer to calculate the number of tokens. ; Set the maximum token length based on the token length constraint. ; If a semantically continuous block If the number of tokens is greater than the maximum token length, then start accumulating tokens from the first sentence of the block, find the maximum total sentence length p, and make p the maximum total sentence length p. and At sentence p, Divided into two new blocks and ,like Still exceeding This process is repeated until a set of text blocks that meet the token length limit is obtained. .

5. The intelligent question-answering method for bidding based on document segmentation as described in claim 4, characterized in that: The step of performing cross-level semantic encoding on text blocks to generate block-level vector encoding includes, Each sentence in the text block is word2vec encoded to generate a sequence of word vectors, and then sentence-level vectors are generated using word-level convolutional attention. ; Sentence-level vectors are aggregated using sentence-level attention to generate block-level vectors.

6. The intelligent question-answering method for bidding based on document segmentation as described in claim 5, characterized in that: The process involves outputting a local summary for each text block based on cross-level encoding results, constructing a set of local summaries, and compressing it to generate a global summary, including... Input the sentence sequence and corresponding sentence-level attention weights of each text block into the large model, extract the obligation, condition, exception, and consequence information, and output a locally compressed summary. Length is limited to ; All The data are concatenated into a set of locally compressed summaries, and then globally compressed. The large model generates a global summary based on logical relationships.

7. The intelligent question-answering method for bidding based on document segmentation as described in claim 6, characterized in that: The bidirectional mapping system between clause numbers and text blocks includes, Perform regular expression matching on each text block to identify all clause numbers, and record the text block and position of each clause number to build a clause index table; Source clause tags are automatically inserted into the generated global summary to form a reverse path from the summary to the original text.

8. The intelligent question-answering method for bidding based on document segmentation as described in claim 7, characterized in that: The process of encoding the question into a vector and calculating semantic similarity yields an evidence set, including: The user-input question is encoded into a question vector, and the semantic similarity between the question vector and each block-level vector is calculated using dot product similarity. The text blocks are sorted in descending order of semantic similarity, and the top n text blocks in the list are selected as the evidence set.

9. The intelligent question-answering method for bidding based on document segmentation as described in claim 8, characterized in that: The process of extracting clause logic to form a logical chain includes, The obligations, conditions, additional examples, and consequences of the local compressed summary of each text block in the evidence set are logically extracted to obtain a logical set. All logical sets are then merged into a logical chain, and conflicts are checked.

10. The intelligent question-answering method for bidding based on document segmentation as described in claim 9, characterized in that: The process involves two rounds of intelligent question-answering generation based on user questions, evidence sets, logical chains, and a global summary, outputting structured question-answering results, including: Perform the first round of reasoning on each text block in the evidence set to obtain a local answer for each text block. ; The local answer is combined with the original question, logical chain, and global summary. Global reasoning is performed using a large model, and the clause numbers pointing to the original text are cited from the clause index table to generate the final answer text. The final output structured question-and-answer structure includes the final answer text, a list of cited bidding terms, evidence text block numbers, and logical chain reference points.

Images