Government purchase question automatic reply method based on text features

Abstract

The invention provides a method for automatically replying government purchase questions based on text features, and the method specifically comprises the following steps: S1, constructing a question item'text-question-answer pair ': based on accumulated question cases, preprocessing the question items, constructing a keyword dictionary, extracting subject terms, mapping the question items to subject subsets through TF-IDF (term frequency-inverse document frequency) and clustering, and establishing a question item'text-question-answer pair'; and finally generating sorted question and answer pairs. And S2, performing automatic reply on the question function based on the general large model: performing ORC textualization processing on a new question function, disassembling the question function into a plurality of question items Q according to keywords, replying the question items Q one by one, through an LDA model subject term, matching a corresponding subject term subset and the general large model, and automatically generating a reply text. According to the method, historical data is converted into structured knowledge which can be understood and utilized by a machine, and a large model is accurately guided to generate professional replies, so that the efficiency and quality of professional work for processing questions are improved.

Description

Technical Field

[0001] This invention relates to the field of text response generation technology, specifically to an automatic response method for government procurement inquiry letters based on text features. Background Technology

[0002] A text-based response generation system is an artificial intelligence-based technology that processes and understands natural language comment text to generate responses that correspond to the comment's description. This system combines Natural Language Processing (NLP) and Computer Vision (CV) techniques, aiming to simulate human creativity and automate response generation.

[0003] Currently, the interpretation and response to inquiry letters in government procurement are all done manually. This process suffers from problems such as unclear breakdown of inquiry issues, omissions, and significant subjective differences in response quality due to limitations in individual knowledge and experience, especially regarding inconsistent response standards for frequently requested issues. A search revealed no reports on automated response methods specifically designed for government procurement inquiry letters; therefore, there is a need to develop an automated response method tailored to the textual characteristics of government procurement inquiry letters. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes an automatic response method for government procurement inquiry letters based on text features. This method transforms historical data into structured knowledge that is understandable and usable by machines, and uses this knowledge to accurately guide large-scale models to generate professional responses, thereby improving the efficiency and quality of handling inquiry letters.

[0005] To achieve the above technical solution, this invention provides an automatic response method for government procurement inquiry letters based on text features, specifically including the following steps:

[0006] S1. Constructing "text-question-answer pairs" for the objections: Based on the accumulated objection cases, the objection letters are preprocessed, a keyword dictionary is constructed, topic words are extracted, and the objections are mapped to topic subsets through TF-IDF and clustering, and finally, ranked question-answer pairs are generated.

[0007] S2. Automatic response to the challenge letter based on the general big model: First, the new challenge letter is processed into ORC text, and then broken down into several challenge items Q according to keywords. Then, each challenge item Q is matched with the subject terms of the LDA model and the corresponding "subset of subject terms". Finally, the top 3-5 groups of "text-question-answer pairs" under the corresponding "subset of subject terms" category are selected as the information source for the big model retrieval enhancement, and the new challenge items are responded to based on the general big model, and the response text is automatically generated.

[0008] Preferably, the specific process of step S1 is as follows:

[0009] S11. Identify and classify the text features of the questions raised in the letter of inquiry, preprocess a large number of existing letters of inquiry and letters of inquiry response to form a collection of inquiry response cases, convert them into text documents through OCR technology, clean and analyze Chinese text in advance, select high-quality documents, and mark the major categories of the text items Q of the question and the text items A of the response to the question.

[0010] S12. Construct a representative keyword from the total number of text keywords in each major category set, extract the constructed keywords, and build a keyword dictionary.

[0011] S13. Using the LDA model, extract keywords from the text of the case set;

[0012] S14. The TF-IDF statistical algorithm is used to calculate the TF value and IDF value of keywords and subject terms respectively. Based on the calculation results, the questioning texts to which the keywords belong are classified into the corresponding subject terms to form a "subset of subject terms".

[0013] S15. Use the repeated binary clustering algorithm to perform clustering calculations on the questionable items Q that have been labeled with text features in step S11, and establish a corresponding mapping relationship between each "questionable item Q" and the "subset of keywords" document based on the calculation results.

[0014] S16. For "Question Item Q" included in the divided "Subject Subset", extract the text of the corresponding response letter for the question item, number it, and organize it into "Text-Question-Answer Pair" of "Question Item Q-Question Response A", and number and label the question-answer pair with basic information.

[0015] Preferably, in step S11, the specific steps for preprocessing a large number of existing inquiry letters and inquiry response letters to form an inquiry response case set include:

[0016] S111. Clean, label, and organize the existing large number of government procurement inquiry letters. Use the LDA method to set keyword sets and subject terms to extract the characteristics of each inquiry in the inquiry letters. Classify the inquiry matters and calculate the inquiry matters that are divided into different subcategories.

[0017] S112. Using the TF-IDF method, evaluate the relationship between each word in the text of the question, calculate the TF-IDF value between texts of similar questions, and then calculate the cluster center representing the specific category of questions based on the TF-IDF value, and assign a single question to a cluster with high similarity.

[0018] S113. The repeated binary clustering algorithm is used to perform cluster analysis on the questionable matters Q that have been labeled with text features, and the results are divided into several subsets.

[0019] Preferably, in step S12, major categories are pre-defined, and several high-frequency words with similar meanings / expressions are grouped into a representative keyword from these categories. The pre-defined major categories for the issues to be questioned are as follows:

[0020]

[0021] Preferably, in step S16, after numbering and labeling the basic information of the question-answer pairs, multiple question-answer pairs within the same subset are sorted, with higher cluster association values ​​ranking higher.

[0022] Preferably, in step S16, the basic information to be annotated includes the source project number, time, and whether it is valid.

[0023] The beneficial effects of the automatic response method for government procurement inquiry letters based on text features provided by this invention are as follows:

[0024] (1) This invention achieves a comprehensive and beneficial effect of high response quality, high efficiency, low cost, good compliance and strong iterability by constructing a structured historical question-response knowledge base and using retrieval-enhanced generation (RAG) technology to drive a large model to respond.

[0025] (2) The responses based on the large model in this invention are not generated out of thin air, but are enhanced by searching historical, real, and high-quality "question-answer pairs". This ensures that the content of the responses conforms to the professional norms and legal provisions of specific fields such as government procurement, and greatly reduces factual errors or "fabrication". Moreover, through multi-layer mapping and correlation ranking of "keyword-subject term-cluster", the most relevant and similar historical cases can be matched for reference for the core questions, making the generated responses more targeted and logically rigorous.

[0026] (3) This invention automates the main processes from OCR, text analysis, classification, retrieval to generation, changing the traditional method that relies entirely on manual review and writing, and increasing processing speed by orders of magnitude. Ordinary staff can use this system to generate high-quality initial drafts of responses, reducing the absolute dependence on a few senior experts, alleviating human resource pressure, and saving time and economic costs.

[0027] (4) The present invention constructs a structured "subset of keywords" and sorted "question-answer pairs," which essentially systematizes and standardizes scattered expert experience and precedents, forming a unified "knowledge code" within the organization. Responses generated based on a standardized knowledge base effectively reduce inconsistencies in responses caused by differences in personal understanding among different personnel, avoid subjective bias, maintain the authority and consistency of responses, and reduce legal risks.

[0028] (5) This invention does not rely solely on large models, but instead employs a hybrid approach: "precise knowledge management using traditional NLP techniques (keyword extraction, LDA, TF-IDF, clustering) + intelligent generation using large models." This approach leverages the understanding and generation capabilities of large models while ensuring the accuracy and controllability of knowledge organization through traditional techniques, thus avoiding the knowledge illusion problem that may exist with large models in specialized fields. Furthermore, the system can provide historical cases (the top 3-5 question-answer pairs) when responding, making the response generation process verifiable, facilitating manual review and traceability, and increasing transparency and credibility. Attached Figure Description

[0029] Figure 1 This is a flowchart of the process of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0031] Example 1: An automatic response method for government procurement inquiry letters based on text features.

[0032] Reference Figure 1 As shown, an automatic response method for government procurement inquiry letters based on text features specifically includes the following steps:

[0033] S1. Based on the accumulated cases of questioning, construct a "text-question-answer pair" for the questioned issues:

[0034] S11. First, the textual features of the questions raised in the inquiry letters are identified and classified. A large number of existing inquiry letters and responses are preprocessed to form a set of inquiry response cases. Specifically, the process involves cleaning, labeling, and organizing the inquiry matters from a large number of existing government procurement inquiry letter cases. The LDA (Latent Dirichlet Allocation) method is used to set keyword sets and subject terms to extract features from each inquiry matter in the inquiry letters, classifying the inquiry matters and calculating the subcategories. Simultaneously, the TF-IDF (term frequency–inverse document frequency) method is used to evaluate the relationship between each word in the text of the inquiry matter, calculating the similarity (TF-IDF value) between texts of similar inquiry matters. Then, based on the TF-IDF value, cluster centers representing specific categories of inquiry matters are calculated, and individual inquiry matters are assigned to clusters with high similarity. A repeated binary clustering algorithm is used to perform cluster analysis on the inquiry matters (Q) that have been labeled with textual features, dividing them into several subsets according to the analysis results. Then, the documents are converted into text documents using OCR technology. After pre-cleaning and Chinese analysis, high-quality documents are selected and categorized into the Q text items for the questions raised and the A text items for the responses to the questions raised.

[0035] S12. Construct a representative keyword from the total number of text keywords in each major category set, extract the constructed keywords, and build a keyword dictionary. This can be done by pre-dividing the major categories and grouping several high-frequency words with similar meanings / expressions within each category into a representative keyword. The pre-defined major category divisions are as follows:

[0036]

[0037] S13. Using the LDA model, extract keywords from the text of the case set.

[0038] S14. The TF-IDF statistical algorithm is used to calculate the TF value and IDF value of keywords and subject terms respectively. Based on the calculation results, the questioning text to which the keyword belongs is classified into the corresponding subject term to form a "subset of subject terms".

[0039] S15. The repeated bisection clustering algorithm is used to perform clustering calculations on the questionable items (Q) that have been labeled with text features in step S11, and a corresponding mapping relationship is established between each "questionable item (Q)" and the "subset of keywords" document based on the calculation results.

[0040] S16. For each "Question (Q)" included in the predefined "Subject Topic Subset," extract the corresponding response text, number it, and organize it into "Question (Q) - Response (A)" text-question-answer pairs. Number and label each question-answer pair with basic information (such as source project number, time, and validity). Sort multiple question-answer pairs within the same subset, prioritizing those with higher cluster association values.

[0041] S2. Automatic response to inquiry letters based on a general large model:

[0042] S21: The new challenge letter is processed into ORC text and broken down into several challenge items (Q) based on keywords.

[0043] S22: For each question (Q), use the LDA model keywords and match the corresponding "subset of keywords".

[0044] S23: Select the top 3-5 "text-question-answer pairs" under the corresponding "subject term subset" category as the information source for enhancing the large-scale model retrieval. Based on the general large-scale model, respond to new questions and automatically generate response text. The general large-scale model pre-sets a high-quality response text for each subset, which corresponds to the standard response text for the question at the corresponding cluster center. Finally, the large-scale model performs semantic fine-tuning based on the standard response text to form the final response text, achieving automatic response.

[0045] This invention is a targeted development based on the textual characteristics of government procurement inquiry letters. It constructs a structured historical inquiry-response knowledge base and utilizes Retrieval Enhanced Generation (RAG) technology to drive a large-scale model for responses, achieving a comprehensive advantage of high-quality, efficient, low-cost, compliant, and highly iterative responses. Furthermore, the responses based on this large-scale model are not generated out of thin air, but rather enhanced by retrieval of historical, authentic, and high-quality question-and-answer pairs. This ensures that the response content aligns with professional norms and legal provisions in specific fields such as government procurement, significantly reducing factual errors or fabrications and improving the quality and accuracy of responses. Moreover, through multi-layered mapping and relevance ranking of "keywords-topics-clustering," the most relevant and similar historical cases can be matched for reference to the core questions, making the generated responses more targeted, logically rigorous, and achieving precise matching.

[0046] This invention automates key processes from OCR, text analysis, classification, and retrieval to generation, changing the traditional method that relies entirely on manual review and writing, resulting in an order-of-magnitude increase in processing speed. Ordinary staff can use this system to generate high-quality initial responses, reducing absolute dependence on a few senior experts, alleviating human resource pressure, and saving time and economic costs. Furthermore, this invention has excellent scalability. New challenge cases and responses, after verification, can be continuously added to the knowledge base, achieving automatic knowledge accumulation and optimization, making the system increasingly "intelligent" with use.

[0047] This invention constructs a structured "subset of keywords" and sorted "question-answer pairs," essentially systematizing and standardizing scattered expert experience and precedents to form a unified "knowledge code" within the organization. Responses generated based on this standardized knowledge base effectively reduce inconsistencies in responses caused by differences in individual understanding among processing personnel, avoid subjective bias, maintain the authority and consistency of responses, and lower legal risks.

[0048] This invention does not rely solely on large models, but instead employs a hybrid approach: "precise knowledge management using traditional NLP techniques (keyword extraction, LDA, TF-IDF, clustering) + intelligent generation using large models." This approach leverages the understanding and generation capabilities of large models while ensuring the accuracy and controllability of knowledge organization through traditional techniques, avoiding the knowledge illusion problem that large models may encounter in specialized fields. Furthermore, the system provides referenced historical cases (the top 3-5 question-answer pairs) when responding, making the response generation process traceable, facilitating manual review and traceability, and increasing transparency and credibility.

[0049] The above description is only a preferred embodiment of the present invention, but the present invention should not be limited to the content disclosed in the embodiments and drawings. Therefore, any equivalent or modified embodiments made without departing from the spirit of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for automatically responding to government procurement inquiry letters based on text features, characterized in that... Specifically, the steps include the following: S1. Constructing "text-question-answer pairs" for the objections: Based on the accumulated objection cases, the objection letters are preprocessed, a keyword dictionary is constructed, topic words are extracted, and the objections are mapped to topic subsets through TF-IDF and clustering, and finally, ranked question-answer pairs are generated. S2. Automatic response to the challenge letter based on the general big model: First, the new challenge letter is processed into ORC text, and then broken down into several challenge items Q according to keywords. Then, challenge items Q are matched one by one with the subject terms of the LDA model and the corresponding "subset of subject terms". Finally, the top 3-5 groups of "text-question-answer pairs" under the corresponding "subset of subject terms" category are selected as the information source for the big model retrieval enhancement, and the new challenge items are responded to based on the general big model, and the response text is automatically generated.

2. The automatic response method for government procurement inquiry letters based on text features as described in claim 1, characterized in that: The specific process of step S1 is as follows: S11. Identify and classify the text features of the questions raised in the letter of inquiry, preprocess a large number of existing letters of inquiry and letters of inquiry response to form a collection of inquiry response cases, convert them into text documents through OCR technology, clean and analyze Chinese text in advance, select high-quality documents, and mark the major categories of the text items Q of the question and the text items A of the response to the question. S12. Construct a representative keyword from the total number of text keywords in each major category set, extract the constructed keywords, and build a keyword dictionary. S13. Using the LDA model, extract keywords from the text of the case set; S14. The TF-IDF statistical algorithm is used to calculate the TF value and IDF value of keywords and subject terms respectively. Based on the calculation results, the questioning texts to which the keywords belong are classified into the corresponding subject terms to form a "subset of subject terms". S15. The repeated binary clustering algorithm is used to perform clustering calculations on the questionable items Q that have been labeled with text features in step S11, and a corresponding mapping relationship is established between each "questionable item Q" and the "subset of keywords" document based on the calculation results. S16. For "Question Item Q" included in the predefined "Subject Subset", extract the text of the corresponding response letter for the question item, number it, and organize it into a "text-question-answer pair" of "Question Item Q - Question Response A", and number and label the question-answer pair with basic information.

3. The automatic response method for government procurement inquiry letters based on text features as described in claim 2, characterized in that, In step S11, the specific steps for preprocessing a large number of existing inquiry letters and inquiry response letters to form a collection of inquiry response cases include: S11. Clean, label, and organize the existing large number of government procurement inquiry letters. Use the LDA method to set keyword sets and subject terms to extract the characteristics of each inquiry in the inquiry letters. Classify the inquiry matters and calculate the inquiry matters that are divided into different subcategories. S12. Using the TF-IDF method, evaluate the relationship between each word in the text of the question, calculate the TF-IDF value between texts of similar questions, and then calculate the cluster center representing the specific category of questions based on the TF-IDF value, and assign a single question to a cluster with high similarity. S13. Use the repeated binary clustering algorithm to perform cluster analysis on the questionable matters Q that have been labeled with text features, and divide them into several subsets according to the analysis results.

4. The automatic response method for government procurement inquiry letters based on text features as described in claim 2, characterized in that, In step S12, major categories are pre-defined, and several high-frequency words with similar meanings / expressions are grouped into a representative keyword from these categories. The pre-defined major categories for the issues to be questioned are as follows:

5. The automatic response method for government procurement inquiry letters based on text features as described in claim 2, characterized in that, In step S16, after numbering and labeling the basic information of the question-answer pairs, multiple question-answer pairs within the same subset are sorted, with higher cluster association values ​​ranking higher.

6. The automatic response method for government procurement inquiry letters based on text features as described in claim 2, characterized in that, In step S16, the basic information to be annotated includes the source project number, time, and whether it is valid.

Images