Recall method and device of retrieval enhancement system

Abstract

The invention relates to the technical field of information retrieval, in particular to a recall method and device of a retrieval enhancement system. The method comprises the following steps: S1, storing knowledge points; s2, carrying out recall enhancement; s3, checking the consistency of recall results; and S4, summarizing the model. The device comprises a knowledge point storage module, a recall enhancement module, a recall result consistency checking module and a model summarization module. According to the method, the features of the knowledge points are extracted and analyzed through the topic model and the clustering model, and the accuracy and availability of the knowledge points are improved. And the ES recall result is converted into the vector, so that the text information is quantized and represented, and the recall accuracy is improved. And the accuracy and the reliability of the recall result are ensured by performing consistency check on the recall result.

Description

Technical Field

[0001] The present invention relates to the field of information retrieval technology, and in particular to a recall method and device for a retrieval enhancement system. Background Art

[0002] In modern information retrieval systems, recall and precision are key indicators for measuring system performance. As users' requirements for information retrieval increase, traditional retrieval methods based on keyword matching have gradually exposed their limitations. For example, the keyword matching method searches for documents in the index database based on the query input by the user to find documents that match the query. However, this method relies on the consistency of the query and the vocabulary in the document, and cannot effectively handle the semantic issues of synonyms, near-synonyms, and different expressions, resulting in insufficient recall and precision.

[0003] In order to improve retrieval performance, many researchers have begun to introduce methods based on vector representation in recent years. For example, methods based on word embeddings, such as Word2Vec and Bert, can map vocabulary to vector space and judge the similarity between texts by calculating the distance or similarity between vectors. This method solves the problem of semantic matching to a certain extent. However, word embedding methods still have some shortcomings, such as the inability to effectively capture the global semantic information of long texts and limited processing effects on short texts or texts in specific fields.

[0004] With the development of deep learning technology, pre-trained language models (such as BERT, GPT-3, etc.) have gradually become the mainstream method for text representation and retrieval tasks. These models can generate text vector representations with rich semantic information through pre-training on large-scale corpora, and adapt to specific tasks through fine-tuning.

[0005] Although the existing retrieval technology has improved the performance of the retrieval system to a certain extent, it still has the following significant defects and shortcomings:

[0006] (1) Limitations of keyword matching: Traditional keyword matching methods rely on the consistency between the query and the vocabulary in the document, and cannot effectively handle semantic issues such as synonyms, near-synonyms, and different expressions. This method has high requirements for the query input by the user, which easily leads to insufficient recall rate, especially in retrieval tasks with complex semantics.

[0007] (2) Limitations of word embedding methods: Although word embedding-based retrieval methods can handle semantic issues to a certain extent, their ability to capture the global semantics of long texts is limited. Word embedding methods can usually only represent semantic relationships at the word level and cannot fully utilize semantic information at the sentence or paragraph level, resulting in poor results when processing long texts and texts in specific fields.

[0008] (3) Pre-trained language models: Although pre-trained language models (such as BERT, GPT-3, etc.) have performed well in many natural language processing tasks, their application still faces some challenges. In certain specific fields or tasks, the generalization ability of pre-trained models may be insufficient, resulting in unsatisfactory retrieval results.

[0009] (4) Consistency of recall results: Existing retrieval methods may return contradictory or inconsistent results when performing a recall. This inconsistency affects the user's retrieval experience and reduces the credibility of the retrieval results. In practical applications, how to ensure the consistency of recall results is an important challenge.

[0010] Based on the above problems, it can be seen that the existing retrieval methods still have a lot of room for improvement in improving system recall, accuracy and result consistency. Summary of the invention

[0011] In view of the problems existing in the background technology, a recall method and device of a retrieval enhancement system are proposed. The characteristics of knowledge points are extracted and analyzed through topic models and clustering models, thereby improving the accuracy and availability of knowledge points. By converting ES recall results into vectors, text information can be quantified and represented, thereby improving the accuracy of recall. By performing consistency checks on the recall results, the accuracy and reliability of the recall results are guaranteed.

[0012] The present invention provides a recall method of a retrieval enhancement system, comprising the following steps:

[0013] S1. Knowledge point storage: Use the model to extract and analyze the features of the knowledge point to obtain the subject words and cluster categories of the knowledge point; while storing the knowledge point itself, the subject words and cluster categories of the knowledge point are stored together as the feature fields of the knowledge point;

[0014] S2, recall enhancement: The query input by the user is first recalled by es keyword matching to obtain es recall results, and then the es recall results are converted into vectors, and then matched from the vector knowledge base to obtain similar results as a way of recall enhancement;

[0015] S3, recall result consistency check: After the previous step of recall enhancement, the recall result is obtained, but the recall result may be contradictory. At this time, the consistency of the model output is used to judge the result;

[0016] S4. Model summary: Put the recall results obtained after the consistency check into the generative model, perform the final result summary, and obtain the final recall results.

[0017] Preferably, in the process of storing knowledge points, two different models are used to further extract and analyze the features of the knowledge points; the first is a topic model with supervised training, and the other is a clustering model.

[0018] Preferably, the topic model is used to help find the key words in the knowledge points.

[0019] Preferably, the knowledge points are classified according to their content and features through a clustering model to obtain the category to which each knowledge point belongs.

[0020] Preferably, when storing knowledge points in the knowledge base, not only the knowledge points themselves are stored, but also the subject words and categories obtained through the topic model and clustering model, so that each knowledge point has its own feature field in the knowledge base; the field includes the subject words and categories of the knowledge point.

[0021] Preferably, based on the similarity of the vectors, the document or information most similar to the ES recall result is found.

[0022] Preferably, a consistency judgment model is used to traverse the query and each recall result to give a consistency judgment result; if the model determines that the information in the recall result is consistent, the recall result is retained, otherwise the recall result is discarded.

[0023] The present invention further proposes a recall device of a retrieval enhancement system, based on the recall method of the retrieval enhancement system, comprising:

[0024] The knowledge point storage module uses the supervised training topic model and clustering model to extract and analyze the features of the knowledge point, and obtain the subject words and cluster categories of the knowledge point; while storing the knowledge point itself, the subject words and cluster categories of the knowledge point are stored together as the feature fields of the knowledge point;

[0025] The recall enhancement module provides an input for the user to input the query, and then feeds back the ES recall result after ES keyword matching and recall. The ES recall result is then converted into a vector. Finally, based on the similarity of the vector, the document or information most similar to the ES recall result is found.

[0026] The recall result consistency check module uses the consistency judgment model to traverse the query and each recall result to make a consistency judgment on the obtained recall results.

[0027] And the model summary module, the model summary module is used to put the recall results obtained after the consistency check into the generative model, and perform the final result summary to obtain the final recall result.

[0028] Compared with the prior art, the present invention has the following beneficial technical effects:

[0029] (1) The present invention significantly improves the recall rate and accuracy of the system by combining the topic model, clustering model and recall enhancement mechanism through the knowledge point storage method. Compared with the traditional keyword matching method, the method based on the present invention can more accurately retrieve the information required by the user, especially when processing complex queries and long texts.

[0030] (2) The present invention uses a consistency discrimination model to perform consistency checks and adjustments on the recall results to ensure semantic consistency of the final results. This design significantly improves the credibility of the retrieval results and user experience, and avoids the common problem of inconsistent results in traditional retrieval systems.

[0031] (3) The present invention summarizes the recall results after consistency verification through a generative model, providing concise and accurate final retrieval results. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A flowchart of a recall method of a retrieval enhancement system in the present invention;

[0033] Figure 2 This is a structural block diagram of the knowledge point storage module in the present invention;

[0034] Figure 3 This is a workflow diagram of the recall enhancement module in the present invention;

[0035] Figure 4 This is a workflow diagram of the recall result consistency verification module in the present invention. DETAILED DESCRIPTION

[0036] Embodiment 1

[0037] like Figure 1 As shown, the present invention proposes a recall method of a retrieval enhancement system, the steps comprising:

[0038] S1. Knowledge point storage: Use the model to extract and analyze the features of the knowledge point to obtain the subject words and cluster categories of the knowledge point; while storing the knowledge point itself, the subject words and cluster categories of the knowledge point are stored together as the feature fields of the knowledge point;

[0039] S2, recall enhancement: The query input by the user is first recalled by es keyword matching to obtain es recall results, and then the es recall results are converted into vectors, and then matched from the vector knowledge base to obtain similar results as a way of recall enhancement;

[0040] S3, recall result consistency check: After the previous step of recall enhancement, the recall result is obtained, but the recall result may be contradictory. At this time, the consistency of the model output is used to judge the result;

[0041] S4. Model summary: Put the recall results obtained after the consistency check into the generative model, perform the final result summary, and obtain the final recall results.

[0042] It should be further explained that in the process of knowledge point storage, two different models are used to further extract and analyze the features of knowledge points; the first is a supervised training topic model, and the other is a clustering model. The topic model is used to help find the subject words in the knowledge points, which can reflect the main content and features of the knowledge points. The clustering model is used to classify the knowledge points according to their content and features to obtain the category to which each knowledge point belongs. When storing knowledge points in the knowledge base, not only the knowledge points themselves are stored, but also the subject words and categories obtained through the topic model and clustering model, so that each knowledge point has its own feature field in the knowledge base; the field includes the subject words and categories of the knowledge points. This method can make the knowledge points in the knowledge base richer and more detailed, and it is also more convenient to quickly and accurately find the required knowledge points in subsequent use and analysis.

[0043] It should be further explained that after the query is input into ES to complete the recall, a series of ES recall results will be obtained. These results are based on keyword matching. However, keyword matching may ignore some semantic similarities, so these ES recall results need to be further processed. Next, these ES recall results are converted into vectors. Then you can look for the results most similar to these vectors in the vector knowledge base. Based on the similarity of the vectors, find the documents or information most similar to the ES recall results. This method can be used as a way to enhance recall. It can supplement the information that may be missed by keyword matching and improve the quality and accuracy of recall. This recall enhancement method based on vector similarity can help better understand the user's query intent and provide more accurate and relevant search results.

[0044] It should be further explained that in the previous step, the recall enhancement technology was used to successfully obtain the recall results. However, this process is not perfect. The recall results may be contradictory or conflicting, which may cause trouble for subsequent processing and analysis. In order to solve this problem, a consistency judgment model is used to traverse the query and each recall result to give a consistency judgment result. The output result of this model is a judgment on the consistency of the recall results. If the model determines that the information in the recall result is consistent, the recall result is retained, otherwise the recall result is discarded.

[0045] Embodiment 2

[0046] like Figure 2-Figure 4 As shown, this embodiment proposes a recall device for a retrieval enhancement system. Based on the recall method for the retrieval enhancement system in Embodiment 1, the device includes a knowledge point storage module, a recall enhancement module, a recall result consistency verification module, and a model summary module.

[0047] The knowledge point storage module uses supervised training topic models and clustering models to extract and analyze the features of knowledge points, and obtain the subject words and cluster categories of the knowledge points; while storing the knowledge points themselves, the subject words and cluster categories of the knowledge points are also stored as feature fields of the knowledge points.

[0048] The recall enhancement module provides an input for the user to enter the query, and after ES keyword matching and recall, it feeds back the ES recall result, then converts the ES recall result into a vector, and finally finds the document or information most similar to the ES recall result based on the similarity of the vector.

[0049] The recall result consistency verification module uses the consistency judgment model to traverse the query and each recall result to make a consistency judgment on the obtained recall results.

[0050] The model summary module is used to put the recall results obtained after the consistency check into the generative model and perform the final result summary to obtain the final recall results.

[0051] The present invention combines the supervised training topic model and clustering model to extract features and classify knowledge points, thereby enhancing the structured information of the knowledge base. When users query, the recall enhancement mechanism of keyword matching and vector matching is used to improve the recall rate and accuracy of the system. By optimizing the calculation process of recall enhancement and using an efficient vector matching algorithm, the query processing process is accelerated. Through the optimization design of the entire process, the real-time response capability of the retrieval system is improved to meet the user's requirements for real-time and efficiency. In the recall result processing stage, a consistency discrimination model is introduced to perform consistency verification and adjustment on the recall results. Ensure that the results finally presented to the user are semantically consistent, and improve the credibility of the retrieval results and user experience. The recall results after the consistency verification are summarized through a generative model to provide concise and accurate final retrieval results. Help users quickly obtain the required information and improve the practicality of the retrieval system.

[0052] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited thereto, and various changes can be made within the knowledge scope of technicians in the relevant technical field without departing from the purpose of the present invention.

Claims

1. A recall method for a retrieval enhancement system, characterized in that the steps include: S1. Knowledge point storage: Use the model to extract and analyze the features of knowledge points to obtain the subject words and cluster categories of knowledge points; When the knowledge point itself is stored, the subject words and cluster categories of the knowledge point are stored together as the feature fields of the knowledge point; S2, recall enhancement: The query input by the user is first recalled by es keyword matching to obtain es recall results, and then the es recall results are converted into vectors, and then matched from the vector knowledge base to obtain similar results as a way of recall enhancement; S3, recall result consistency check: After the previous step of recall enhancement, the recall result is obtained, but the recall result may be contradictory. At this time, the consistency of the model output is used to judge the result; S4. Model summary: Put the recall results obtained after the consistency check into the generative model, perform the final result summary, and obtain the final recall results.

2. The recall method of the retrieval enhancement system according to claim 1, characterized in that: In the process of knowledge point storage, two different models are used to further extract and analyze the features of knowledge points; The first is a topic model with supervised training, and the other is a clustering model.

3. The recall method of the retrieval enhancement system according to claim 2, characterized in that: The topic model is used to help find the key words in the knowledge points.

4. The recall method of the retrieval enhancement system according to claim 3, characterized in that: The knowledge points are classified according to their content and characteristics through the clustering model to obtain the category to which each knowledge point belongs.

5. The recall method of the retrieval enhancement system according to claim 4, characterized in that: When storing knowledge points in the knowledge base, not only the knowledge points themselves are stored, but also the subject words and categories obtained through the topic model and clustering model, so that each knowledge point has its own feature field in the knowledge base; the field includes the subject words and categories of the knowledge point.

6. The recall method of the retrieval enhancement system according to claim 1, characterized in that: Based on the similarity of the vectors, find the documents or information that are most similar to the ES recall results.

7. The recall method of the retrieval enhancement system according to claim 1, characterized in that: Use the consistency judgment model to traverse the query and each recall result and give a consistency judgment result; if the model determines that the information in the recall result is consistent, the recall result will be retained, otherwise the recall result will be discarded.

8. A recall device of a retrieval enhancement system, characterized in that: The recall method based on the retrieval enhancement system in claim 1 comprises: The knowledge point storage module uses the supervised training topic model and clustering model to extract and analyze the features of the knowledge point, and obtain the subject words and cluster categories of the knowledge point; while storing the knowledge point itself, the subject words and cluster categories of the knowledge point are stored together as the feature fields of the knowledge point; The recall enhancement module provides an input for the user to input the query, and then feeds back the ES recall result after ES keyword matching and recall. The ES recall result is then converted into a vector. Finally, based on the similarity of the vector, the document or information most similar to the ES recall result is found. The recall result consistency check module uses the consistency judgment model to traverse the query and each recall result to make a consistency judgment on the obtained recall results. And the model summary module, the model summary module is used to put the recall results obtained after the consistency check into the generative model, and perform the final result summary to obtain the final recall result.

Images