Long document dense retrieval method and system based on topic clustering global features

Abstract

The application relates to the field of information retrieval in natural language processing technology, in particular to a long document dense retrieval method and system based on topic clustering global features, which cuts a target long document into several segments, obtains a local relevance score of the target long document based on the relevance between a preset query and dense vectors of each segment of the target long document; iteratively clusters the dense vectors of each segment of the target long document, and records clustering centers for representing topic information of the target long document in each iteration round; and obtains a global relevance score of the target long document based on the relevance between a preset query dense vector and dense vectors of each clustering center; and obtains a dense retrieval score of the target long document based on the local relevance score and the global relevance score. The application considers both local features and document topic global features in long documents, better captures complex hierarchical structures and topic features of long documents in the document retrieval process, and can improve the text information retrieval efficiency and performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information retrieval in natural language processing technology, and particularly relates to a long document dense retrieval method and system based on topic clustering global features. BACKGROUND

[0002] Document retrieval aims to search the most relevant documents to respond to queries, and is widely used in many fields of natural language processing, such as question answering, dialogue, recommendation system, etc. In the past few decades, sparse retrieval models (such as BM25) have dominated the field of information retrieval, mainly relying on the representation and discrete features of overlapping keywords in queries and documents to obtain relevant documents, which is difficult to fully capture the complex contextual relationship between texts and has limited ability to understand the problem of synonyms and polysemy in queries and documents. With the development of deep learning technology, researchers have tried to use deep neural networks to obtain continuous dense representation vectors of texts, thereby introducing the concept of dense retrieval models.

[0003] With the development of deep learning technology, especially pre-trained language models (such as BERT, GPT, etc.), it is possible to make deeper semantic modeling of documents and queries. By representing documents and queries as dense embedding vectors, deep learning models can better understand semantic relationships and thus improve the accuracy of retrieval. Compared with traditional sparse retrieval, dense retrieval no longer represents text as discrete words or features, but maps queries and documents in a low-dimensional space as continuous and dense vectors, which can better capture the semantic information between queries and documents and improve the effect of document retrieval. However, due to the length limitation of most pre-trained models on text input sequences, dense retrieval mainly focuses on the fragment retrieval task.

[0004] Pre-training models (PLMs) represented by BERT and large language models (LLMs) represented by ChatGPT are pre-trained through large-scale corpora, which enables them to have strong semantic understanding capabilities and learn more rich semantic information. This capability is crucial for improving the performance of information retrieval tasks. In particular, large models have made significant improvements in input sequence length compared to traditional dense retrieval models. However, there are still some potential difficulties in processing long document retrieval tasks with large models. On the one hand, long documents usually contain a large amount of information, some of which may not be important for a specific information retrieval task. This may cause the model to be disturbed by unnecessary information in the document, thereby reducing performance. On the other hand, due to the complex hierarchical structure of long text, including topic features, relationships between segments, and chapter structure, the complex hierarchical structure of the document affects the comprehensiveness of capturing document information, and there is a risk of information loss. Therefore, there is an urgent need for a solution that can balance the complex hierarchical structure of the document and the comprehensiveness of the information to optimize the effect of document information retrieval. SUMMARY

[0005] To this end, the present application provides a long document dense retrieval method and system based on topic clustering global features, which balances local features and document topic global features in long documents to better capture the complex hierarchical structure and topic features of long documents during document retrieval, thereby improving text information retrieval efficiency and performance.

[0006] According to the design scheme provided by the present application, on the one hand, a long document dense retrieval method based on topic clustering global features is provided, which comprises:

[0007] The target long document is divided into several segments, and the dense vectors of each segment in the long document are obtained and stored offline by a dense retrieval model. The score of the segment most relevant to the preset query dense vector in the target long document is obtained based on the similarity between the preset query and the dense vectors of each segment of the target long document, and the score of the segment is taken as the local relevance score of the target long document.

[0008] The dense vectors of each segment of the target long document are iteratively clustered, and the cluster centers representing the topic information of the target long document in each iteration round are recorded until the cluster centers in the long document are not updated; and the global relevance score of the target long document is obtained according to the relevance between the preset query dense vector and the dense vectors of each cluster center.

[0009] The local relevance score and the global relevance score of the target long document with the preset query are fused by a preset balanced fusion strategy to obtain the dense retrieval score of the target long document.

[0010] As the long document dense retrieval method based on the topic clustering global feature of the application, further, the target long document is cut into several segments, including:

[0011] Firstly, the matching score of the preset query and each document in the long document set is obtained by using the BM25 algorithm, the documents in the long document set are sorted according to the matching score, and the target long document related to the preset query is selected according to the threshold, wherein the long document set is composed of several documents in the information retrieval knowledge base;

[0012] Then, the target long document is cut into several segments of a preset length, and there is an overlap of a specified length between adjacent segments.

[0013] As the long document dense retrieval method based on the topic clustering global feature of the application, further, the most relevant segment score in the target long document to the preset query is obtained based on the relevance between the preset query dense vector and the dense vector of each segment of the target long document, including:

[0014] Firstly, the preset query dense vector and the dense vector of each segment of the target long document are obtained based on the query encoder and the document encoder respectively;

[0015] Then, the similarity score between the preset query dense vector and the dense vector of each segment of the target long document is calculated based on the cosine similarity, so as to obtain the local relevance score of the target long document according to the most relevant segment.

[0016] As the long document dense retrieval method based on the topic clustering global feature of the application, further, the dense vector of each segment of the target long document is iteratively clustered, and the iterative clustering process includes:

[0017] A segment dense vector is randomly selected from the dense vectors of each segment as an initial clustering center, the distance between the clustering center and other segment dense vectors is calculated, and the clustering result is split and merged according to the distance, so as to select the clustering center of the next iteration, until the clustering center is not updated.

[0018] As the long document dense retrieval method based on the topic clustering global feature of the application, further, the global relevance score of the target long document is obtained according to the relevance between the preset query dense vector and the dense vector of each clustering center, including:

[0019] Firstly, the relevance score between the preset query dense vector and the dense vector of each clustering center is calculated based on the cosine similarity;

[0020] Then, the relevance scores between the preset query dense vector and the dense vector of each clustering center are aggregated to obtain the global relevance score of the target long document.

[0021] As the long document dense retrieval method based on the topic clustering global feature of the application, further, the process of fusing the local relevance score and the global relevance score of the target long document with the preset query through the preset balance fusion strategy is represented as: Wherein, S d is the dense retrieval score of the target long document d, S max is the local relevance score, is the global relevance score, and a is an influence factor.

[0022] As the long document dense retrieval method based on the topic clustering global feature of the application, further, the process of fusing the local relevance score and the global relevance score of the target long document with the preset query through the preset balance fusion strategy is represented as:

[0023] The BM25 algorithm is used to obtain the matching score of the target long document related to the preset query, and the matching score is taken as the sparse retrieval score of the target long document.

[0024] According to the dense retrieval score and the sparse retrieval score of the target long document and through a preset adjustment factor, the final score of the retrieval reordering of the target long document is obtained.

[0025] In another aspect, the application also provides a long document dense retrieval system based on topic clustering global features, comprising: a local feature acquisition module, a global feature acquisition module and a feature fusion calculation module, wherein,

[0026] The local feature acquisition module is used to split the target long document into several segments, obtain and offline store the dense vectors of each segment of the long document through a dense retrieval model, obtain the segment score of the most relevant segment of the target long document with the preset query based on the similarity between the preset query and the dense vectors of each segment of the target long document, and take the segment score as the local relevance score of the target long document.

[0027] The global feature acquisition module is used to iteratively cluster the dense vectors of each segment of the target long document, record the clustering centers for representing the topic information of the target long document in each iteration round, and obtain the global relevance score of the target long document according to the relevance between the dense vector of the preset query and the dense vectors of each clustering center.

[0028] The feature fusion module is used to fuse the local relevance score and the global relevance score of the target long document with the preset query through a preset balance fusion strategy, and obtain the dense retrieval score of the target long document.

[0029] The application has the following beneficial effects:

[0030] The application represents a long document as a set of overlapping segments, and calculates semantic similarity between a query and each segment in the segment set, to obtain local information of the long document based on the semantic similarity; topic clustering vectors of the long document are obtained by clustering analysis on the segment set, and similarity between each topic vector and a query term group is calculated, and global information of the long document is obtained by aggregating all topic vector similarities, and long document dense retrieval scores are calculated by balancing the fusion strategy of local features and global topic features in the long document to balance the local information and global information of the long document, so that the long document retrieval can more comprehensively consider the document topic, and the long text retrieval performance can be improved at the premise of ensuring the document retrieval efficiency, which is convenient for deployment and implementation in NLP specific application fields such as question answering, computer vision, machine translation, and has good application prospect. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 A long document dense retrieval process based on topic clustering global features in the embodiment is shown;

[0032] Figure 2 A long document dense retrieval principle architecture in the embodiment is shown. DETAILED DESCRIPTION

[0033] To make the purpose, technical scheme and advantages of the application clearer and more apparent, the application will be further described in detail below with reference to the drawings and technical scheme.

[0034] In the information retrieval task, according to a target document set D, D={d1, d2,..., d i ,...,d n}, a query index is established, and the index includes term, document frequency, term frequency and other information, and the embodiment of the application, as shown in Figure 1 provides a long document dense retrieval method based on topic clustering global features, which includes:

[0035] S101, the target long document is divided into several segments, the dense vectors of each segment in the long document are obtained and stored offline by a dense retrieval model, the similarity between the preset query and the dense vectors of each segment of the target long document is obtained, the segment score most related to the preset query dense vector in the target long document is obtained, and the segment score is taken as the local relevance score of the target long document.

[0036] Specifically, the target long document is divided into several segments, which can be designed to include the following contents:

[0037] First, the BM25 algorithm is used to obtain the matching score between the preset query and each document in the long document set. The documents in the long document set are sorted according to the matching score, and the target long document related to the preset query is selected according to the threshold. The long document set consists of several documents in the information retrieval knowledge base.

[0038] Then, based on the preset segment length, the target long document is divided into several segments of the preset length, and there is an overlap of a specified length between adjacent segments.

[0039] When a user submits a query q, the BM25 algorithm in sparse retrieval can be used to calculate the BM25 match score between the query and each document. The documents are then sorted according to their BM25 match scores to obtain the first-round sorted search list, and the most relevant Top documents for the query are retrieved. k This document will be used as the target long document for the second round of dense retrieval.

[0040] See Figure 2 As shown, for the target long document d i Based on the analysis of its fragment length, the target long document d i It can be segmented into overlapping segments containing 150 terms each, {seg1, seg2, ..., seg...} i ,...,seg n},|seg i |=150,1≤i≤n, where the length of the overlapping part of the segment can be set to 75 terms.

[0041] Among them, the score of the segment in the target long document that is most relevant to the dense vector of the preset query is obtained based on the correlation between the preset query and the dense vector of each segment of the target long document, including:

[0042] First, based on the query encoder and document encoder, the preset query dense vector and the dense vector of each segment of the target long document are obtained respectively;

[0043] Then, based on dense vectors and using cosine similarity, the score of the segment most relevant to the preset query in the target long document is calculated, so as to obtain the local relevance score of the target long document based on the most relevant segment.

[0044] To improve the retrieval efficiency of long documents, in this embodiment, a document encoder (Document Encoder, E) can be used. D Encode each segment in the target long document into a fixed-length dense vector. The scores of the segments most relevant to the query terms are stored to select key local information for long documents. Similarly, query encoding (E) is used. Q The generator encodes the preset query into a corresponding dense vector E.q .

[0045] The long document local information uses cosine function to calculate the similarity score of the query condition word group and all segment vectors in the target long document, and selects the segment score (S max ) most related to the query condition word group as the local key feature score of the entire target long document, and the calculation process can be represented as:

[0046]

[0047] S102, iteratively cluster each segment dense vector of the target long document, record the clustering center used to represent the theme information of the target long document in each iteration round, and obtain the global relevance score of the target long document according to the relevance between the preset query dense vector and each clustering center dense vector.

[0048] Specifically, the iteration clustering process of each segment dense vector of the target long document can be designed to include:

[0049] A segment dense vector is randomly selected from each segment dense vector as an initial clustering center, the distance between the clustering center and other segment dense vectors is calculated, and the clustering result is split and merged according to the distance, so as to select the clustering center of the next iteration, until the clustering center is not updated.

[0050] For the query condition q and the target long document d i , based on the segment dense vector of the target long document The segment embedding set in the long document is analyzed by using the clustering iteration algorithm to obtain the clustering center vector of the long document, that is, the theme information of the target long document, wherein the iteration clustering algorithm can be described as:

[0051] Let A segment representation is randomly selected from the segment dense representation vector set as an initial clustering center C1, the distance between the remaining segment dense representation vectors in the segment dense representation vector and the clustering center is calculated, and the probability of being selected as the next clustering center is calculated. Finally, through multiple iterations, the k clustering centers of the long document are obtained, C={C1', C2',..., C i ',..., C k}.

[0052] Wherein, the global relevance score of the target long document is obtained according to the relevance between the preset query dense vector and each clustering center dense vector, which can be designed to include the following contents:

[0053] First, the relevance score between the preset query dense vector and each clustering center dense vector is calculated based on the cosine similarity;

[0054] Then, the relevance scores between the preset query dense vector and the cluster center dense vector are aggregated to obtain a global relevance score of the target long document.

[0055] The calculation process of the score of the query and each cluster center for the query condition q can be represented as:

[0056]

[0057] wherein E q represents the dense representation vector of the query, represents the dense representation vector of the cluster center of the document.

[0058] The global information score of the long document can be obtained by aggregating all topic vector information in the long document The aggregation process is represented as:

[0059]

[0060] S103, the local relevance score and the global relevance score of the target long document for the preset query are fused by a preset balance fusion strategy to obtain a dense retrieval score of the target long document.

[0061] The global feature based on the topic clustering can completely summarize all important information of the long document. In order to meet the needs of a specific query, a fusion strategy for balancing the local feature and the global topic feature in the long document is proposed in the embodiment, which fuses the key local feature of the target long document meeting the specific query condition requirement and the global feature based on the topic clustering. On the one hand, the needs of the query can be met to some extent, and the overall content in the long document is considered completely to avoid the "topic drift" phenomenon. On the other hand, the dense retrieval score S d of the long document is obtained by balancing the local feature and the global feature based on the topic clustering in the long document, so that the topics of the document can be considered more comprehensively and in detail when the long document is retrieved.

[0062] The specific calculation formula of the dense retrieval score of the long document can be represented as follows:

[0063]

[0064] wherein a is an influence factor, which plays a regulating role in the local feature of the long document to prevent the occurrence of the "topic drift" phenomenon.

[0065] The dense retrieval score of the long document can be obtained through the dense retrieval described above. In order to further improve the reordering performance of the long document retrieval, the sparse retrieval score based on the sparse retrieval model and the dense retrieval score can be combined in the embodiment of the present case, and the combination can be designed to include the following contents:

[0066] The BM25 algorithm is used to obtain the matching score of the target long document related to the preset query, and the matching score is taken as the sparse retrieval score of the target long document;

[0067] The final score of the target long document retrieval reordering is obtained according to the dense retrieval score and the sparse retrieval score of the target long document and through a preset adjustment factor.

[0068] The BM25 algorithm in the sparse retrieval is used to calculate the sparse retrieval score of the target long document according to the frequency of the preset query in the target long document and the phrase distribution in the target long document.

[0069] The final score of the long document is calculated by combining the sparse retrieval score based on the exact match of the term and the dense retrieval score The formula can be designed as follows:

[0070]

[0071] Among them, S d represents the dense retrieval score of semantic relevance, BM25(d) represents the sparse retrieval score in the exact match, and γ is an adjustment factor of the sparse retrieval score and the dense retrieval score.

[0072] Further, based on the above method, the embodiment of the present application also provides a long document dense retrieval system based on topic clustering global features, which comprises a local feature acquisition module, a global feature acquisition module and a feature fusion calculation module, wherein,

[0073] The local feature acquisition module is used to divide the target long document into several segments, obtain and offline store the dense vectors of each segment of the long document through the dense retrieval model, obtain the segment score of the target long document most relevant to the preset query dense vector based on the similarity between the preset query and the dense vector of each segment of the target long document, and take the segment score as the local relevance score of the target long document;

[0074] The global feature acquisition module is used to iteratively cluster the dense vectors of each segment of the target long document, record the clustering centers for representing the topic information of the target long document in each iteration round, and obtain the global relevance score of the target long document according to the relevance between the preset query dense vector and the dense vector of each clustering center;

[0075] The feature fusion module is configured to fuse the local relevance score and the global relevance score of the target long document with the preset query by a preset balanced fusion strategy to obtain a dense retrieval score of the target long document.

[0076] A series of comparative experiments were performed on two standard TREC collections, namely, Robust04 and GOV2 datasets, and detailed information of the datasets is shown in Table 1.

[0077] Table 1, statistical information of experimental datasets

[0078]

[0079]

[0080] To verify the performance of the dense retrieval DRSTC in the scheme, the DRSTC dense retrieval fusing global information of different numbers of topic features of long documents was compared with other information retrievals, and experimental results with significant changes were presented, and experimental results in different cases are shown in Table 2 and Table 3.

[0081] Table 2, performance results of the DRSTC dense retrieval method under different numbers of cluster centers on the GOV2 dataset and comparison with the baseline model

[0082]

[0083] Table 3, performance results of the DRSTC dense retrieval method under different numbers of cluster centers on the Robust04 dataset and comparison with the baseline model

[0084]

[0085] According to Table 2 and Table 3, on the datasets GOV2 and Robust04, the performance results of the DRSTC in the scheme under four different numbers of cluster centers are all better than those of the baseline model.

[0086] In addition, to reflect the contribution of global features based on topic clustering to performance improvement, the retrieval results of the dense retrieval DRSTC of the global features (global topic information STC) of the topic clustering in the scheme and the original dense retrieval are listed, and specific experimental results are shown in Table 4 and Table 5.

[0087] Table 4, comparison of semantic retrieval results of the DRSCM under four different numbers of cluster centers on the GOV2 dataset and the original dense retrieval model

[0088]

[0089] Table 5. Comparison of semantic retrieval results of DRSCM with original dense retrieval model on Robust04 dataset under four different numbers of cluster centers

[0090]

[0091] From Table 4 and Table 5, it can be seen that the dense retrieval scheme in the present application which fuses the topic clustering feature can better understand the central idea of long documents, making the retrieval results more accurate.

[0092] In summary, the above experimental results show that in the long document retrieval task, the present application scheme fully considers the global information in the long document, and the computational complexity is reduced from O(n 2 ) to O(n), the retrieval efficiency is improved by 3 orders of magnitude compared with the model based on cross-encoder, which is convenient for deployment and implementation in practical application fields, and has good application prospect.

[0093] Unless otherwise specified, the relative steps, numerical expressions and numerical values of the components and steps set forth in these embodiments do not limit the scope of the present application.

[0094] Each embodiment in the specification is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between each embodiment can be referred to each other. For the system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method part.

[0095] The units and method steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been described in the above description in general terms. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation does not exceed the scope of the present application.

[0096] Those skilled in the art can understand that all or part of the steps of the above method can be instructed by a program to complete the relevant hardware, and the program can be stored in a computer readable storage medium, such as a read-only memory, a magnetic disk or an optical disk, etc. Alternatively, all or part of the steps of the above embodiments can also be implemented using one or more integrated circuits, and accordingly, each module / unit in the above embodiments can be implemented in the form of hardware or in the form of a software function module. The present application is not limited to any specific form of combination of hardware and software.

[0097] Finally, it should be noted that the above-described embodiments are merely specific embodiments of the present application, which are used to illustrate the technical solutions of the present application, but not to limit the same. The protection scope of the present application is not limited thereto. Although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those skilled in the art that any person skilled in the art can still modify or easily think of changes to the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to some of the technical features, within the technical scope disclosed by the present application. The modifications, changes or replacements do not cause the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for dense retrieval of long documents based on global features of topic clustering, characterized in that, Include: The target long document is divided into several segments. Dense vectors of each segment in the long document are obtained and stored offline through a dense retrieval model. The score of the segment in the target long document that is most relevant to the dense vector of the preset query is obtained based on the similarity between the preset query and the dense vector of each segment in the target long document. The score of this segment is used as the local relevance score of the target long document. Iterative clustering is performed on the dense vectors of each segment of the target long document, and the cluster centers representing the topic information of the target long document are recorded in each iteration. The long document is obtained through multiple iterations. Cluster centers ; The global relevance score of the target long document is obtained based on the correlation between the preset query dense vector and the dense vector of each cluster center. The iterative clustering process includes: randomly selecting a fragment dense vector from each fragment dense vector as the initial cluster center, calculating the distance between the cluster center and other fragment dense vectors respectively, and splitting and merging the clustering results according to the distance to select the cluster center for the next round of iteration, until the current cluster center has not been updated. By using a pre-defined balanced fusion strategy, the local relevance scores and global relevance scores of the target long document with respect to the pre-defined query are fused together to obtain the dense retrieval score of the target long document.

2. The method for dense retrieval of long documents based on global features of topic clustering according to claim 1, characterized in that, The target long document is divided into several segments, including: First, the BM25 algorithm is used to obtain the matching score between the preset query and each document in the long document set. The documents in the long document set are sorted according to the matching score, and the target long document related to the preset query is selected according to the threshold. The long document set consists of several documents in the information retrieval knowledge base. Then, based on the preset segment length, the target long document is divided into several segments of the preset length, and there is an overlap of a specified length between adjacent segments.

3. The method for dense retrieval of long documents based on global features of topic clustering according to claim 1 or 2, characterized in that, Based on the correlation between the preset query and the dense vectors of each segment of the target long document, the dense vector of the segment most relevant to the dense vector of the preset query in the target long document is obtained, including: First, based on the query encoder and document encoder, the preset query dense vector and the dense vector of each segment of the target long document are obtained respectively; Then, based on cosine similarity, the similarity score between the preset query dense vector and the dense vector of each segment in the target long document is calculated, so as to obtain the local relevance score of the target long document based on the most relevant segment.

4. The method for dense retrieval of long documents based on global features of topic clustering according to claim 1, characterized in that, The global relevance score of the target long document is obtained based on the correlation between the preset query dense vector and the dense vectors of each cluster center, including: First, the correlation score between the preset query dense vector and the dense vector of each cluster center is calculated based on cosine similarity; Then, the relevance scores between the preset query dense vector and the dense vectors of each cluster center are aggregated to obtain the global relevance score of the target long document.

5. The method for dense retrieval of long documents based on global features of topic clustering according to claim 1, characterized in that, The process of fusing the local relevance score and the global relevance score of the target long document with the preset query using a preset balanced fusion strategy can be represented as follows: ,in, For dense retrieval of target long documents d, For local correlation scores, For global relevance score, It is the impact factor.

6. The method for dense retrieval of long documents based on global features of topic clustering according to claim 1, characterized in that, Also includes: The BM25 algorithm is used to obtain the target long document matching score related to the preset query, and this matching score is used as the target long document sparse retrieval score. The final score for re-ranking the target long document retrieval is obtained based on the dense retrieval score and sparse retrieval score of the target long document and through a preset adjustment factor.

7. A dense retrieval system for long documents based on global features of topic clustering, characterized in that, It includes: a local feature acquisition module, a global feature acquisition module, and a feature fusion calculation module. in, The local feature acquisition module is used to divide the target long document into several segments, acquire and store the dense vectors of each segment in the long document offline through a dense retrieval model, and obtain the score of the segment in the target long document that is most relevant to the dense vector of the preset query based on the similarity between the preset query and the dense vectors of each segment in the target long document. The score of this segment is used as the local relevance score of the target long document. The global feature acquisition module iteratively clusters the dense vectors of each segment of the target long document, recording the cluster centers representing the topic information of the target long document in each iteration. Through multiple iterations, the module obtains the cluster centers of the long document. Cluster centers ; The global relevance score of the target long document is obtained based on the correlation between the preset query dense vector and the dense vector of each cluster center. The iterative clustering process includes: randomly selecting a fragment dense vector from each fragment dense vector as the initial cluster center, calculating the distance between the cluster center and other fragment dense vectors respectively, and splitting and merging the clustering results according to the distance to select the cluster center for the next round of iteration, until the current cluster center has not been updated. The feature fusion module is used to fuse local relevance scores and global relevance scores through a preset balanced fusion strategy to obtain the dense retrieval score of the target long document.

8. An electronic device, characterized in that, include: At least one processor, and a memory coupled to said at least one processor; The memory stores a computer program that can be executed by the at least one processor to implement the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed, enables the implementation of the method as described in any one of claims 1 to 6.

Images