Document processing method and device and information retrieval method and device

Abstract

The embodiment of the invention provides a document processing method and device and an information retrieval method and device. According to the document processing method, the text content of a target document can be subjected to fragmentation processing to obtain a plurality of text fragments, then each text fragment is input into a first large language model for model processing to obtain a processing result, the processing result comprises a plurality of knowledge fragments, and then the target document is processed according to the knowledge fragments. And generating first tag data corresponding to a plurality of preset tag categories, and correspondingly storing each knowledge fragment and the first tag data in a first database, so that the first tag data is used for knowledge question and answer retrieval. In the information retrieval method, after a user question in a text form is determined according to an information retrieval request of a user, second label data corresponding to the plurality of label categories are generated according to the user question, and then the user question is retrieved based on first correlation between at least part of first label data in a first database and the second label data. Knowledge fragments related to the user question are retrieved.

Description

Technical Field

[0001] The embodiments of this specification relate to the field of computer technology, and in particular, to a document processing method and apparatus, and an information retrieval method and apparatus. Background Art

[0002] Currently, some documents contain complex and specialized content, making them difficult for many people to understand. For example, the complexity and specialized nature of insurance clauses in insurance contracts often hinder average consumers from understanding and choosing appropriate insurance products. Furthermore, the professionalism of insurance agents varies widely, making it difficult for them to provide high-quality customer service. Insurance agents have a limited capacity to handle customer needs, and when insurance platforms experience significant user growth, they are unable to effectively handle this demand, leading to user churn.

[0003] There is an urgent need for a reasonable and reliable solution that can help users quickly obtain the information they need from documents. Summary of the Invention

[0004] The embodiments of this specification provide document processing and information retrieval solutions that can help users quickly obtain required information from documents.

[0005] In a first aspect, an embodiment of the present specification provides a document processing method, comprising: segmenting the text content of a target document to obtain a plurality of text segments; for each of the plurality of text segments, inputting the text segment into a first large language model for model processing to obtain a processing result, which includes a plurality of knowledge segments; generating first label data corresponding to a plurality of preset label categories according to each knowledge segment; and storing each knowledge segment and the first label data in a first database so that the first label data can be used for knowledge question and answer retrieval.

[0006] In some embodiments, the text content of the target document is segmented to obtain multiple text fragments, including: segmenting the text content into multiple initial fragments based on a preset text segmentation strategy; determining the correlation between adjacent initial fragments in the multiple initial fragments, and merging the initial fragments whose correlation reaches a preset threshold, thereby obtaining the multiple text fragments.

[0007] In some embodiments, the several label categories include keyword labels; generating first label data corresponding to the preset several label categories according to each knowledge fragment includes: for each target fragment in the each knowledge fragment, based on a preset keyword recognition strategy, performing keyword recognition on multiple entities included in the target fragment to obtain a first recognition result; wherein, the multiple entities are obtained by performing entity recognition on the target fragment using a second largest language model; generating a first keyword label according to the first recognition result, and incorporating the first keyword label into the first label data of the target fragment.

[0008] In some embodiments, the several label categories include slot labels, and the slot labels involve several slot categories; generating first label data corresponding to the preset several label categories according to each knowledge fragment includes: for each target fragment in the each knowledge fragment, identifying several entities corresponding to the several slot categories from the multiple entities included in the target fragment; wherein the multiple entities are obtained by performing entity recognition on the target fragment using a second largest language model; generating a first slot label based on the several entities, and incorporating the first slot label into the first label data of the target fragment.

[0009] In some embodiments, the several label categories include intent labels; generating first label data corresponding to the preset several label categories according to each knowledge fragment includes: for each target fragment in each knowledge fragment, using a third language model to perform intent recognition on the target fragment to obtain a second recognition result; generating a first intent label based on the second recognition result, and incorporating the first intent label into the first label data of the target fragment.

[0010] In some embodiments, the method also includes: obtaining multiple first historical questions associated with the target document; for each target segment in the knowledge segments, classifying the first historical questions among the multiple first historical questions whose correlation with the target segment reaches a preset threshold into the question group corresponding to the target segment; selecting a question from the question group as a standard question, and using other questions in the question group as similar questions to the standard question; storing the standard question and the similar questions in correspondence in a second database, and at least storing the standard question in association with the target segment in the first database for knowledge question and answer retrieval.

[0011] In some embodiments, the target document is associated with multiple original second historical questions; obtaining multiple first historical questions associated with the target document includes: using the multiple second historical questions as the multiple first historical questions; or rewriting the multiple second historical questions to obtain the multiple first historical questions.

[0012] In some embodiments, the rewriting process includes at least one of the following: grammar correction, vocabulary replacement, and sentence structure adjustment.

[0013] In some embodiments, the processing result also includes a number of predicted questions corresponding to each of the several knowledge fragments; before selecting a question from the question group as the standard question, it also includes: classifying the predicted questions corresponding to the target fragment into the question group.

[0014] In some embodiments, the target document includes any one of the following documents: an agreement document, a news document, a social media document; wherein the agreement document includes an insurance contract document.

[0015] In a second aspect, an embodiment of the present specification provides an information retrieval method, comprising: determining a user question in text form based on a user's information retrieval request; generating second label data corresponding to several preset label categories based on the user question; determining a first correlation between at least part of the first label data in a first database and each of the second label data; wherein the first database stores a plurality of knowledge fragments and the plurality of knowledge fragments each correspond to the first label data of the several label categories; based on the first correlation, determining several first knowledge fragments related to the user question from the knowledge fragments corresponding to at least part of the first label data and classifying them into an answer set; and determining a retrieval result based on the answer set.

[0016] In some embodiments, determining the user question in text form based on the user's information retrieval request includes: in response to the information retrieval request including a question in text form, using the question as the initial question; in response to the information retrieval request including a question in non-text form, converting the question into the initial question in text form; and using the initial question or a rewritten result of the initial question as the user question.

[0017] In some embodiments, before determining the retrieval results based on the answer set, it also includes: determining the second correlation between each of the knowledge fragments and the user question; based on the second correlation, determining several second knowledge fragments related to the user question from the knowledge fragments and including them in the answer set.

[0018] In some embodiments, the first database also stores standard questions corresponding to each of the multiple knowledge fragments, and the standard questions and their similar questions are also stored in the second database; before determining the retrieval results based on the answer set, it also includes: determining the third correlation between at least some of the similar questions in the second database and the user question; based on the third correlation, determining several similar questions related to the user question from the at least some of the similar questions; based on the standard questions corresponding to each of the several similar questions, obtaining the third knowledge fragment corresponding to the standard question from the first database and incorporating it into the answer set.

[0019] In some embodiments, determining the retrieval result based on the answer set includes: sorting the answers in the answer set based on a target sorting method to obtain a first sorting result; and determining the retrieval result based on the first sorting result.

[0020] In some embodiments, the target sorting method is sorting by relevance; the sorting of the answers in the answer set includes: sorting the answers in the answer set in order of target relevance from high to low; wherein, the target relevance includes any one of the following: a first correlation between the first label data and the second label data of the answer in the answer set, and a second correlation between the answer and the user question.

[0021] In some embodiments, determining the search result based on the first ranking result includes: optimizing the ranking of the first ranking result based on a preset rearrangement strategy to obtain a second ranking result; and generating a search result including the second ranking result.

[0022] In some embodiments, the ranking optimization of the first ranking result based on a preset re-ranking strategy includes: obtaining the historical behavior data and / or preference data of the user related to information retrieval, and ranking optimization of the first ranking result according to the historical behavior data and / or preference data.

[0023] In some embodiments, the first sorting result is sorted and optimized based on a preset re-arrangement strategy, including: obtaining the text content of the target event; wherein the target event is a recent emergency or hot event; determining the target answer with the highest correlation with the text content in the first sorting result, and moving the target answer to the head of the first sorting result.

[0024] In a third aspect, an embodiment of the present specification provides a document processing device, comprising: a segmentation unit, configured to segment the text content of a target document to obtain a plurality of text fragments; a processing unit, configured to input each text fragment of the plurality of text fragments into a first large language model for model processing to obtain a processing result, which includes a plurality of knowledge fragments; a generation unit, configured to generate first label data corresponding to a plurality of preset label categories according to each knowledge fragment; and a storage unit, configured to store the knowledge fragments and the first label data in a first database so that the first label data can be used for knowledge question and answer retrieval.

[0025] In a fourth aspect, an embodiment of the present specification provides an information retrieval device, comprising: a first determination unit, configured to determine a user question in text form based on a user's information retrieval request; a generation unit, configured to generate second label data corresponding to several preset label categories based on the user question; a second determination unit, configured to determine a first correlation between at least part of the first label data in the first database and the second label data; wherein the first database stores a plurality of knowledge fragments and the plurality of knowledge fragments each correspond to the first label data of the several label categories; a third determination unit, configured to determine, based on the first correlation, from the knowledge fragments corresponding to at least part of the first label data, several first knowledge fragments related to the user question and include them in an answer set; a fourth determination unit, configured to determine a retrieval result based on the answer set.

[0026] In a fifth aspect, an embodiment of this specification provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed in a computer, the computer is caused to execute the method described in any one of the implementations of the first and second aspects.

[0027] In a sixth aspect, an embodiment of this specification provides a computing device comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, it implements the method described in any one of the implementation methods of the first and second aspects.

[0028] In a seventh aspect, an embodiment of this specification provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the method described in any one of the implementation modes of the first and second aspects.

[0029] In the solution provided by the above embodiment of this specification, the text content of the target document can be segmented to obtain multiple text fragments, and then each text fragment is input into the first large language model for model processing to obtain a processing result, which includes several knowledge fragments, and then based on each knowledge fragment, a first label data corresponding to a plurality of preset label categories is generated, and each knowledge fragment and the first label data are stored in a first database so that the first label data can be used for knowledge question and answer retrieval. Thus, with the help of the deep semantic understanding ability of the large language model and combined with the labeling technology, a first database containing multiple knowledge fragments and the first label data corresponding to the plurality of preset label categories of the plurality of knowledge fragments can be constructed based on a large amount of documents, and then the first database can be used for information retrieval to provide a knowledge fragment retrieval service based on the label data, helping users to quickly obtain the required information from the document. Specifically, in the information retrieval stage, after determining the user question in text form according to the user's information retrieval request, second label data corresponding to the plurality of label categories is generated according to the user question, and then based on the first correlation between at least part of the first label data and the second label data in the first database, the knowledge fragments related to the user question are retrieved, thereby quickly obtaining the retrieval results. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions of the multiple embodiments disclosed in this specification, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings described below are only the multiple embodiments disclosed in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0031] Figure 1 It is a schematic diagram of the processing process of the document processing stage and the information retrieval stage in the embodiment of this specification;

[0032] Figure 2 is a flow chart of the document processing method in the embodiment of this specification;

[0033] Figure 3 is a flow chart of the information retrieval method according to the embodiment of this specification;

[0034] Figure 4 Schematic diagram of the process of mining standard question-similar question pairs in the embodiment of this specification;

[0035] Figure 5 This is a flowchart of the process of mining standard question-similar question pairs in an embodiment of this specification;

[0036] Figure 6 is another flow chart of the information retrieval method according to an embodiment of this specification;

[0037] Figure 7 is a schematic structural diagram of a document processing device according to an embodiment of this specification;

[0038] Figure 8 It is a structural diagram of the information retrieval device in the embodiment of this specification. DETAILED DESCRIPTION

[0039] To help those skilled in the art better understand the technical solutions in this specification, the following will provide a clear and complete description of the technical solutions in the embodiments of this specification, in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of this specification, not all of them. All other embodiments derived by those skilled in the art based on the embodiments in this specification without creative effort shall fall within the scope of protection of this specification.

[0040] As mentioned earlier, some documents contain complex and specialized content, making them difficult for many to understand. For example, the complexity and specialized nature of insurance clauses in insurance contracts often hinder average consumers from understanding and choosing appropriate insurance products. Furthermore, the professionalism of insurance agents varies widely, making it difficult for them to provide high-quality customer service. Insurance agents have a limited capacity to handle customer needs, and when insurance platforms experience significant user growth, they are unable to effectively handle this demand, leading to user churn.

[0041] The embodiments of this specification provide a document processing and information retrieval solution that can help users quickly obtain the required information from documents. The solution can be divided into the document processing stage and the information retrieval stage. Figure 1 It is a schematic diagram of the processing process of the document processing stage and the information retrieval stage in the embodiment of this specification.

[0042] like Figure 1 As shown, the processing process of the document processing stage is shown above the dotted line, and the processing process of the information retrieval stage is shown below the dotted line. Among them, the document processing stage involves several documents to be processed, a large language model M1, and a database D1. It should be noted that the several documents can be documents in any retrieval scenario. For example, in an agreement terms retrieval scenario, the several documents can be several agreement documents. Furthermore, when the agreement terms retrieval scenario is an insurance terms retrieval scenario, the several agreement documents are specifically several insurance contract documents. For another example, in a news retrieval scenario, the several documents can be several news documents. For another example, in a social media content retrieval scenario, the several documents can be several social media documents.

[0043] The large language model M1 is used, for example, but not limited to, to extract knowledge fragments from input text fragments. The content of these knowledge fragments is easier to understand than the original text fragments. It should be noted that the large language model M1 can be trained based on multiple sample fragments in text form and the corresponding knowledge fragments for each of the multiple sample fragments.

[0044] The database D1 can be called a document library, which is used to store the knowledge fragments produced in the document processing stage and the associated information of the knowledge fragments, which includes but is not limited to the following: Figure 1 The label data L1 shown in .

[0045] like Figure 1 As shown, in the document processing stage, for each of the documents to be processed, the document can be used as the target document, and the text content of the target document can be segmented to obtain multiple text fragments. Afterwards, for each of the multiple text fragments, the text fragment can be input into the large language model M1 for model processing to obtain a processing result, which includes several knowledge fragments. Then, based on each knowledge fragment, label data L1 corresponding to several preset label categories can be generated. Among them, the several label categories include at least one of keyword labels, slot labels, and intent labels. Then, each knowledge fragment and its respective label data L1 can be stored in the database D1 accordingly, so that the label data L1 can be used for knowledge question and answer retrieval.

[0046] In the information retrieval stage, we can provide knowledge fragment retrieval services based on label data based on database D1. Figure 1 As shown, after receiving the user's information retrieval request, the user's question in text form can be determined based on the information retrieval request. Afterwards, label data L2 corresponding to the above-mentioned several label categories can be generated based on the user's question. Then, based on the correlation between at least part of the label data L1 in the database D1 and the label data L2 (hereinafter referred to as the first correlation), several knowledge fragments related to the user's question (hereinafter referred to as several first knowledge fragments) can be determined from the knowledge fragments corresponding to the at least part of the label data L1 and classified into the answer set, and the retrieval result can be determined based on the answer set. Subsequently, the retrieval result can be provided to the user.

[0047] It's important to note that the process described in this article, when applied to insurance policy search scenarios, enables both consumers and insurance professionals to more quickly find key information when reading insurance policies, reducing information sifting time and improving decision-making efficiency and quality. Furthermore, this process can be easily extended to search scenarios such as news and social media content, demonstrating strong generalization and compatibility.

[0048] Next, combine Figure 2 , introduces the specific processing process of the target document in the document processing stage. Among them, Figure 2 This is a flow chart of a document processing method in an embodiment of this specification. The method can be executed by any device, platform or device cluster with computing and processing capabilities, and includes steps S201-S207 as shown below.

[0049] like Figure 2 As shown, in step S201, the text content of the target document is segmented to obtain multiple text segments.

[0050] Specifically, the text content can be extracted from the target document first, and then the text content can be segmented to obtain multiple text fragments. The file format of the target document can be PDF, TXT, or picture, etc. When the target document is not an editable text document, such as a PDF document or a picture document, it can be converted into an editable text document first, and then the text content can be extracted. Furthermore, when the target document is a PDF document, the target document can be converted using PDF parsing technology. When the target document is a picture document, the target document can be converted using OCR (Optical Character Recognition) technology.

[0051] When the text content of the target document is segmented, the text content can be first segmented into multiple initial segments based on a preset text segmentation strategy. The text segmentation strategy can, for example, indicate that the entire document is divided into segments of the smallest granularity based on punctuation marks or document structure. It should be understood that the text segmentation strategy can be configured according to actual needs and is not specifically limited here. Thereafter, the correlation between adjacent initial segments in the multiple initial segments can be determined, and the initial segments whose correlation reaches a preset threshold can be merged to obtain multiple text segments.

[0052] In practice, the correlation between adjacent initial segments can be measured by comparing their similarities. Common correlation calculation methods include cosine similarity, Jaccard similarity, and TF-IDF (Term Frequency-Inverse Document Frequency). In addition, for ease of calculation, adjacent initial segments can be converted into numerical vector representations first, and then the correlation between the vector representations of adjacent initial segments can be calculated. It should be pointed out that in order to improve the effect of text vectorization, in the embodiments of this specification, a large language model can be combined to convert the text into a vector representation. Since the large language model can capture the complex relationships and long-term dependencies between words and convert the text into a high-dimensional dense vector, the vector not only contains word meaning information, but also contains syntactic structure and contextual meaning. Therefore, using a large language model to vectorize the text can achieve deep contextual vectorization of the text.

[0053] After the text content of the target document is divided into multiple text segments, information mining can be performed on the multiple text segments. As an implementation method, step S203 can be performed to mine knowledge segments from the text segments.

[0054] Specifically, in step S203, for each of the multiple text segments, the text segment is input into the large language model M1 for model processing to obtain a processing result, which includes several knowledge segments.

[0055] It should be noted that with the help of the deep semantic understanding ability of the large language model M1, several easy-to-understand knowledge fragments can be mined from each text fragment.

[0056] Next, in step S205 , label data L1 corresponding to a plurality of preset label categories are generated according to each knowledge segment.

[0057] Specifically, for each knowledge segment in each knowledge segment, the knowledge segment can be used as a target segment, and based on the target segment, label data L1 corresponding to the above-mentioned label categories is generated for the target segment. Among them, the above-mentioned label categories may include at least one of keyword labels, slot labels, and intent labels. The extraction of keyword labels can be achieved through the deep semantic understanding ability of the large language model, which can accurately identify the core concepts and themes in the text and provide a basis for subsequent retrieval. The intent label is used to represent the main purpose or intention of the text and can be determined by semantic analysis of the text. Slot labels usually involve several slot categories, which may include at least one of the amounts, products, and dates, for example. Slot labels can be used to identify specific entities in the text, such as amounts, products, and / or dates. By identifying these slots, the machine can better understand the text and provide targeted services.

[0058] When the above-mentioned several label categories include keyword labels and / or slot labels, for the target segment as mentioned above, the large language model M2 for entity recognition can be first used to perform entity recognition on the target segment to obtain multiple entities included in the target segment, and then the keyword label K1 and / or slot label S1 of the target segment are constructed based on the multiple entities, and then the keyword label K1 and / or slot label S1 are included in the label data L1 of the target segment.

[0059] When constructing the keyword tag K1 for the target segment, keyword recognition can first be performed on the multiple entities included in the target segment based on a preset keyword recognition strategy, resulting in a recognition result Rr1. This keyword recognition strategy may, for example, indicate that entities whose occurrences reach a threshold number of times are to be recognized as keywords. It should be understood that this keyword recognition strategy can be configured based on actual needs and is not specifically limited here. Subsequently, the keyword tag K1 for the target segment can be generated based on the recognition result Rr1.

[0060] When constructing the slot label S1 of the target segment, several entities corresponding to the above-mentioned several slot categories can be identified from the multiple entities included in the target segment, and then the slot label S1 of the target segment is generated according to the several entities.

[0061] When the aforementioned label categories include intent labels, the large language model M3 for intent recognition can be used to first perform intent recognition on the target segment, obtaining a recognition result Rr2. Based on the recognition result Rr2, an intent label I1 for the target segment can then be generated and included in the target segment's label data L1.

[0062] It's important to note that the large language model M3 is also used to identify user query intent during the subsequent information retrieval phase. Leveraging the large language model's deep semantic understanding capabilities, this solution can more accurately parse user query intent. Even for complex, ambiguous queries, or those containing industry-specific terminology, this solution accurately grasps user needs, improving the accuracy and comprehensiveness of query interpretation.

[0063] Next, in step S207, each knowledge segment and its label data L1 are stored in correspondence in the database D1, so that the label data L1 can be used for knowledge question and answer retrieval.

[0064] In one embodiment, to facilitate user retrieval of a specific document, after obtaining each knowledge fragment by executing step S203, an association relationship between the target document and each knowledge fragment can be established. As an example, the document identifier of the target document can be associated with each knowledge fragment and stored in database D1.

[0065] Figure 2 The solution provided by the corresponding embodiment can, with the help of the deep semantic understanding ability of the large language model and combined with labeling technology, build a database D1 based on massive documents, which contains multiple knowledge fragments and label data L1 corresponding to several preset label categories. The database D1 can then be used for information retrieval to provide knowledge fragment retrieval services based on label data, helping users quickly obtain the required information from documents.

[0066] Next, combine Figure 3 , introduces the knowledge fragment retrieval process based on label data. Among them, Figure 3 This is a flow chart of the information retrieval method in the embodiment of this specification. The method can be executed by any device, platform or device cluster with computing and processing capabilities, including steps S301-S309 as shown below.

[0067] like Figure 3 As shown, in step S301, a user question in text form is determined according to the user's information retrieval request.

[0068] It should be noted that, to meet the needs of different users, the solutions provided in the embodiments of this specification can support multiple input methods to achieve multimodal input. These multiple input methods include, but are not limited to, multiple of voice recognition, text input, and image recognition. These input methods can be combined to provide users with a more flexible and convenient operating experience.

[0069] Since multimodal input is supported, the questions included in the user's information retrieval request may be in text form or in non-text form (such as voice or image). In such a case, the initial question in text form can be determined based on the information retrieval request, and then the final user question can be determined based on the initial question. As an example, in response to the information retrieval request including a question in text form, the question can be used as the initial question, and in response to the information retrieval request including a question in non-text form, the question can be converted into an initial question in text form. Afterwards, the initial question or the rewritten result of the initial question can be used as the final user question.

[0070] In practice, to make it easier for machines to understand and process user requests, the initial question can be rewritten one or more times to improve the language expression of the user input. When rewriting the initial question, for example, at least one of grammatical correction, vocabulary replacement, and sentence structure adjustment can be performed.

[0071] Next, in step S303, label data L2 corresponding to several preset label categories are generated according to the user question.

[0072] Among them, the several label categories are the same as the several label categories in step S205 in the previous text, and the method for generating label data L2 is similar to the method for generating label data L1 described in the previous text. Please refer to the relevant instructions in the previous text and will not be repeated here.

[0073] Next, label recall can be performed by executing steps S305 and S307, thereby recalling several first knowledge fragments related to the user question.

[0074] Specifically, in step S305, the first correlation between at least part of the label data L1 in the database D1 and the label data L2 is determined; wherein the database D1 stores multiple knowledge fragments and the multiple knowledge fragments each correspond to the label data L1 of the above-mentioned several label categories.

[0075] Specifically, when a user's information retrieval request includes a document identifier, each knowledge fragment associated with the document identifier can be selected from database D1, and the label data L1 of each knowledge fragment can be used in the relevance calculation. When the user's information retrieval request does not include a document identifier, all label data L1 in database D1 can be used in the relevance calculation.

[0076] When determining the first correlation between each of the at least part of the label data L1 and the label data L2, the first correlation between the vector representation of each of the at least part of the label data L1 and the vector representation of the label data L2 may be specifically calculated.

[0077] In step S307 , based on the first correlation, several first knowledge fragments related to the user question are determined from the knowledge fragments corresponding to at least part of the label data L1 and are included in the answer set.

[0078] As an example, the knowledge fragments associated with the first correlation that reaches a preset threshold value among the knowledge fragments can be determined as the first knowledge fragments related to the user's question, and then the first knowledge fragments can be included in the answer set as the answer. As another example, the top K knowledge fragments associated with the highest first correlation among the knowledge fragments can all be determined as the first knowledge fragments related to the user's question, and then the first knowledge fragments can be included in the answer set as the answer. K can be a natural number greater than 1. It should be understood that the value of K can be configured according to actual needs and is not specifically limited here.

[0079] In practice, in addition to adopting the label recall method, other recall methods can also be adopted to achieve multi-way recall. As an example, the other recall method includes a recall method based on the relationship matching between user questions and knowledge fragments. Specifically, when adopting this recall method, the correlation between each knowledge fragment corresponding to at least part of the above-mentioned label data L1 and the user question (hereinafter referred to as the second correlation) can be determined. For example, the second correlation between the vector representation of each knowledge fragment and the vector representation of the user question is calculated, and then according to the second correlation, a number of knowledge fragments related to the user question (hereinafter referred to as the second knowledge fragments) are determined from the knowledge fragments and included in the answer set. Furthermore, after determining the several first knowledge fragments and the several second knowledge fragments as mentioned above, the merged result of the several first knowledge fragments and the several second knowledge fragments is included in the answer set. It should be noted that by calculating the second correlation between the vector representation of each knowledge fragment and the vector representation of the user question, a fast and high-quality preliminary recall can be achieved. This process not only takes into account the direct matching of vocabulary, but also deeply captures the similarity of context and syntactic structure, significantly improving the coverage and relevance of the recall.

[0080] Next, in step S309, the search result is determined according to the answer set.

[0081] Specifically, the answers in the answer set may be sorted based on the target sorting method to obtain a sorting result Sr1 , and the retrieval result may be determined based on the sorting result Sr1 .

[0082] The target sorting method can be the system default sorting method or the user-selected sorting method. The target sorting method can be any of the following: sorting by time, sorting by importance, sorting by relevance, etc. The relevance can be the relevance between tag data, or the relevance between user questions and knowledge fragments, and is not specifically limited here.

[0083] It should be noted that the solutions provided in the embodiments of this specification can provide users with a multi-dimensional filtering function, including various sorting methods such as sorting by time, sorting by importance, and sorting by relevance. Before submitting an information retrieval request, the user can select the desired sorting method. Based on this, when the information retrieval request includes a user-selected sorting method, the user-selected sorting method is used as the target sorting method. When the information retrieval request does not include a user-selected sorting method, the default sorting method is used as the target sorting method.

[0084] Taking the target sorting method of sorting by relevance as an example, the answers in the answer set can be sorted in descending order of target relevance. The target relevance includes any of the following: a first relevance between the label data L1 and the label data L2 of the answer in the answer set, and a second relevance between the answer and the user question.

[0085] By sorting the answers in the answer set according to the target sorting method, highly accurate sorting can be achieved, ensuring that the knowledge fragments that best meet user needs are presented first.

[0086] When determining the search results based on the ranking result Sr1, as one implementation method, the search results including the ranking result Sr1 can be generated. As another implementation method, the ranking result Sr1 can be first optimized based on a preset re-ranking strategy to obtain the ranking result Sr2, and then the search results including the ranking result Sr2 can be generated.

[0087] In one example, the reordering strategy may indicate optimization of the ranking based on the user's historical behavior data and / or preference data. Based on this, when optimizing the ranking result Sr1 based on the reordering strategy, the user's historical behavior data and / or preference data related to information retrieval may be obtained and used to optimize the ranking of the ranking result Sr1. Using such a reordering strategy, the order of displayed results can be dynamically adjusted, achieving personalized sorting and prioritizing the most relevant and valuable information, greatly enhancing the personalization and satisfaction of the user experience.

[0088] In another example, the reordering strategy may indicate that answers related to recent emergencies or hot topics should be placed at the top. Based on this, when optimizing the ranking result Sr1 based on the reordering strategy, the text content of the target event (the target event being the recent emergencies or hot topics) can be obtained. The target answer with the highest relevance to the text content in the ranking result Sr1 is then determined and moved to the top of the ranking result Sr1.

[0089] It should be understood that the content of the re-ranking strategy can be configured according to actual needs and is not specifically limited here. By optimizing the ranking result Sr1, the quality and timeliness of the search results can be further improved, ensuring that the system always maintains high availability and adaptability.

[0090] Figure 3 The corresponding embodiment provides a solution that, after determining a textual user question based on a user's information retrieval request, generates label data L2 corresponding to the aforementioned multiple label categories based on the user question. Then, based on the first correlation between at least a portion of the label data L1 in database D1 and the label data L2, retrieves knowledge fragments related to the user question, thereby obtaining retrieval results. By matching relationships between label data and conducting multi-label joint queries, more accurate knowledge fragment retrieval is achieved. The solution is particularly suitable for processing complex queries, helping users quickly obtain the information they need from documents.

[0091] Combined with the previous article Figure 2 This paper introduces the mining of knowledge fragments. In practice, during the document processing stage, question-answer pairs (QA) with knowledge fragments as answers can be mined. A represents the answer, and Q represents the question, specifically a Q-Q pair consisting of a standard question (abbreviated as the Q) and similar questions to the standard question (abbreviated as similar questions). By mining Q-Q pairs corresponding to knowledge fragments, similar question recall can also be performed during the information retrieval stage. For example, based on the correlation between the user's question and similar questions, knowledge fragments related to the user's question can be retrieved.

[0092] Take the example of a knowledge segment j extracted by the large language model M1 from a text segment i of the target document. Figure 4 As shown, the target document can be associated with multiple historical questions, which are derived from historical information retrieval requests related to the target document by several users. Figure 4 This is a schematic diagram of the question-similar question pair mining process in the embodiments of this specification. During the document processing phase, historical questions related to knowledge segment j can be identified from the multiple historical questions and assigned to the question group corresponding to knowledge segment j. Furthermore, the large language model M1 can also be used to predict questions corresponding to knowledge segments. Therefore, the processing results output for text segment i can also include several predicted questions corresponding to knowledge segment j, which can also be assigned to the question group corresponding to knowledge segment j.

[0093] After obtaining the question set corresponding to knowledge segment j, one question from the set can be selected as the standard question, and the other questions in the set can be designated as similar questions to the standard question. The standard question and the similar questions are then stored in database D2 as a set, and at least the standard question is stored in database D1 in association with knowledge segment j for use in knowledge question and answer retrieval. Database D2 can be referred to as a question library, which stores standard question-similar question pairs.

[0094] Next, we take the example of using the large language model M1 to predict the questions corresponding to the knowledge fragments, and each knowledge fragment corresponds to several prediction questions. Figure 5 , introduces the specific mining process of the labeled question-similar question pairs corresponding to the knowledge fragments. Figure 5 FIG. 5 is a flow chart of a process for mining standard question-similar question pairs in an embodiment of this specification. The mining process includes steps S501-S507 as shown below.

[0095] like Figure 5 As shown, in step S501, a plurality of first historical questions associated with the target document are obtained.

[0096] In practice, the target document is associated with multiple original historical questions (hereinafter referred to as multiple second historical questions). In one example, the multiple second historical questions can be directly used as multiple first historical questions to participate in the calculation of the correlation between the knowledge fragments. In another example, in order to improve the language expression of the user input and thus improve the accuracy of the calculation results, the multiple second historical questions can be rewritten to obtain multiple rewritten first historical questions. The rewriting process includes at least one of grammatical correction, vocabulary replacement, and sentence structure adjustment.

[0097] Then, in step S503, for each target segment in each knowledge segment obtained based on the target document, the first historical questions whose correlation with the target segment reaches a preset threshold, and the predicted questions corresponding to the target segment, are classified into the question group corresponding to the target segment.

[0098] Specifically, the correlation between the vector representation of each first historical question and the vector representation of the target segment can be calculated, and then the first historical question whose correlation with the target segment reaches a preset threshold is selected from the first historical questions, and then the first historical question and the predicted questions corresponding to the target segment are grouped into a question group corresponding to the target segment.

[0099] Next, in step S505 , one question is selected from the question group as a standard question, and the other questions in the question group are regarded as similar questions to the standard question.

[0100] When determining the standard question, for example, a question may be randomly selected from the question set as the standard question, or a question with the highest relevance to the target segment may be selected from the question set as the standard question. It should be understood that the standard question selection strategy can be configured based on actual needs and is not specifically limited here.

[0101] Next, in step S507, the standard question and similar questions are stored in the database D2 in correspondence, and at least the standard question is associated with the target segment and stored in the database D1 for use in knowledge question and answer retrieval.

[0102] Furthermore, when storing the standard question in the database D1, a standard question-similar question pair consisting of the standard question and its similar questions may be specifically associated with the target segment and stored in the database D1.

[0103] In one embodiment, to facilitate user retrieval of specific documents, after mining the target segment's question-similar question pair, an association relationship can be established between the target document and the question-similar question pair. As an example, the question-similar question pair can be associated with the document identifier of the target document and stored in database D2.

[0104] By also performing the following in the document processing phase Figure 5 The mining process shown can at least combine label recall and similar question recall in the information processing stage to achieve multi-way recall and improve the accuracy and comprehensiveness of information retrieval results.

[0105] Below, we take the three recall methods of label recall, similar question recall and recall based on the relationship matching between user questions and knowledge fragments as examples in the information retrieval stage. Figure 6 , introduces the information retrieval process. Among them, Figure 6 This is another flow chart of the information retrieval method in the embodiment of this specification. The method can be executed by any device, platform or device cluster with computing and processing capabilities, and includes steps S601-S615 as shown below.

[0106] like Figure 6 As shown, in step S601, a user question in text form is determined according to the user's information retrieval request.

[0107] In step S603, label data L2 corresponding to several preset label categories are generated according to the user question.

[0108] In step S605, the first correlation between at least part of the label data L1 in the database D1 and the label data L2 is determined; wherein the database D1 stores multiple knowledge fragments, standard questions of the multiple knowledge fragments and label data L1 corresponding to the above-mentioned multiple label categories.

[0109] In step S607 , based on the first correlation, several first knowledge fragments related to the user question are determined from the knowledge fragments corresponding to at least part of the label data L1 .

[0110] In step S609, the second correlation between each of the above knowledge segments and the user question is determined, and based on the second correlation, several second knowledge segments related to the user question are determined from the above knowledge segments.

[0111] For the explanation of steps S601-S609, please refer to the relevant description in the previous text, which will not be repeated here.

[0112] In step S611, the third correlation between at least some of the similar questions in database D2 and the user question is determined, and several similar questions related to the user question are determined from the at least some of the similar questions based on the third correlation, and according to the standard questions corresponding to each of the several similar questions, the third knowledge fragment corresponding to the standard question is obtained from database D1.

[0113] When the user's information retrieval request includes a document identifier, similar questions associated with the document identifier can be selected from database D2 to participate in the relevance calculation. When the user's information retrieval request does not include a document identifier, all similar questions in database D2 can be included in the relevance calculation.

[0114] When determining the correlation between each of the at least some of the similar questions and the user question (hereinafter referred to as the third correlation), the third correlation between the vector representation of each of the at least some of the similar questions and the vector representation of the user question can be specifically calculated. After calculating the third correlation, similar questions whose third correlation with the user question reaches a preset threshold can be determined as similar questions related to the user question; or, the top N similar questions with the highest third correlation with the user question can all be determined as similar questions related to the user question. Wherein, N can be a natural number greater than 1. It should be understood that the value of N can be configured according to actual needs and is not specifically limited here.

[0115] In step S613, the combined results of the first knowledge fragments, the second knowledge fragments and the third knowledge fragments are classified into an answer set.

[0116] In step S615, the search result is determined according to the answer set.

[0117] For the explanation of step S615, please refer to the relevant description of step S309 in the previous text, which will not be repeated here.

[0118] Combined with the previous article Figure 3 and Figure 6This section introduces the information retrieval solution provided by the embodiments of this specification. After providing information retrieval results to users, this solution also supports instant user feedback on the retrieval results. Based on this, this solution also allows for user feedback related to the retrieval results. This feedback is used to optimize the ranking algorithm, enabling the system to evolve over time, maintaining efficiency and accuracy.

[0119] Figure 7 : is a schematic diagram of the structure of the document processing device in the embodiment of this specification. The device can be applied to any device, platform or device cluster with computing and processing capabilities, and includes: a sharding unit 701, configured to perform sharding processing on the text content of the target document to obtain multiple text segments; a processing unit 702, configured to input each text segment of the multiple text segments into a first large language model for model processing to obtain a processing result, which includes multiple knowledge segments; a generating unit 703, configured to generate first label data corresponding to multiple preset label categories based on each knowledge segment; a storage unit 704, configured to store each knowledge segment and the first label data in a first database, so that the first label data can be used for knowledge question and answer retrieval.

[0120] In one embodiment, the segmentation unit 701 can be further configured to: segment the text content of the target document into multiple initial segments based on a preset text segmentation strategy; determine the correlation between adjacent initial segments in the multiple initial segments, and merge the initial segments whose correlation reaches a preset threshold, thereby obtaining the above-mentioned multiple text segments.

[0121] In one embodiment, the above-mentioned several tag categories include keyword tags; the generation unit 703 can be further configured to: for each target segment in each knowledge segment, based on a preset keyword recognition strategy, perform keyword recognition on multiple entities included in the target segment to obtain a first recognition result; wherein, the multiple entities are obtained by performing entity recognition on the target segment using the second largest language model; generate a first keyword tag based on the first recognition result, and classify the first keyword tag into the first tag data of the target segment.

[0122] In one embodiment, the above-mentioned several label categories include slot labels, and the slot labels involve several slot categories; the generation unit 703 can be further configured to: for each target segment in each knowledge segment, identify several entities corresponding to the several slot categories from the multiple entities included in the target segment; wherein the multiple entities are obtained by performing entity recognition on the target segment using the second largest language model; generate a first slot label based on the several entities, and include the first slot label in the first label data of the target segment.

[0123] In one embodiment, the above-mentioned several label categories include intent labels; the generation unit 703 can be further configured to: for each target segment in each knowledge segment, use the third language model to perform intent recognition on the target segment to obtain a second recognition result; generate a first intent label based on the second recognition result, and classify the first intent label into the first label data of the target segment.

[0124] In one embodiment, the above-mentioned device may also include: an acquisition unit (not shown in the figure), configured to acquire multiple first historical questions associated with the target document; a question mining unit (not shown in the figure), configured to, for each target segment in each knowledge segment, classify the first historical questions among the multiple first historical questions whose correlation with the target segment reaches a preset threshold into the question group corresponding to the target segment, select one question from the question group as the standard question, and use other questions in the question group as similar questions of the standard question; the storage unit 704 may also be configured to: store the standard question and the similar question in correspondence in the second database, and at least store the standard question in association with the target segment in the first database for knowledge question and answer retrieval.

[0125] In one embodiment, the target document is associated with multiple original second historical questions; the acquisition unit can be further configured to: use the multiple second historical questions as the multiple first historical questions; or rewrite the multiple second historical questions to obtain the multiple first historical questions.

[0126] In one embodiment, the rewriting process includes at least one of grammar correction, vocabulary replacement, and sentence structure adjustment.

[0127] In one embodiment, the above processing results also include a number of prediction questions corresponding to each of the above knowledge segments; the above question mining unit can also be configured to: classify the prediction questions corresponding to the target segment into the above question group.

[0128] In one embodiment, the target document includes at least one of an agreement document, a news document, and a social media document, etc.; wherein the agreement document includes an insurance contract document.

[0129] Figure 8: is a structural diagram of the information retrieval device in an embodiment of the present specification. The device can be applied to any device, platform or device cluster with computing and processing capabilities, and includes: a first determining unit 801, configured to determine a user question in text form according to a user's information retrieval request; a generating unit 802, configured to generate second label data corresponding to a plurality of preset label categories according to the user question; a second determining unit 803, configured to determine a first correlation between at least part of the first label data in the first database and the second label data; wherein the first database stores a plurality of knowledge fragments and the plurality of knowledge fragments each correspond to the first label data of the plurality of label categories; a third determining unit 804, configured to determine, according to the first correlation, from the knowledge fragments corresponding to the at least part of the first label data, a plurality of first knowledge fragments related to the user question and classify them into an answer set; a fourth determining unit 805, configured to determine a retrieval result according to the answer set.

[0130] In one embodiment, the first determination unit 801 can be further configured to: in response to the information retrieval request including a question in text form, use the question as the initial question; in response to the information retrieval request including a question in non-text form, convert the question into an initial question in text form; and use the initial question or a rewritten result of the initial question as the user question.

[0131] In one embodiment, the second determination unit 803 can also be configured to: determine the second correlation between each of the above-mentioned knowledge fragments and the user question; the third determination unit 804 can also be configured to: determine, based on the second correlation, several second knowledge fragments related to the user question from the above-mentioned knowledge fragments and include them in the answer set.

[0132] In one embodiment, the first database also stores standard questions corresponding to each of the above-mentioned multiple knowledge fragments, and the standard questions and their similar questions are also stored in the second database; the second determination unit 803 can also be configured to: determine the third correlation between at least some of the similar questions in the second database and the user question; the third determination unit 804 can also be configured to: determine several similar questions related to the user question from the at least some of the similar questions based on the third correlation, and according to the standard questions corresponding to each of the several similar questions, obtain the third knowledge fragment corresponding to the standard question from the first database and include it in the answer set.

[0133] In one embodiment, the fourth determining unit 805 may be further configured to: sort the answers in the answer set based on a target sorting method to obtain a first sorting result, and determine the search result according to the first sorting result.

[0134] In one embodiment, the target sorting method is sorting by relevance; the fourth determination unit 805 can be further configured to: sort the answers in the answer set in order of target relevance from high to low; wherein the target relevance includes any of the following: a first correlation between the first label data and the second label data of the answer in the answer set, and a second correlation between the answer and the user question.

[0135] In one embodiment, the fourth determining unit 805 may be further configured to: optimize the ranking of the first ranking result based on a preset rearrangement strategy to obtain a second ranking result; and generate a search result including the second ranking result.

[0136] In one embodiment, the fourth determining unit 805 may be further configured to: obtain historical behavior data and / or preference data of the user related to information retrieval, and optimize the ranking of the first ranking result according to the historical behavior data and / or preference data.

[0137] In one embodiment, the fourth determination unit 805 can be further configured to: obtain the text content of the target event; wherein the target event is a recent emergency event or hot event; determine the target answer with the highest correlation with the text content in the first sorting result, and move the target answer to the head of the first sorting result.

[0138] The embodiment of the present specification also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed in a computer, the computer is caused to execute the following Figure 2 、 Figure 3 、 Figure 5 、 Figure 6 Described method.

[0139] The embodiment of this specification also provides a computing device, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the following is achieved: Figure 2 、 Figure 3 、 Figure 5 、 Figure 6 Described method.

[0140] The embodiments of this specification also provide a computer program product, including a computer program / instruction, which, when executed by a processor, implements the following Figure 2 、 Figure 3 、 Figure 5 、 Figure 6 Describe the steps of the method.

[0141] Those skilled in the art will appreciate that, in one or more of the above examples, the functions described in the various embodiments disclosed in this specification may be implemented using hardware, software, firmware, or any combination thereof. When implemented using software, these functions may be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium.

[0142] In some cases, the actions or steps recited in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. Additionally, the processes depicted in the accompanying drawings do not necessarily require the particular order shown or sequential order to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0143] The specific implementation methods described above further illustrate in detail the purposes, technical solutions and beneficial effects of the multiple embodiments disclosed in this specification. It should be understood that the above description is only the specific implementation methods of the multiple embodiments disclosed in this specification, and is not intended to limit the protection scope of the multiple embodiments disclosed in this specification. Any modifications, equivalent replacements, improvements, etc. made on the basis of the technical solutions of the multiple embodiments disclosed in this specification should be included in the protection scope of the multiple embodiments disclosed in this specification.

Claims

1. A document processing method, comprising: Segment the text content of the target document to obtain multiple text segments; For each text segment of the plurality of text segments, input the text segment into a first large language model for model processing to obtain a processing result, which includes a plurality of knowledge segments; According to each knowledge fragment, first label data corresponding to a plurality of preset label categories are generated; The knowledge fragments and the first label data are stored correspondingly in a first database, so that the first label data can be used for knowledge question and answer retrieval.

2. The method according to claim 1, wherein: The text content of the target document is segmented to obtain multiple text segments, including: Based on a preset text segmentation strategy, segment the text content into a plurality of initial segments; The correlation between adjacent initial segments among the multiple initial segments is determined, and the initial segments whose correlation reaches a preset threshold are merged to obtain the multiple text segments.

3. The method according to claim 1, wherein: The several tag categories include keyword tags; The step of generating first label data corresponding to a plurality of preset label categories according to each knowledge fragment comprises: For each target segment in each of the knowledge segments, based on a preset keyword recognition strategy, keyword recognition is performed on multiple entities included in the target segment to obtain a first recognition result; wherein the multiple entities are obtained by performing entity recognition on the target segment using a second language model; A first keyword tag is generated according to the first recognition result, and the first keyword tag is included in the first tag data of the target segment.

4. The method according to claim 1, wherein: The several label categories include slot labels, and the slot labels relate to several slot categories; The step of generating first label data corresponding to a plurality of preset label categories according to each knowledge fragment comprises: For each target segment in each of the knowledge segments, identifying a plurality of entities corresponding to the plurality of slot categories from a plurality of entities included in the target segment; wherein the plurality of entities are obtained by performing entity recognition on the target segment using a second largest language model; A first slot tag is generated according to the plurality of entities, and the first slot tag is included in the first tag data of the target segment.

5. The method according to claim 1, wherein: The several tag categories include intent tags; The step of generating first label data corresponding to a plurality of preset label categories according to each knowledge fragment comprises: For each target segment in the knowledge segments, using the third language model to perform intent recognition on the target segment to obtain a second recognition result; A first intent tag is generated according to the second recognition result, and the first intent tag is included in the first tag data of the target segment.

6. The method according to claim 1, further comprising: Acquire a plurality of first historical questions associated with the target document; For each target segment in each of the knowledge segments, the first historical questions whose correlation with the target segment reaches a preset threshold among the plurality of first historical questions are classified into a question group corresponding to the target segment; Selecting one question from the question group as a standard question, and using other questions in the question group as similar questions to the standard question; The standard question and the similar question are stored in a second database in correspondence, and at least the standard question is associated with the target segment and stored in the first database for knowledge question and answer retrieval.

7. The method according to claim 6, wherein: The target document is associated with the original plurality of second historical questions; The obtaining of a plurality of first historical questions associated with the target document includes: Using the plurality of second historical questions as the plurality of first historical questions; or, The plurality of second historical questions are rewritten to obtain the plurality of first historical questions.

8. The method according to claim 7, wherein: The rewriting process includes at least one of the following: grammar correction, vocabulary replacement, and sentence structure adjustment.

9. The method according to claim 6, wherein: The processing result also includes a number of prediction questions corresponding to each of the several knowledge fragments; Before selecting a question from the question set as the standard question, it also includes: Each prediction problem corresponding to the target segment is classified into the problem group.

10. The method according to claim 1, wherein: The target document includes any of the following documents: an agreement document, a news document, and a social media document; wherein the agreement document includes an insurance contract document.

11. An information retrieval method comprising: Determine the user's question in text form based on the user's information retrieval request; Generating second label data corresponding to a plurality of preset label categories according to the user question; Determine a first correlation between at least part of the first label data in the first database and the second label data; wherein the first database stores a plurality of knowledge fragments and the plurality of knowledge fragments each correspond to the first label data of the plurality of label categories; According to the first correlation, from the knowledge fragments corresponding to at least part of the first tag data, determining several first knowledge fragments related to the user question to be included in the answer set; A search result is determined according to the answer set.

12. The method according to claim 11, wherein: Determining the user's question in text form according to the user's information retrieval request includes: In response to the information retrieval request including a question in text form, using the question as an initial question; In response to the information retrieval request including a question in non-text form, converting the question into an initial question in text form; The initial question or a rewritten result of the initial question is used as the user question.

13. The method according to claim 11, wherein: Before determining the search result according to the answer set, the method further includes: Determining a second correlation between each of the knowledge fragments and the user question; According to the second correlation, several second knowledge fragments related to the user question are determined from the various knowledge fragments and included in the answer set.

14. The method according to claim 11, wherein: The first database also stores standard questions corresponding to each of the plurality of knowledge fragments, and the standard questions and similar questions are also correspondingly stored in the second database; Before determining the search result according to the answer set, the method further includes: determining a third correlation between each of at least some of the similar questions in the second database and the user question; According to the third correlation, determining a number of similar questions related to the user question from the at least some of the similar questions; According to the standard questions corresponding to the several similar questions, the third knowledge fragments corresponding to the standard questions are obtained from the first database and included in the answer set.

15. The method according to any one of claims 11 to 14, wherein: Determining the search result according to the answer set includes: Based on the target sorting method, sort the answers in the answer set to obtain a first sorting result; A search result is determined according to the first sorting result.

16. The method according to claim 15, wherein: The target sorting method is sorting by relevance; The sorting of the answers in the answer set includes: The answers in the answer set are sorted in descending order of target relevance; wherein the target relevance includes any one of the following: a first correlation between the first label data and the second label data of the answer in the answer set, and a second correlation between the answer and the user question.

17. The method according to claim 15, wherein: The determining of the search result according to the first ranking result includes: Based on a preset rearrangement strategy, the first sorting result is sorted and optimized to obtain a second sorting result; Generate a search result including the second ranked result.

18. The method according to claim 17, wherein: The step of optimizing the sorting of the first sorting result based on a preset rearrangement strategy includes: Acquire historical behavior data and / or preference data of the user related to information retrieval, and optimize the first sorting result according to the historical behavior data and / or preference data.

19. The method according to claim 17, wherein: The step of optimizing the sorting of the first sorting result based on a preset rearrangement strategy includes: Obtaining the text content of a target event; wherein the target event is a recent emergency or hot event; A target answer having the highest correlation with the text content is determined in the first sorting results, and the target answer is moved to the head of the first sorting results.

20. A document processing device, comprising: A fragmentation unit is configured to fragment the text content of the target document to obtain multiple text fragments; A processing unit is configured to input each text segment of the plurality of text segments into a first large language model for model processing to obtain a processing result including a plurality of knowledge segments; A generating unit is configured to generate first label data corresponding to a plurality of preset label categories according to each knowledge fragment; The storage unit is configured to store the knowledge fragments and the first label data in a first database in correspondence, so that the first label data can be used for knowledge question and answer retrieval.

21. An information retrieval device, comprising: A first determining unit is configured to determine a user question in text form according to the user's information retrieval request; A generating unit, configured to generate second label data corresponding to a plurality of preset label categories according to the user question; A second determining unit is configured to determine a first correlation between at least part of the first label data in the first database and the second label data; wherein the first database stores a plurality of knowledge fragments and the plurality of knowledge fragments each correspond to the first label data of the plurality of label categories; A third determining unit is configured to determine, according to the first correlation, from the knowledge fragments corresponding to at least part of the first tag data, that a number of first knowledge fragments related to the user question are included in the answer set; The fourth determining unit is configured to determine a search result according to the answer set.

22. A computer program product, comprising a computer program / instruction, which, when executed by a processor, implements the steps of the method according to any one of claims 1 to 19.

23. A computing device comprising a memory and a processor, wherein: The memory stores executable codes, and when the processor executes the executable codes, the method according to any one of claims 1 to 19 is implemented.

Images