Retrieval method and electronic equipment

Abstract

The invention discloses a retrieval method and electronic equipment. The retrieval method comprises the steps that on the basis of a query request, the query request comprises structure information such as entities, attributes and constraints, and the structure information indicates that a single recall channel cannot cover a complete evidence set, multiple target recall channels are determined from preset recall channels; wherein the preset recall channel comprises a row-level recall channel, a column-level recall channel, a table recall channel and a text recall channel; retrieving to-be-retrieved data in parallel based on the plurality of target recall channels to obtain candidate retrieval results corresponding to the target recall channels; and outputting the target retrieval result. Wherein the target retrieval result is determined from the candidate retrieval results, and the coverage rate and the accuracy rate of complex document retrieval are improved.

Description

Technical Field

[0001] This application relates to the field of information retrieval technology, specifically to a retrieval method and electronic device. Background Technology

[0002] With the advent of the information age, information retrieval has become increasingly important, and users' demands for information retrieval are also growing. Currently, information retrieval typically employs single-channel recall strategies, such as keyword or vector-based recall. However, single-channel recall strategies often result in poor recall results, meaning lower retrieval accuracy and a poor user experience. Summary of the Invention

[0003] This application provides a retrieval method and an electronic device to improve retrieval accuracy.

[0004] On the one hand, embodiments of this application provide a retrieval method, the method comprising: Based on the query request, multiple target recall channels are determined from the preset recall channels; wherein, the preset recall channels include row-level recall channels, column-level recall channels, table-level recall channels and text-level recall channels; Based on each of the target recall channels, the data to be retrieved is retrieved to obtain the candidate retrieval results corresponding to each of the target recall channels; Output the target search result; wherein the target search result is determined from the candidate search results.

[0005] In some embodiments, determining a plurality of target recall channels from preset recall channels includes: Calculate the matching degree between the query request and each preset recall channel; Multiple target recall channels are determined from the preset recall channels based on the matching degree.

[0006] In some embodiments, determining a plurality of target recall channels from preset recall channels includes: Calculate the matching degree between the query request and each preset recall channel; Multiple target recall channels are determined from the preset recall channels based on the matching degree.

[0007] In some embodiments, the aforementioned multiple target recall channels include a main recall channel and an auxiliary recall channel; accordingly, based on the main recall channel and the auxiliary recall channel, the data to be retrieved is retrieved in parallel to obtain the candidate retrieval results corresponding to the main recall channel and the candidate retrieval results corresponding to the auxiliary recall channel, respectively.

[0008] Optionally, the recall ratio of the primary recall channel is higher than that of the secondary recall channel to ensure the accuracy of the recall results.

[0009] Optionally, the model size of the primary recall channel is larger than that of the secondary recall channel to ensure the accuracy and coverage of the primary recall channel.

[0010] In addition, the resource allocation priority of the primary recall channel is higher than that of the secondary recall channel to ensure the performance of the primary recall channel.

[0011] Optionally, determining multiple target recall channels from the preset recall channels based on the matching degree includes: Use the preset recall channel with the highest matching degree as the main recall channel; Auxiliary recall channels are determined from the remaining preset recall channels other than the preset recall channel with the highest matching degree. For example, all remaining preset recall channels are used as auxiliary recall channels; or, for example, some remaining preset recall channels are used as auxiliary recall channels, such as the preset recall channel with the highest matching degree among the remaining preset recall channels.

[0012] In some embodiments, the above row-level recall channel is indexed based on row-level blocks of tables in the data to be retrieved.

[0013] The column-level recall channel is based on indexes created using field titles and unit information.

[0014] The table-based recall channel performs overall recall based on the table's contextual features.

[0015] The text recall channel performs semantic retrieval based on a vector embedding model.

[0016] In some embodiments, before outputting the target retrieval result, the method includes: The candidate search results are fused based on multiple dimensions to obtain a fusion score for each candidate search result; wherein, the multiple dimensions include at least two of the following: semantic relevance between the candidate search results and the query request, attribute consistency between the candidate search results and the query request, and data coverage. M candidate search results are selected from the candidate search results to obtain the target search result; wherein the fusion score of the M candidate search results is higher than the fusion score of the candidate search results other than the M candidate search results.

[0017] In some embodiments, the step of fusing the candidate search results based on multiple dimensions to obtain a fusion score corresponding to each candidate search result includes: For each of the candidate search results, the following is adopted: Calculate the fusion score corresponding to the candidate search result; wherein, the Score represents the fusion score corresponding to the candidate search result; This indicates the semantic relevance between the candidate search results and the query request. This indicates that the candidate search results are consistent with the attributes of the query request. The data coverage is represented by w1, w2, and w3.

[0018] On the other hand, embodiments of this application also provide a computer program product, including a computer program or instructions, which, when executed by a processor, implement the steps in any of the retrieval methods provided in embodiments of this application.

[0019] On the other hand, embodiments of this application also provide a computer-readable storage medium storing a computer program or instructions thereon, including a computer program or instructions that, when executed by a processor, implement the steps in any of the retrieval methods provided in embodiments of this application.

[0020] The retrieval method provided in this application embodiment involves an electronic device determining multiple target recall channels matching the query request from a preset recall channel pool. These preset recall channels include row-level recall channels, column-level recall channels, table-level recall channels, and text-level recall channels. The electronic device can then perform parallel retrieval of the data to be retrieved based on these multiple target recall channels, obtaining candidate retrieval results corresponding to each target recall channel. This multi-channel retrieval enables the use of different recall methods, ensuring comprehensiveness and improving retrieval coverage. Finally, the electronic device determines the target retrieval result from the candidate retrieval results, ensuring the accuracy of the target retrieval result and thus improving retrieval accuracy and user satisfaction. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of a multi-channel recall fusion system provided in the embodiments of this application; Figure 2 This is a flowchart illustrating a retrieval method provided in an embodiment of this application; Figure 3 This is an illustration of a task execution scenario provided in an embodiment of this application. Figure 1 ; Figure 4 This is an illustration of a task execution scenario provided in an embodiment of this application. Figure 2 ; Figure 5 This is a schematic diagram of a fusion process provided in an embodiment of this application; Figure 6 This is a schematic diagram of a weight determination process provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0023] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] In the following description, specific embodiments of the invention will be illustrated with reference to steps and symbols performed by one or more computers, unless otherwise stated. Therefore, these steps and operations will be referred to several times as being performed by a computer, and computer execution as referred to herein includes operations by a computer processing unit representing electronic signals of data in a structured format. This operation transforms the data or maintains it at a location in the computer's memory system, which can be reconfigured or otherwise alter the operation of the computer in a manner well known to those skilled in the art. The data structure maintained by the data is the physical location of the memory, which has specific characteristics defined by the data format. However, the principles of the invention described above are not intended to be limiting, and those skilled in the art will understand that many of the steps and operations described below can also be implemented in hardware.

[0025] The terms "module" or "unit" as used herein can be considered as software objects executing on the computing system. The various components, modules, engines, and services described herein can be considered as implementations on the computing system. While the apparatus and methods described herein are preferably implemented in software, they can also be implemented in hardware, both of which are within the scope of this invention.

[0026] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when an element is “connected” or “coupled” to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein may include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0027] In document scenarios such as industry, energy, manufacturing, and government, information is often distributed across different granular structural levels: row level (records or samples); column level (parameters or indicators); table level (contextual statistics); and text paragraphs (descriptive content). In some embodiments, a single-channel recall scheme is typically used to retrieve information from documents, such as through keyword or vector recall. However, single-channel recall schemes have the following problems: 1) The recall results are insufficiently comprehensive, making it difficult to cover both structured tables and text paragraphs; 2) The lack of a multi-granularity fusion strategy makes it impossible to assess the contextual integrity of the recall results, which means it is impossible to assess the coverage.

[0028] 3) It cannot dynamically balance semantic relevance with attribute (or alternative description) consistency.

[0029] Therefore, to address the above problems, this application provides a multi-channel recall fusion system to achieve a multi-channel recall fusion method, such as... Figure 1 As shown, the multi-channel recall fusion system consists of an intent recognition module, a multi-channel recall module, a fusion scoring module, and an output module.

[0030] The intent recognition module is used to identify the query intent entered by the user. The multi-channel recall module performs recall through parallel row-level, column-level, whole-table, and text recall channels to obtain the recall results corresponding to the query intent.

[0031] The fusion scoring module introduces a fusion scoring mechanism to perform multi-granular fusion scoring on the recall results, determine the target recall results (i.e. target retrieval results) that are highly matched with the query intent, achieve high recall rate and high relevance retrieval for complex documents, ensure the coverage of recall results, and improve the accuracy of recall results.

[0032] The output module can output the target search results.

[0033] The following section will detail the implementation process of the aforementioned multi-channel recall fusion method, which is essentially the retrieval method. Figure 2 As shown, the implementation process may include: S201, The electronic device receives a query request.

[0034] The query request includes a query statement, which can be sent by other devices or entered by the user. The electronic device performs task intent recognition based on the query statement to determine the query intent.

[0035] In some embodiments, electronic devices can also perform structured analysis of query requests through an intelligent agent to determine request structure information.

[0036] The request structure information includes multiple core entities (Anchor), target attributes (Attribute), and constraints.

[0037] Core entities include entities in the query request, such as well name, project number, object identifier, etc.

[0038] The target attribute includes one or more of the following: numeric attributes, conditional attributes, and conclusion attributes in the query request.

[0039] Numeric attributes represent quantifiable data or parameters that appear directly in the query request.

[0040] Conditional attributes represent the background, status, or prerequisites carried in the query request.

[0041] Conclusive attributes indicate the final retrieval result indicated by the query request, and are typically evaluative, decision-making, or descriptive of the final state.

[0042] For example, if measure A is implemented starting in 2023, whether the output of product A significantly increased before and after the measure is an output-type attribute, the implementation of measure A starting in 2023 is a conditional attribute, and whether the output increased is a conclusion-type attribute.

[0043] Constraints represent the constraint information carried in the query request, that is, the constraints that the final search results (such as candidate search results) must follow. Optionally, constraints may include one or more of the following: structural constraints, search scope constraints, and numerical constraints. Examples include time range, chapter restrictions, maximum, minimum, and comparison requirements.

[0044] Furthermore, the aforementioned electronic devices can be terminals (such as mobile phones, tablets, laptops, etc.) or cloud servers. For example, taking a mobile phone as an example, ... Figure 3 As shown, electronic device 110 receives a query request input by the user. For example, if the electronic device is a server, such as... Figure 4As shown, electronic device 120 can receive query requests sent by mobile phone 130.

[0045] S202. Based on the query request, the electronic device determines multiple target recall channels from the preset recall channels. Among them, the preset recall channels include row-level recall channels, column-level recall channels, table-level recall channels, and text-level recall channels.

[0046] Each of the aforementioned preset recall channels uses an independent index. Specifically, the row-level recall channel builds its index based on row-level blocks in the table of the data to be retrieved.

[0047] The column-level recall channel is based on indexes created using field titles and unit information.

[0048] The table recall channel (or whole table recall channel) is based on the contextual features of the table for overall recall.

[0049] The text recall channel (or plain text recall channel) performs semantic retrieval based on a vector embedding model.

[0050] In addition, each of the above-mentioned preset recall channels also adopts an independent recall strategy. Simply put, it uses an independent algorithm / model for retrieval and recall.

[0051] In this embodiment, the electronic device can determine the data to be retrieved corresponding to the query request, such as the document to be retrieved. This data to be retrieved can be user-specified, or it can be automatically determined by the electronic device. For example, the information of the data to be retrieved (such as location and name) is carried in the query request, or it can be data matching the query request, such as a document, used as the data to be retrieved.

[0052] Furthermore, the electronic device determines multiple target recall channels matching the query request from a preset recall channel set, thereby achieving the determination of multiple recall channels. The query request includes structural information such as entities, attributes, and constraints, and this structural information indicates that a single recall channel cannot cover the complete evidence set; therefore, multiple target recall channels are determined from the preset recall channels.

[0053] In some embodiments, the target recall channels can be determined based on matching scores. The electronic device calculates the matching score between the query request and each preset recall channel. Then, the electronic device determines multiple target recall channels from the preset recall channels based on the matching scores, thereby determining the target recall channels that match the query intent, thus ensuring recall accuracy.

[0054] For example, each preset recall channel has a corresponding example task template. Electronic devices can calculate the semantic similarity between the query request and the example task model corresponding to the preset recall channel to determine the matching degree between the query request and the preset recall channel. In other words, the matching degree can represent semantic similarity.

[0055] Optionally, the electronic device can determine the target recall channel from the preset recall channels in descending order of matching degree. For example, a preset number of preset recall channels can be selected as target recall channels in descending order of matching degree. Alternatively, the electronic device can select preset recall channels with a matching degree higher than a preset matching degree as target recall channels.

[0056] Optionally, the aforementioned multiple target recall channels include a primary recall channel and secondary recall channels. The electronic device can use the preset recall channel with the highest matching degree as the primary recall channel. Furthermore, the electronic device can determine secondary recall channels from the remaining preset recall channels other than the preset recall channel with the highest matching degree. For example, all remaining preset recall channels can be used as secondary recall channels; or, for example, some remaining preset recall channels can be used as secondary recall channels, such as using the preset recall channel with the highest matching degree among the remaining preset recall channels as the secondary recall channel.

[0057] For example, if the query request above indicates a search for table content, then the table-level recall channel with the highest match to the query request can be used as the primary recall channel. Alternatively, if the query request above indicates a search for parameter values, then the column-level recall channel with the highest match to the query request can be used as the primary recall channel.

[0058] Optionally, the recall ratio of the primary recall channel is higher than that of the secondary recall channel to ensure the accuracy of the recall results.

[0059] Optionally, the model size of the primary recall channel is larger than that of the secondary recall channel to ensure the accuracy and coverage of the primary recall channel.

[0060] In addition, the resource allocation priority of the primary recall channel is higher than that of the secondary recall channel to ensure the performance of the primary recall channel.

[0061] The process of determining the target recall channel has been introduced above. The process of using the target recall channel for recall will be introduced below.

[0062] S203. The electronic device searches the data to be searched in parallel based on multiple target recall channels to obtain the candidate search results corresponding to each target recall channel.

[0063] In this embodiment, the electronic device can activate a multi-channel recall system to enable parallel recall using multiple target recall channels. This means that based on the index and recall strategy of the multi-target recall channels, the data to be detected is processed, and the retrieval results (i.e., recall results) corresponding to each target recall channel are obtained. The electronic device can use the retrieval results corresponding to each target recall channel as candidate detection results, thus achieving multi-channel recall.

[0064] S204. The electronic device determines the target search result from the candidate search results.

[0065] In this embodiment, the electronic device can determine the target retrieval results with a high degree of matching with the query request from the candidate detection results corresponding to different target recall channels, ensuring the accuracy of the target retrieval results and guaranteeing user experience. Furthermore, the electronic device utilizes multiple recall channels matching the query request, including row-level, column-level, table-level, and text-level recall channels, to perform recall. This allows for recall from the data to be detected using different indexes and recall strategies, achieving recall across different dimensions. It can accommodate both structured tables and text paragraphs, ensuring comprehensiveness and coverage of the recall, and achieving high recall rate and high relevance retrieval for complex documents.

[0066] In addition, even if the data to be detected includes tables (such as nested tables or complex tables containing images), electronic devices can still retrieve data through the table retrieval channel, avoiding low accuracy in table retrieval due to the inability to accurately extract the table data to be detected.

[0067] In some embodiments, the electronic device can evaluate the confidence level of candidate search results based on a fusion scoring mechanism. The electronic device fuses the candidate search results based on multiple dimensions to obtain a fusion score for each candidate search result; wherein, the multiple dimensions include at least two of the following: semantic relevance between the candidate search result and the query request, attribute consistency between the candidate search result and the query request, and data coverage. Then, the electronic device selects M candidate search results from the candidate search results in descending order of their fusion scores to obtain the target search result. In other words, the fusion score of these M candidate search results is higher than the fusion score of the candidate search results other than the M candidate search results.

[0068] Optionally, the process for determining the fusion score may include: For each candidate search result, the electronic device uses Calculate the fusion score corresponding to the candidate retrieval result.

[0069] Here, Score represents the fusion score corresponding to the candidate retrieval result. w1 represents the first weight, which is the weight corresponding to semantic relevance.

[0070] w2 represents the second weight, which is the weight corresponding to attribute consistency.

[0071] w3 represents the third weight, which is the weight corresponding to the data coverage.

[0072] This indicates the semantic relevance between the candidate search results and the query request (i.e., the similarity between the candidate search results and the query request).

[0073] This indicates the consistency (i.e., structural consistency) between the candidate search results and the query request in terms of attributes. For example, if the query request indicates a query for age, then if the candidate search results are related to age, the attribute consistency is high. If the candidate search results are not related to age, such as if the candidate search results indicate personality, then the attribute consistency is low.

[0074] This refers to data coverage (or context coverage). For example, if the data to be tested consists of five documents, and four documents are retrieved, the data coverage is 80%. Or, if the data to be tested consists of one document divided into 10 chapters, and two chapters are retrieved, the data coverage is 20%. Additionally, data coverage can also be determined based on the sufficiency of candidate search results.

[0075] In some embodiments, the electronic device may perform sufficiency checks on all candidate search results based on the request structure information and in accordance with sufficiency verification rules.

[0076] The sufficiency test includes one or more of the following: core entity hit detection, valid evidence source detection corresponding to the target attribute, and constraint condition satisfaction detection; valid evidence source refers to the location in the document.

[0077] Core entity hit detection is used to check whether all execution results hit all core entities in the query request.

[0078] The detection of valid evidence sources corresponding to target attributes is used to detect whether each target attribute corresponds to at least one valid piece of evidence.

[0079] The constraint satisfaction test is used to check whether the candidate search results meet the constraints. Simply put, it checks whether the range integrity requirements are met, such as whether the maximum value, minimum value, and all candidate values ​​required for interval calculation are available.

[0080] In addition, the aforementioned adequacy test may also include testing whether evidence structures spanning multiple pages or paragraphs are identified and processed.

[0081] After conducting a sufficiency search, electronic devices can use models or algorithms to calculate the results based on the core entity hit detection, the detection of valid evidence sources corresponding to the target attributes, and the detection of constraint satisfaction.

[0082] In the embodiments of this application, such as Figure 5 As shown, the semantic relevance, attribute consistency, and data coverage of the candidate retrieval results to the query request are used as inputs for the fusion scoring, and the fusion score is used as the output of the fusion scoring. In other words, the score of the candidate retrieval results is jointly determined by semantic relevance, attribute consistency, and data coverage, realizing a multi-granularity fusion strategy. It can also evaluate the semantic relevance, attribute consistency, and contextual integrity of the recall results, thus achieving accurate evaluation of the recall results.

[0083] Optionally, on the one hand, the aforementioned weights (such as the first weight, the second weight, and the third weight) are preset. On the other hand, the aforementioned weights are adaptively adjusted based on the query request and the candidate search results (i.e., candidate fragments). These weights are adapted to the query request, target attributes, and candidate results; that is, these weights can be adaptively adjusted according to the query request, target attributes, and the structural integrity of the candidate search results.

[0084] For example, the weighting coefficients w1, w2, and w3 of the fusion scoring can be calculated using a weighting function. Sure.

[0085] Among them, f i (q, D) represents the importance function of the input query request q and the candidate fragment D on the i-th feature, which can be obtained through learning or rule-based computation. This feature includes the semantic relevance, attribute consistency, and data coverage mentioned above.

[0086] In the embodiments of this application, such as Figure 6 As shown, the weight adaptation mechanism may include: S601, The electronic device extracts the vector q of the query request and the vector D of the candidate fragment.

[0087] S602. The electronic device calculates wi based on the weight calculation function.

[0088] S603, the electronic device outputs dynamic weights w1, w2 and w3.

[0089] In this embodiment of the application, the electronic device can query requests to automatically adjust the fusion weights of different recall channels in order to dynamically evaluate the accuracy of the recall results.

[0090] S205. The electronic device outputs the target retrieval results.

[0091] In this embodiment of the application, after determining the target search result, the electronic device can output the target search result, such as displaying the target search result or sending the target search result to the terminal.

[0092] In some embodiments, after obtaining candidate search results, the electronic device can perform sufficiency and consistency checks on the candidate search results. If both the sufficiency and consistency checks indicate that the results are normal, the electronic device can determine the target search result from the candidate search results.

[0093] In cases where the adequacy test results or the consistency test results are abnormal, the electronic device will be subject to a second recall.

[0094] Optionally, as mentioned above, sufficiency testing includes one or more of the following: core entity hit detection, valid evidence source detection corresponding to the target attribute, and constraint satisfaction detection.

[0095] If not all core entities are hit, the core entity hit detection can be considered a failure. If all core entities are hit, the core entity hit detection can be considered a success.

[0096] If each target attribute corresponds to at least one valid piece of evidence, the detection of a valid evidence source can be considered successful. If no valid evidence corresponds to a target attribute, the detection of a valid evidence source can be considered unsuccessful. For example, {"coverage_status": "insufficient", "missing_attributes": ["target attribute A"], "reason": "No complete candidate set was found that can be used to calculate the extrema." Therefore, it can be determined that the detection of valid evidence sources failed.

[0097] If the constraints are not met, the constraint satisfaction check can be considered a failure. If the constraints are met, the constraint satisfaction check can be considered a success.

[0098] Optionally, the above adequacy test results can indicate whether there is an abnormality. If any adequacy test fails, the electronic device can determine that the adequacy test results indicate an abnormality. If all adequacy tests are successful, the electronic device can determine that the adequacy test results indicate normality.

[0099] Taking sufficiency testing, which includes core entity hit detection, detection of valid evidence sources corresponding to target attributes, and constraint satisfaction detection, as an example, if core entity hit detection fails, detection of valid evidence sources corresponding to target attributes fails, or constraint satisfaction detection fails, the electronic device can determine that the sufficiency testing result indicates an anomaly.

[0100] If the core entity hit detection is successful, the valid evidence source corresponding to the target attribute is successfully detected, and the constraint conditions are met, the electronic device can determine that the sufficiency test result indicates normal operation.

[0101] The above describes the process of sufficiency testing for candidate search results. The following section will introduce the relevant content of consistency testing.

[0102] The aforementioned consistency checks include one or more of the following: evidence consistency checks, structural consistency checks, and semantic consistency checks.

[0103] Evidence consistency detection refers to the detection of valid evidence sources corresponding to candidate search results. That is, it verifies whether each candidate search result can be traced back to at least one valid evidence source. Furthermore, if all candidate search results can be traced back to valid evidence, the electronic device can determine that the evidence consistency detection has succeeded. If there are candidate search results that cannot be traced back to valid evidence, the electronic device can determine that the evidence consistency detection has failed.

[0104] Structural consistency testing refers to the structural consistency between candidate search results and valid evidence; that is, verifying whether all candidate search results conform to the structural constraints of the evidence, for example: Do the table fields correspond to each other? Has the time sequence been incorrectly reversed? Do comparative questions distinguish between different states or points in time?

[0105] Furthermore, if all candidate search results conform to the structural constraints of the evidence, the electronic device can determine that the structural consistency detection was successful. If any candidate search result does not conform to the structural constraints of the evidence, the electronic device can determine that the structural consistency detection failed.

[0106] Semantic consistency detection involves detecting anomalous expressions in candidate search results, such as verifying the presence of conceptual confusion, implicit unit variations, or generalized expressions among all candidate search results. If all candidate search results are expressed correctly, the electronic device can determine that the semantic consistency detection was successful. If any candidate search results exhibit anomalous expressions, the electronic device can determine that the semantic consistency detection failed.

[0107] In this embodiment of the application, the electronic device may adopt a combined verification mechanism, using a rule verification module to perform deterministic rule judgments on numerical calculations, unit consistency, comparison conditions, etc., and using a model-assisted audit module to audit the "correspondence between conclusion fields and evidence fields" based on a language model, but without generating new conclusions.

[0108] Similar to the previous examples, the consistency check results can indicate whether an anomaly has occurred. If any consistency check fails, the electronic device can determine that the consistency check results indicate an anomaly. If all consistency checks succeed, the electronic device can determine that the consistency check results indicate normal operation.

[0109] Taking consistency detection, which includes evidence consistency detection, structural consistency detection, and semantic consistency detection, as an example, if evidence consistency detection fails, structural consistency detection fails, or semantic consistency detection fails, the electronic device can determine that the consistency detection result indicates an anomaly.

[0110] If the evidence consistency test, structural consistency test, and semantic consistency test are all successful, the electronic device can determine that the consistency test results indicate normal operation.

[0111] Alternatively, the above consistency check results can also be in a structured form, for example: {"consistency_status": "conflict","conflicted_fields": [{"field": "target attribute B","issue": "multiple pieces of evidence exist but aggregation is not complete"}]}.

[0112] In some embodiments, the electronic device can not only output the search results for each target, but also the search information corresponding to each target search result. The search information includes multiple parameters such as the fusion score, weight distribution (e.g., the first weight, second weight, and third weight mentioned above), and data coverage for each target search result. The weight distribution can also be referred to as the channel contribution ratio. Based on this, interpretable information related to the target search results is provided, enabling relevant personnel to clearly understand the relevant information of the target search results.

[0113] In some embodiments, before outputting the target search results, that is, before determining that the electronic device can perform confidence testing and assess contextual integrity of the target search results to evaluate the accuracy of the target search results, the electronic device can separately determine whether each target search result is greater than or equal to a preset score threshold, and whether the data coverage is greater than or equal to a preset coverage rate.

[0114] If the fusion score corresponding to the target retrieval result is greater than or equal to the preset score threshold, and the data coverage is greater than or equal to the preset coverage, output the target retrieval result that is greater than or equal to the preset score threshold. If the fusion scores corresponding to the target search results are all less than the preset score threshold or the data coverage is less than the preset coverage, a secondary recall action is performed to supplement the missing information.

[0115] In some embodiments, the aforementioned secondary recall may be based on adjusting search suggestions, remapping the vector of the query request, or redetermining the target recall channel. Alternatively, the aforementioned secondary recall may be based on reflection results, adjusting the retrieval process, etc. For example, the reflection results may include improvement strategies and remedial suggestions. For instance, improvement strategies may include expanding the search scope; triggering table structuring processing; adjusting the reasoning path; rejecting answers and outputting explanations.

[0116] Optionally, the above table structuring process may include: Based on page level, the electronic device identifies candidate table regions in the document (or document to be processed). Cross-page continuity detection is performed on candidate table regions on adjacent pages in the document to be processed to determine candidate table regions corresponding to the same table. A reconstruction operation is performed on the candidate table regions corresponding to the same table; wherein, the reconstruction operation includes header inheritance and / or merging cell restoration and hierarchical mapping; Based on the content of the candidate areas of the reconstructed table, structured table information is generated.

[0117] In this embodiment of the application, the electronic device can perform layout analysis on the document to be processed to determine the layout elements on each page of the document to be processed. The layout elements may include a table area (i.e., a table candidate area) so that the table in the document to be processed can be extracted using the layout elements, thereby enabling the table to be processed and the task to be executed.

[0118] For example, electronic devices can use LayoutParser, YOLO detection, or other models to perform layout analysis on the document to be processed. Additionally, layout elements can also include table areas, headings, headers and footers, images, etc.

[0119] Optionally, the electronic device can also determine metadata about candidate table areas on each page. For example, the metadata may include at least one of the following: Page number; The location of the candidate area in the table; Row and column quantity characteristics; Does it include a table header? Textual and geometrical distribution features of table content.

[0120] Among them, geometric distribution features (also known as geometric features, table geometric features, etc.) represent at least one of the following features: column border alignment of the candidate table area, table width similarity, and page spacing. Semantic features (also known as text features) represent at least one of the following features: table header keywords, units, and column headings.

[0121] Optionally, after determining the metadata of the candidate table region, the electronic device can construct a table structure state description of the candidate table region based on the metadata. This table structure state description can be used for cross-page continuity detection. Of course, the electronic device can also directly use the metadata for cross-page continuity detection.

[0122] For example, the table structure status description may include at least one of the following information: Does the current page contain field definition information? Whether the first row of the current page is a data row, that is, whether the candidate table area in the current page includes the table header; Does the row content maintain semantic continuity with the previous page? Alignment of the candidate area in the table, such as column alignment or column alignment.

[0123] The text distribution characteristics of the candidate regions in the table (i.e., the text features mentioned above).

[0124] In some embodiments, the document to be processed may be a bidding document. Accordingly, the tables in the document to be processed may include the scoring table and qualification table in the bidding document. Alternatively, the document to be processed may be an engineering technical document. Accordingly, the tables in the document to be processed may include the parameter and indicator table in the engineering technical document.

[0125] Alternatively, the document to be processed can be a production document from the oil and gas or energy industry. Correspondingly, the tables in the document to be processed can include production data tables from production documents in the oil and gas or energy industry.

[0126] Alternatively, the document to be processed can be a manufacturing, testing, or quality report. Correspondingly, the tables in the document to be processed can include complex tables from manufacturing, testing, or quality reports. Of course, the document to be processed can also be other types of documents, and this application does not impose any limitations on them.

[0127] It should be noted that a table can consist of a header and data rows. The header is the table's "title bar," which defines the meaning, category, or attribute of each column of data. Data rows are the entries in the table that carry the actual data. For example, in Table 1, "Name" and "Class" are the header, while "Zhang San, Class 1" are the data rows.

[0128] Table 1

[0129] In this embodiment, the electronic device can determine the cross-page continuity of candidate table areas on adjacent pages according to page number order to identify whether the candidate table areas on adjacent pages belong to the same table, thereby filtering out the candidate table areas corresponding to each independent table in the document to be processed. Furthermore, candidate table areas corresponding to the same table can be uniformly identified as a logical table object. This logical table object represents table instances belonging to the same business semantic entity before and after a page crossover, and is used for subsequent table structure reconstruction, field mapping, and structured output.

[0130] Alternatively, the above-mentioned cross-page continuity detection can be determined through the following two implementation methods.

[0131] In one implementation, the electronic device can perform cross-page detection by setting a preset continuity condition. The electronic device determines whether the candidate table regions on adjacent pages meet the preset continuity condition.

[0132] If the table candidate areas on adjacent pages meet a preset continuity condition, it indicates that the table candidate area on the next page is a continuation of the table candidate area on the previous page. In other words, the table candidate areas on adjacent pages belong to different portions of the same table. Therefore, the electronic device can determine that the table candidate areas on adjacent pages correspond to the same table.

[0133] If the table candidate areas on adjacent pages do not meet the preset continuity condition, it indicates that the table candidate area on the next page of the two adjacent pages is not a continuation of the table candidate area on the previous page. In other words, the table candidate areas on adjacent pages do not belong to different table parts of the same table, but belong to two independent tables. Therefore, the electronic device determines that the table candidate areas on adjacent pages do not correspond to the same table.

[0134] The preset continuity conditions include several of the following conditions: The table candidate areas on adjacent pages correspond to the same chapter or the same structural path; The current page containing the candidate table area lacks field definitions, and the data row density is greater than the preset density value; The table information in the candidate table areas of adjacent pages is the same; the table information includes the number of rows, the number of columns, and the column width distribution. The similarity of the table content in the candidate table areas of adjacent pages is greater than the preset similarity.

[0135] The above structure path represents the path of the header field and / or data cell in the candidate area of ​​the table.

[0136] In addition, it is understandable that the above-mentioned cross-page continuity detection process does not depend on the presence of a table header.

[0137] In another implementation, electronic devices can perform cross-page detection using a cross-page continuity score (CCS). The electronic device calculates the cross-page continuity score for candidate table regions on adjacent pages based on multiple feature dimensions. These multiple feature dimensions include several of the following: table geometric feature similarity, table text similarity (or semantic feature similarity), and unit consistency score.

[0138] The electronic device then determines whether the cross-page continuity score of the table candidate area on adjacent pages is greater than or equal to a preset score.

[0139] If the cross-page continuity score is greater than or equal to a preset score, it indicates that the table candidate area on the next page is a continuation of the table candidate area on the previous page. In other words, the table candidate areas on adjacent pages belong to different portions of the same table. Therefore, the electronic device determines that the table candidate areas on adjacent pages correspond to the same table. If the cross-page continuity score is less than a preset score, it indicates that the table candidate area on the next page is not a continuation of the table candidate area on the previous page. In other words, the table candidate areas on adjacent pages do not belong to different parts of the same table, but to two independent tables. Therefore, the electronic device determines that the table candidate areas on adjacent pages do not correspond to the same table.

[0140] Optionally, the process of determining the cross-page continuity score of the table candidate areas of the adjacent pages may include: weighted summation of multiple feature dimensions to obtain the cross-page continuity score.

[0141] Taking multiple feature dimensions, including table geometric feature similarity, table text similarity, and unit consistency score, as an example, electronic devices can accordingly adopt... Achieve weighted summation.

[0142] Here, α, β, and γ are weighting coefficients, satisfying α + β + γ = 1. Furthermore, α, β, and γ can be preset.

[0143] This indicates the similarity of the geometric features of the tables. This indicates the similarity of the table text (or alternatively, the similarity of the table header text). This indicates the consistency score of the units.

[0144] Optionally, table geometric feature similarity represents the similarity between at least one of the following features: column border alignment, table width similarity, and page spacing of candidate table regions on adjacent pages. Simply put, table geometric feature similarity represents the geometric alignment of table boundaries on adjacent pages.

[0145] Semantic feature similarity represents the similarity between table header keyword overlap, unit matching, and at least one feature in column headings. Simply put, table geometric feature similarity represents the semantic overlap between tables on adjacent pages.

[0146] Here, header keyword overlap indicates the similarity between keywords in the header sections of candidate tables on adjacent pages. Unit matching indicates the similarity between units in candidate data rows on adjacent pages.

[0147] In this embodiment, for each dimension feature, the electronic device can use a relevant algorithm or model to calculate that dimension feature. Taking the calculation of table geometric feature similarity, where table geometric feature similarity represents the similarity between the column border alignment, table width similarity, and page spacing of the table candidate areas of adjacent pages, as an example, the electronic device can input the column border alignment, table width similarity, and page spacing of the table candidate areas of adjacent pages into the network model, so that the network model can calculate the similarity and obtain the table geometric feature similarity.

[0148] Alternatively, electronic devices can convert the column border alignment, table width similarity, and page spacing of candidate table regions on adjacent pages into geometric feature vectors, which can then be used to obtain the geometric feature similarity of tables by combining the candidate table regions on adjacent pages with cosine similarity.

[0149] In this embodiment of the application, for all candidate table regions corresponding to different tables in the document to be processed, the electronic device performs reconstruction operations such as header integration, cell merging restoration and hierarchical mapping on all candidate table regions corresponding to the same table, so as to restore the semantics of the same table and generate a table semantic tree.

[0150] Optionally, the implementation process of the above-mentioned header inheritance may include: The electronic device copies the table headers from the candidate table area on the previous page to the candidate table area on the current page, and performs semantic alignment and unit verification. Unit verification checks the consistency of the units.

[0151] In this embodiment, the electronic device extracts all header fields from the candidate table area of ​​the previous page. Then, the electronic device calculates the similarity between the header fields and the first row fields at corresponding positions in the candidate table area of ​​the current page, obtaining the similarity score corresponding to the first row field. In this way, the electronic device can determine the similarity score corresponding to each first row field in the candidate table area of ​​the current page.

[0152] Then, the electronic device determines whether the similarity of the first row of fields is greater than or equal to the preset similarity threshold.

[0153] If the similarity of the first row field is greater than or equal to the preset similarity threshold, the electronic device copies the header field from the previous page corresponding to the first row field to the current page, which is used as the header field corresponding to the first row field.

[0154] If the similarity of the first row field is less than a preset similarity threshold, it indicates that the header field in the previous page corresponding to that first row field does not match it. Therefore, the header field in the previous page cannot be used as the header field corresponding to the first row field. Instead, the header of the first row field is completed according to the dictionary to achieve semantic alignment. Alternatively, a template can be used for header completion.

[0155] Specifically, the electronic device can find the field with the highest semantic similarity to the first row field from a preset dictionary and use it as the header field corresponding to the first row field.

[0156] After the current page includes the complete table header, the electronic device can continue to verify whether the units of the table header fields on the current page are consistent with the units of the data in the corresponding data cells, in order to achieve unit consistency verification.

[0157] Under consistent conditions, electronic devices can determine that the header of the current page is correct and generate a complete header semantic mapping, that is, associate the header fields with the data in their corresponding data cells.

[0158] In the event of inconsistencies, the electronic device can correct the corresponding header field of the data cell based on the semantics of the data within the data cell, in conjunction with a dictionary. The electronic device can then generate a complete header semantic mapping based on the corrected header.

[0159] Based on this, electronic devices automatically restore the header inheritance, enabling header reproduction and avoiding the inability to correctly link content between pages.

[0160] In some embodiments, the process of restoring merged cells and mapping hierarchical levels includes: Electronic devices identify merged cells in a candidate area of ​​a table based on the geometric coverage relationship of cells; wherein merged cells include row merges and / or column merges; Restore the parent-child relationship between fields in merged cells in electronic devices; The electronic device maps the header of the candidate area of ​​the table to a multi-level header based on the parent-child relationship.

[0161] In this embodiment, for tables with merged cells, the electronic device identifies row and column merging based on cell geometric coverage relationships. Then, the electronic device can construct a header hierarchy, restoring the parent-child relationships between the fields corresponding to the merged cells, thereby mapping the headers in the candidate area of ​​the table to multi-level headers based on these parent-child relationships. The fields corresponding to the merged cells include the merged cell itself and its subordinate fields.

[0162] In addition, electronic devices can map multi-level headers to structured field paths, such as converting two-dimensional headers into multi-level field description structures.

[0163] For example, the "Family" cell is a row merge, and the fields corresponding to the merged cell include the headers "Father" and "Mother". Therefore, in order to restore the complete cell structure, the electronic device restores the parent-child relationship between the fields corresponding to the merged cells, that is, directly establishes a parent-child relationship between the merged cell and its subordinate fields.

[0164] Optionally, before identifying merged cells in the candidate area of ​​a table based on the geometric coverage relationship of cells, the coordinates of cells spanning multiple pages can be converted into a unified coordinate system, and the coordinates of the candidate area of ​​the table within the original page can be normalized and mapped to the global coordinate system spanning multiple pages to ensure the correct restoration of the parent-child relationship, realize cell relocation and structural reconstruction, and thus ensure that the cell hierarchy and merged state remain consistent during the reconstruction process.

[0165] In some embodiments, the reconstruction operation described above also includes table semantic alignment. This table semantic alignment operation may be performed after the above-described header inheritance, merged cell restoration, and hierarchy mapping.

[0166] For example, the process of semantic alignment of the above tables may include: Electronic devices align multi-level headers with each data cell in the candidate area of ​​the table, i.e., perform semantic field alignment of the table, to ensure that each data cell can be mapped to a unique field path, and that cross-page continuation can inherit the field definitions of the previous page and prevent field semantic drift due to layout changes.

[0167] Understandably, for tables in the document to be processed that do not span multiple pages, the reconstruction operation performed by the electronic device on the corresponding table candidate area may not include table inheritance. Furthermore, for tables in the document to be processed that do not contain merged cells, the reconstruction operation may not include restoring merged cells; instead, hierarchical mapping can be performed directly.

[0168] In some embodiments, the format of the structured table information described above may be JSON.

[0169] In this embodiment of the application, after completing cross-page merging and complex structure parsing, the electronic device can use the content in the candidate area of ​​the table corresponding to the reconstructed same table to generate the structured table information corresponding to the table, and can output unified structured table information to achieve accurate extraction of table content.

[0170] Optionally, the structured table information mentioned above may include several of the following: Unique identifier for the table; Table header field paths; Structured key-value mapping of data rows; Table source information (such as page range, chapter path).

[0171] The header field path represents the path of the header field, which reflects the relationship of the header field. For example, the header field path of the header field "father" is family-father.

[0172] The structured key-value mapping of the data rows described above can reflect the data (i.e., key values) in the data row cells and their corresponding header fields.

[0173] Therefore, clearly and straightforwardly displaying the structure of a table can improve the speed at which the required information can be obtained from structured table information.

[0174] For example, the structured table information above is as follows: { "table_id": "T2025_001", "pages": [5, 6], "headers": ["serial number", "test item", "unit", "value"], "rows": [ {"Serial Number": "1", "Test Item": "Density", "Unit": "g / cm3", "Value": "1.23"}, {"Serial Number": "2", "Test Item": "Viscosity", "Unit": "mPa·s", "Value": "45.6"} ], "ccs_score": 0.88, "confidence_score": 0.94, "structure_hash": "8f9a1b3c..." }

[0175] Here, `structure_hash` represents the hash value corresponding to the structured table information. For example, after determining the structured table information, the electronic device can calculate the hash value based on the content of that structured table information (such as data in data cells, header fields, etc.) to obtain the actual hash value corresponding to the table. Then, the electronic device can compare the actual hash value corresponding to the table with the standard hash value corresponding to that table to achieve hash verification.

[0176] The standard hash value corresponding to the table represents the correct hash value for that table.

[0177] Successful verification indicates correct table extraction, allowing the electronic device to directly output the corresponding structured table information. Successful verification indicates table extraction failed, requiring the electronic device to perform a second extraction. Additionally, the `confidence_score` can represent the confidence level, which can be understood as a score indicating continuity across pages.

[0178] Figure 7 This is a schematic diagram of the structure of an electronic device provided in some embodiments of this application. Figure 7 The dashed line in the text indicates that the unit or module is optional. Figure 7 The electronic device 700 can be used to implement the methods described in the above method embodiments. The electronic device 700 can be a chip, a terminal device, or a server.

[0179] Electronic device 700 may include one or more processors 710. The processor 710 can support the electronic device 700 in implementing the methods described in the preceding method embodiments. The processor 710 can be a general-purpose processor or a special-purpose processor. For example, the processor can be a Central Processing Unit (CPU). Alternatively, the processor can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0180] The electronic device 700 may also include one or more memories 720. Computer programs are stored on the memories 720. The memories 720 may be independent of the processor 710 or integrated into the processor 710.

[0181] The electronic device 700 may also include a transceiver 730. The processor 710 can communicate with other devices or chips via the transceiver 730. For example, the processor 710 can send and receive data with other devices or chips via the transceiver 730.

[0182] The computer program in memory 720 can be executed by processor 710, causing processor 710 to perform the retrieval method as described above.

[0183] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0184] Therefore, embodiments of this application provide a computer-readable storage medium storing a computer program thereon, which is loaded by a processor to execute the steps described in the above-described method embodiments of this application. For example, the computer program, when loaded by a processor, can execute the retrieval method as described above.

[0185] For details on the implementation of each of the above operations / steps, please refer to the previous examples, which will not be repeated here.

[0186] The computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0187] Since the computer program stored in the computer-readable storage medium can execute the steps in any of the above method embodiments provided in the embodiments of this application, the beneficial effects that the methods described in any of the above method embodiments can achieve can be realized, as detailed in the preceding embodiments, and will not be repeated here.

[0188] This application also provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the electronic device to perform the methods provided in the various optional implementations of the above embodiments.

[0189] The foregoing has provided a detailed description of a retrieval method and electronic device provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A retrieval method, characterized in that, include: Based on the query request, multiple target recall channels are determined from the preset recall channels; wherein, the preset recall channels include row-level recall channels, column-level recall channels, table-level recall channels and text-level recall channels; Based on the multiple target recall channels, the data to be retrieved is retrieved in parallel to obtain the candidate retrieval results corresponding to each target recall channel; Output the target search results; wherein the target search results are determined from the candidate search results.

2. The method according to claim 1, characterized in that, The step of determining multiple target recall channels from preset recall channels includes: Calculate the matching degree between the query request and each preset recall channel; Multiple target recall channels are determined from the preset recall channels based on the matching degree.

3. The method according to claim 1 or 2, characterized in that, Before outputting the target retrieval result, the method includes: The candidate search results are fused based on multiple dimensions to obtain a fusion score for each candidate search result; wherein, the multiple dimensions include at least two of the following: semantic relevance between the candidate search results and the query request, attribute consistency between the candidate search results and the query request, and data coverage. According to the fusion score in descending order, M candidate search results are selected from the candidate search results to obtain the target search result.

4. The method according to claim 3, characterized in that, The process of fusing the candidate search results based on multiple dimensions to obtain a fusion score for each candidate search result includes: For each of the candidate search results, the following is adopted: Calculate the fusion score corresponding to the candidate search result; wherein, the Score represents the fusion score corresponding to the candidate search result; This indicates the semantic relevance between the candidate search results and the query request. This indicates that the candidate search results are consistent with the attributes of the query request. The data coverage is represented by w1, w2, and w3, which represent the first, second, and third weights, respectively.

5. The method according to claim 4, characterized in that, The output of the target retrieval result includes: Output the target retrieval result and its corresponding retrieval information; wherein, the retrieval information includes multiple of the following: the fusion score, weight distribution, and data coverage corresponding to the target retrieval result.

6. The method according to claim 4 or 5, characterized in that, The output target retrieval results include: If the fusion score corresponding to the target retrieval result is greater than or equal to a preset score threshold, and the data coverage is greater than or equal to a preset coverage, the target retrieval result that is greater than or equal to the preset score threshold will be output. The method further includes: If the fusion scores corresponding to the target retrieval results are all less than a preset score threshold or the data coverage is less than a preset coverage, a secondary recall action is performed.

7. The method according to claim 1, characterized in that, Before outputting the target retrieval result, the method further includes: A sufficiency test is performed on the candidate search results to obtain a sufficiency test result, and a consistency test is performed on the candidate search results to obtain a consistency test result; If the sufficiency test result indicates that the result is normal, and the consistency test result indicates that the result is normal, the target search result is determined from the candidate search results.

8. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program or instructions, which, when executed by the processor, cause the processor to perform the method as described in any one of claims 1 to 8.

9. A computer-readable storage medium, characterized in that, It stores a computer program or instructions thereon, which, when executed by a processor, implement the method as described in any one of claims 1 to 8.

10. A computer program product, characterized in that, When the computer program product is run on an electronic device, it causes the electronic device to perform the method as described in any one of claims 1 to 8.

Images