Archive text automatic abstracting and knowledge question-answering system combined with large model

Abstract

The invention discloses an archive text automatic abstracting and knowledge question-answering system combined with a large model, and relates to the technical field of text automatic abstracting and question-answering. Comprising a text scanning module for performing text scanning on a target archive text, and extracting text structured data through an entity recognition model to obtain text information items; and the information screening module is used for carrying out information screening on the basis of the text information items, screening the information into judgment data and non-judgment data, taking the judgment information as input data, and generating matching information corresponding to the input data through a large language model. According to the method, the file noise interference is resisted through a triple mechanism, the cross-modal alignment technology performs joint analysis on the OCR text, the seal image and the table data, so that the fuzzy character recognition accuracy is improved, the mapping relation between the key information item and the alignment information item is established, and the spelling dislocation problem of the digital file is solved.

Description

Technical Field

[0001] This invention relates to the field of automatic text summarization and question answering technology, specifically to an automatic text summarization and knowledge question answering system for archives that combines a large model. Background Technology

[0002] Automatic text summarization and knowledge-based question answering refer to two types of text-based automated processing tasks, typically used in information systems such as archives, documents, and databases. Automatic text summarization quickly extracts core information, key points, or important facts to form a concise summary, with common forms including paragraph-level summaries, sentence-level summaries, and keyword summaries. It is used for archive organization, rapid retrieval, information summarization, and assisted reading. Knowledge-based question answering, on the other hand, answers specific questions posed by users based on information within the archival text, without requiring users to read the entire document themselves. Automatic text summarization and knowledge-based question answering are often interconnected.

[0003] A text automatic summarization method, apparatus, and electronic device, with patent publication number CN109101489A, calculates the score of each sentence by utilizing the relevance and similarity between sentences, and comprehensively considers the word collinearity rate and semantic relevance of sentences, thereby improving the accuracy of sentence scoring. The proposed summarization method is convenient to calculate, has strong universality, and is applicable to text content summarization extraction in multiple fields.

[0004] In the process of automatic text summarization and knowledge question answering, the above-mentioned and similar technical solutions suffer from various forms of text noise, such as colloquial dialect expressions, blurred characters in handwriting recognition, and spelling errors common in online texts. The presence of these noises seriously affects the accuracy of semantic parsing. When the system attempts to understand text containing a lot of noise, it often fails to correctly identify key information, leading to omissions or misjudgments. Traditional summarization and question answering rely on full-text semantic parsing, which is easily interfered with by text noise, resulting in omissions or misjudgments of key information. Summary of the Invention

[0005] The purpose of this invention is to provide an automatic text summarization and knowledge question answering system that combines large-scale models to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an automatic text summarization and knowledge question answering system for archives combining a large model, comprising:

[0007] Text scanning module: performs text scanning on the target file text, extracts structured text data through entity recognition model, and obtains text information items;

[0008] Information screening module: Based on text information items, information is screened into decision data and non-decision data. Decision information is used as input data. The large language model generates matching information corresponding to the input data to obtain the input matching set. The input matching set includes at least one input matching item corresponding to the input data.

[0009] Model creation module: Based on the input matching set, a text comparison model is created to obtain the judgment model. Non-judgment data is used as the input of the judgment model to obtain output feedback information. The output feedback information is compared with the text information items to obtain the comparison matching results as summary data to obtain summary information items. Then, the target archive text is automatically summarized and output, reducing the impact of text noise on summary acquisition.

[0010] Data output module: Based on the judgment model, it outputs question and answer data, including archive text data and metadata. Based on the text data and metadata, it generates question and answer data to obtain question and answer information items, which prevents the omission of key information and reduces information misjudgment.

[0011] Furthermore, the structured data includes timestamps, geographic locations, person entities, organization entities, and topic tags, and the methods for obtaining text information items include:

[0012] A preliminary scan of the archival text was performed to obtain keyword information, including geographical keywords, personal keywords, institutional keywords, and main tag keywords, thus obtaining key information items;

[0013] Simultaneously parse OCR text, seal images, and table data, perform cross-modal alignment, and obtain aligned information items corresponding to key information items;

[0014] The text information items are obtained by comparing and combining key information items and alignment information items.

[0015] Furthermore, the method for obtaining the text information item also includes:

[0016] A preliminary scan of the archival text is performed to obtain blurred and missing characters, thus identifying interfering information items.

[0017] A damaged text repair model is created. Based on the damaged text repair model, combined with contextual calendar calculation, blurred and missing text in the scanned document is repaired, and interfering information items are repaired to obtain the repaired information items.

[0018] Text information items are obtained by comparing and combining repair information items, key information items, and alignment information items.

[0019] Furthermore, the determination information includes person entities, organization entities, and topic tags, and the method for obtaining the input matching set includes:

[0020] Based on the judgment information, a three-dimensional mapping of entity relationships is created. Different processing logics are matched based on different input entities to obtain entity mapping items.

[0021] Implicit relationship mining is performed based on entity mapping terms. Reasoning is carried out through spatiotemporal overlap relationships to obtain spatial intersection information and temporal causal information, output extended information, and then obtain the input matching set.

[0022] Furthermore, the method for obtaining the determination model includes:

[0023] The information model architecture diagram is created based on the input matching set, including spatiotemporal grid coding, dynamic graph construction, federated verification pipeline, multi-granularity output, and incremental optimization.

[0024] Based on spatiotemporal grid coding, discrete spatiotemporal data is transformed into computable coordinates. Based on dynamic graph construction, a spatiotemporal sensitive network of entity relationships is established. Based on federated verification pipeline, the credibility of reverse data is ensured. Based on multi-granularity output, reverse data of different precisions is generated on demand. Finally, based on incremental optimization, the model is self-updated to obtain the judgment model.

[0025] Furthermore, the non-determination data includes timestamps and geographical locations, and the method for obtaining the output feedback information includes:

[0026] Based on the decision model, using timestamps and geographic locations as model inputs, the output information of the decision model is obtained, resulting in timestamp output datasets and geographic location output datasets;

[0027] Information is correlated based on the timestamp output dataset and the geographic location output dataset to obtain overlapping and non-overlapping information, resulting in a first information set and a second information set. The first information set is used as feedback data to obtain output feedback information.

[0028] Furthermore, the method for obtaining the summary information item includes:

[0029] By comparing the information output items with the text information items, the overlapping data between the judgment data in the text information items and the output feedback information is obtained, and the overlapping data is used as the summary information items.

[0030] Furthermore, the question-and-answer data generation includes annotation generation, and the methods for obtaining question-and-answer information items include:

[0031] Data annotation is performed on the archival text data and the target archival text to obtain question and answer annotation items. These annotation items are then used to generate question and answer data, which in turn yields question and answer information items.

[0032] Furthermore, the question-and-answer data generation also includes evidence generation, and the method for obtaining question-and-answer information items also includes:

[0033] Set a data ratio value, remove data from metadata based on the data ratio value, obtain the retained data, and get the retained information items;

[0034] By using the retained data items as evidence labels for question and answer, data association is performed on the question and answer labeled items to obtain the question and answer information items.

[0035] Compared with the prior art, the beneficial effects of the present invention are:

[0036] This system, which combines large-scale models with automatic text summarization and knowledge-based question answering for archives, employs a triple mechanism to combat archival noise interference. First, cross-modal alignment technology performs joint analysis of OCR text, seal images, and tabular data, improving the accuracy of fuzzy character recognition and establishing a mapping relationship between key information items and aligned information items to solve the problem of spelling misalignment in digitized archives. Second, dynamic text repair, through a dual-channel repair engine, can repair damage caused by insect infestation or ink diffusion, with the repaired information items complementing the original key information. Third, an interference information stripping mechanism uses high-precision defect location to label interference items, reducing the misjudgment rate of dialect expressions.

[0037] Simultaneously, a verifiable intelligent decision-making chain was constructed. Multi-dimensional relationship mining includes entity relationship three-dimensional mapping technology to resolve implicit power structures, a spatiotemporal overlapping reasoning mechanism to increase the discovery of implicit associations in archives, encrypted spatiotemporal grid IDs of the federated verification system to achieve cross-institutional evidence chain aggregation, and contradiction arbitration rules to automatically eliminate erroneous relationship edges. Finally, through dynamic optimization of output, a multi-granularity output engine generates macroscopic time slices on demand, and an incremental optimization mechanism triggers local model updates through new evidence, thereby improving the timeliness of institutional evolution reasoning in archives. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0039] Figure 2 This is a schematic diagram illustrating the structured data display of the present invention;

[0040] Figure 3 This is a schematic diagram of the text information item acquisition process of the present invention;

[0041] Figure 4 This is a schematic diagram of the feedback data acquisition process of the present invention;

[0042] Figure 5 This is a schematic diagram of the output feedback information connection in file A of the present invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0044] Due to the unique nature of archival texts, their noise forms are diverse and complex. For example, colloquial dialect expressions are common in historical archives, and blurred characters generated during the recognition of handwritten documents are unavoidable. Digital archives may also contain spelling errors common in online texts. The presence of this noise severely impacts the accuracy of semantic analysis, thus significantly negatively affecting the performance of automatic summarization and question-answering systems. Traditional automatic summarization and question-answering methods often rely on semantic analysis of the entire text, attempting to extract key information and answer user questions through a comprehensive understanding of the text content. However, when processing texts containing a large amount of noise, the system may fail to correctly identify key information due to noise interference, leading to omissions or misjudgments. For example, when processing texts containing dialects… When expressing text, the system may fail to accurately understand its meaning, thus overlooking important historical events or relationships between figures. Similarly, ambiguous characters may lead to errors in recognizing key numbers or dates, thereby affecting the understanding of the timeline of historical events. The technical solution provided in this application uses a triple mechanism to resist archival noise interference. Among them, cross-modal alignment technology performs joint analysis on OCR text, seal images, and table data, improving the accuracy of ambiguous character recognition and establishing a mapping relationship between key information items and aligned information items, solving the problem of spelling misalignment in digitized archives; dynamic repair of damaged text, through a dual-channel repair engine, can repair damage caused by insects and ink diffusion, and the repaired information items complement the original key information; the interference information stripping mechanism uses high-precision defect location to mark interference items, reducing the misjudgment rate of dialect expressions, such as... Figure 1 As shown, it includes a text scanning module, an information screening module, a model creation module, and a data output module.

[0045] Text scanning module: performs text scanning on the target archive text, extracts structured text data through entity recognition model, and obtains text information items.

[0046] It is important to note that, such as Figure 2As shown, the structured data includes timestamps, geographic locations, person entities, organization entities, and topic tags. The methods for obtaining text information items include: performing a preliminary scan of the archival text to obtain keyword information, including geographic keywords, person keywords, organization keywords, and subject tag keywords, to obtain key information items; simultaneously parsing OCR text, seal images, and table data, performing cross-modal alignment to obtain aligned information items corresponding to the key information items; and obtaining text information items based on the comparison and combination of key information items and aligned information items.

[0047] Specifically, during the initial scanning of the archival text, keywords such as timestamps, geographical locations, individuals, organizations, and thematic tags were extracted. These were then enhanced using OCR text extraction. When processing the seal images, intelligent seal recognition was employed, including seal localization, 3D correction, seal script recognition, and entity mapping. First, overlapping characters and ink patterns were separated using seal outline feature detection. Then, a perspective transformation algorithm was used to eliminate seal pressure and tilting deformation. Next, a deep residual network was used to analyze ancient seal scripts such as Nine-Fold Seal Script and Jade Chopstick Seal Script. Finally, the seal-like text was... The data is transformed into organizational entities. When parsing table data, the table headers are associated with the data through table semantic parsing. For example, when recognizing table headers such as "salary silver, taels", the "two hundred" data column below is automatically bound. Then, cross-cell merging is performed, and finally, cross-modal alignment is performed through a cross-modal alignment engine. First, keyword anchoring is required. Anchor point rules include time entities and geographic entities. Among them, time entities are preferentially selected from Gregorian calendar dates after calendar conversion. Geographic entities are obtained by mapping locations to modern coordinates, thereby obtaining alignment information items corresponding to key information items, and finally obtaining the text structured data in the target archive text.

[0048] It is important to note that, such as Figure 3 As shown, the method for obtaining text information items also includes: performing a preliminary scan of the archive text to obtain blurred and missing characters in the archive text, thus obtaining interference information items; creating a damaged text repair model, and based on the damaged text repair model, combining contextual calendar calculations to repair blurred and missing characters in the scanned document, repairing interference information items, thus obtaining repaired information items; and obtaining text information items based on the comparison and combination of repaired information items, key information items, and alignment information items.

[0049] Specifically, when acquiring blurred and missing fonts in archival text, high-precision defect localization is achieved by scanning the archival image using edge detection algorithms to identify blurred areas, such as ink diffusion and yellowing of paper, and missing areas, such as insect damage and tears. Interference information items are then labeled in a format such as: location-type-original text. In creating the damaged text repair model, a dual-channel repair engine is designed, including a visual repair channel and a semantic reasoning channel. The technical principle of the visual repair channel is to train a historical font library using generative adversarial networks, predicting and filling in missing strokes through residual pixels, and training data... The sample consists of official documents from the target era, covering official scripts such as regular script and official script. The repair logic is as follows: input 50% of the strokes of a damaged character, output 3-5 candidate characters, sorted by similarity. The semantic reasoning channel includes a calendar calculation module and an official title verification module. When the missing text contains time information, the calendar calculation module generates a reasonable value based on the context date range. The official title verification module verifies the entity logic of the institution based on the official title database of the target era, and then repairs the interfering information items to obtain the repaired information items. Based on the comparison and combination of the repaired information items, key information items, and alignment information items, the text information items are obtained.

[0050] Information screening module: Based on text information items, information is screened into decision data and non-decision data. Decision information is used as input data, and matching information corresponding to the input data is generated through a large language model to obtain the input matching set.

[0051] It is important to note that the input matching set includes at least one input matching item corresponding to the input data. The judgment information includes person entities, organization entities, and topic tags. The method for obtaining the input matching set includes: creating a three-dimensional mapping of entity relationships based on the judgment information; matching different processing logics based on different input entities to obtain entity mapping items; performing implicit relationship mining based on entity mapping items; reasoning through spatiotemporal overlap relationships to obtain spatial intersection information and temporal causal information; outputting extended information; and thus obtaining the input matching set.

[0052] Specifically, the input entities in creating the 3D mapping of entity relationships include person entities, organization entities, and topic tags. The corresponding processing logic is as follows: associating spatiotemporal coordinates with social relationship chains, analyzing organizational evolution and power structure, binding historical events and legal validity, and then mining implicit relationships. Spatiotemporal overlapping relationship reasoning includes spatial intersection and temporal causality. Spatial intersection refers to people or organizations appearing in the same location, while temporal causality refers to changes in organizations before and after an event, thereby outputting extended information.

[0053] Model creation module: Creates a text comparison model based on the input matching set to obtain the decision model. Uses non-decision data as input to the decision model to obtain output feedback information. Based on the output feedback information, compares it with the text information items to obtain the comparison matching results as summary data to obtain the summary information items.

[0054] It is important to note that, based on the summary information items, the target file text is automatically summarized, thereby reducing the impact of text noise on the summary acquisition.

[0055] It is important to note that the method for obtaining the decision model includes: creating an information model architecture diagram based on the input matching set, including spatiotemporal grid encoding, dynamic graph construction, federated verification pipeline, multi-granularity output, and incremental optimization; converting discrete spatiotemporal data into computable coordinates based on spatiotemporal grid encoding; establishing a spatiotemporal sensitive network of entity relationships based on dynamic graph construction; ensuring the credibility of reverse data based on the federated verification pipeline; generating reverse data of different precisions as needed based on multi-granularity output; and finally performing model self-updating based on incremental optimization to obtain the decision model.

[0056] Specifically, spatiotemporal grid coding constructs a four-dimensional coordinate index, including calendar fusion conversion, geographic hierarchy mapping, and spatiotemporal fusion IDs. Calendar fusion conversion automatically converts historical calendars such as the lunar calendar, Gregorian calendar, and sexagenary cycle year names to Julian days. Geographic hierarchy mapping converts historical place names into modern GIS coordinates through a spatiotemporal database, generating a 9-level GeoHash code. The spatiotemporal fusion ID generates a unique grid identifier. Dynamic graph construction is a causal-driven entity relationship network. The core technologies are spatiotemporally sensitive relationship weights and automatic causal chain discovery and conflict resolution mechanisms. Spatiotemporally sensitive relationship weights calculate relationship strength and relationship edges carrying time attributes based on co-occurrence frequency and time decay factors. Automatic causal chain discovery... The system uses a time-series graph convolutional network to identify event causality. A conflict resolution mechanism automatically removes erroneous relationship edges and labels the source of evidence. The federated verification pipeline aggregates evidence chains across institutions. The workflow involves encrypted query sending, collaborative verification of evidence packages, and conflict arbitration rules. First, the local system sends an encrypted spatiotemporal grid ID to the federated network. Then, collaborative verification of evidence packages is performed; for example, Archive A returns data, Archive B verifies the data, and finally, conflict arbitration is conducted. The multi-granularity output engine generates backtracking views on demand, including three output modes: macroscopic time slices, mesoscopic spatial heatmaps, and microscopic entity relationship chains. The incremental optimization mechanism is a feedback-driven model evolution, and the self-updating logic is that new evidence triggers local updates, ultimately resulting in a judgment model.

[0057] It is important to note that, such as Figure 4As shown, the non-decision data includes timestamps and geographic locations. The method for obtaining the output feedback information includes: based on the decision model, using timestamps and geographic locations as model inputs, obtaining the output information of the decision model to obtain timestamp output datasets and geographic location output datasets; performing information association based on timestamp output datasets and geographic location output datasets to obtain overlapping and non-overlapping information, obtaining a first information set and a second information set, using the first information set as feedback data, and then obtaining the output feedback information.

[0058] Specifically, after obtaining the judgment model, the timestamp and geographic location are used as separate inputs to obtain the model output. Since the output of the judgment model is different under different timestamps and different geographic locations, timestamp output datasets and geographic location output datasets are obtained separately. The timestamp output dataset and geographic location output dataset have overlapping and non-overlapping information. Therefore, the overlapping information is used as feedback data to obtain the output feedback information.

[0059] In the specific implementation process, such as Figure 5 As shown, based on the timestamps, geographical locations, person entities, organization entities, and topic tags obtained from file A, using person entities, organization entities, and topic tags as input data, a large language model is used to generate matching information corresponding to the input data, and a comparison model is created to obtain a judgment model. The timestamps in file A are three time points: 1, 2, and 3, and the geographical locations are three locations: a, b, and c. At this time, the timestamps and geographical locations are used as inputs to the model separately to obtain the model output results. The locations corresponding to the timestamps are a, q; b, d; w, and the time points corresponding to the geographical locations are 1, 4; 2, 5, 8; 9. Therefore, timestamp output datasets and geographical location output datasets are obtained respectively. The overlapping information of the timestamp output dataset and the geographical location output dataset is 1-a, 2-b, and the non-overlapping information is 1-q, 4-a, 2-d, 5-b, 8-b, 3-w, and 9-c. At this time, 1-a and 2-b are used as feedback data to obtain the output feedback information of file A.

[0060] It should be noted that the method for obtaining the summary information item includes: comparing the information output item with the text information item, obtaining the overlapping data between the judgment data in the text information item and the output feedback information, and using the overlapping data as the summary information item.

[0061] Data output module: Based on the decision model, it outputs question and answer data, including archive text data and metadata. Based on the text data and metadata, it generates question and answer data to obtain question and answer information items.

[0062] It is important to note that question-and-answer data generation based on text data and metadata can prevent the omission of key information and reduce misjudgment of information. Question-and-answer data generation includes annotation generation. The methods for obtaining question-and-answer information items include: annotating the archival text data and the target archival text to obtain question-and-answer annotation items, using the question-and-answer annotation items as question-and-answer data generation, and then obtaining question-and-answer information items.

[0063] Specifically, after creating the decision model, when question and answer output is performed based on the decision model, since the data output by the decision model simultaneously incorporates the content of the target archive text and the content generated by the large language model, the output data includes archive text data and metadata. The archive text data corresponds to the content of the target archive text, while the metadata corresponds to the content of the speech model. Therefore, data annotation is performed based on the archive text data and the target archive text to obtain question and answer annotation items. The question and answer annotation items are used as question and answer data to generate question and answer information items.

[0064] It is important to note that question-and-answer data generation also includes evidence generation. The methods for obtaining question-and-answer information items include: setting a data ratio value, removing metadata based on the data ratio value to obtain retained data, and obtaining retained information items; using the retained data items as question-and-answer evidence tags, associating question-and-answer annotation items with data, and then obtaining question-and-answer information items.

[0065] Specifically, the set data ratio is 50%. Since the data output by the judgment model combines the content of the target archive text and the content generated by the large language model, and the content generated by the large language model is large, some data is removed by setting the data ratio, and important data is retained. The retained data is used as evidence labels, and together with the data annotation of the archive text data and the target archive text data, data association is carried out to prevent misjudgment of information.

[0066] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended embodiments and their equivalents.

Claims

1. An archive text automatic summarization combined with a large model and a knowledge question and answer system, comprising: a text scanning module: performing text scanning on a target archive text, extracting text structured data through an entity recognition model, and obtaining text information items; characterized in that it further comprises: an information screening module: screening information based on the text information items, screening into determined data and non-determined data, taking the determined information as input data, generating matching information corresponding to the input data through a large language model, obtaining an input matching set, and the input matching set at least including one input matching item corresponding to the input data; a model creation module: creating a text comparison model based on the input matching set, obtaining a determination model, taking the non-determined data as the input item of the determination model, obtaining output feedback information, comparing the output feedback information with the text information items, obtaining a comparison matching result as summary data, obtaining summary information items, and then automatically summarizing and outputting the target archive text, reducing the influence of text noise on summary acquisition; a data output module: performing question and answer output based on the determination model, and the output data including archive text data and metadata, generating question and answer data based on the text data and the metadata, and obtaining question and answer information items, preventing key information from being missed while reducing information misjudgment.

2. The system of claim 1, wherein: The structured data includes timestamps, geographic locations, person entities, institution entities, and topic labels, and the method for obtaining the text information items comprises: performing preliminary scanning on the archive text to obtain keyword information, including geographic keywords, person keywords, institution keywords, and subject label keywords, and obtaining key information items; synchronously analyzing OCR text, seal images, and table data, performing cross-modal alignment, and obtaining alignment information items corresponding to the key information items; comparing and combining the key information items and the alignment information items to obtain the text information items.

3. The system of claim 2, wherein: The method for obtaining the text information items further comprises: performing preliminary scanning on the archive text to obtain fuzzy and missing fonts in the archive text, and obtaining interference information items; creating a broken text repair model, repairing the fuzzy and missing fonts in the scanned document based on the broken text repair model and the context calendar, repairing the interference information items, and obtaining repaired information items; comparing and combining the repaired information items, the key information items, and the alignment information items to obtain the text information items.

4. The system of claim 2, wherein: The determination information includes person entities, institution entities, and topic labels, and the method for obtaining the input matching set comprises: based on the determination information, creating an entity relationship three-dimensional mapping, matching different processing logics based on different input entities, and obtaining entity mapping items; based on the entity mapping items, performing implicit relationship mining, reasoning through spatiotemporal overlap relationships, obtaining spatial intersection information and temporal causal information, outputting extended information, and then obtaining the input matching set.

5. The system of claim 1, wherein: The method for obtaining the determination model comprises: creating an information model architecture diagram based on the input matching set, including spatiotemporal grid encoding, dynamic graph construction, federated verification pipeline, multi-granularity output, and incremental optimization; The non-deterministic data includes a timestamp and a geographic position, and the output feedback information acquisition method includes:

6. The system of claim 1, wherein: Based on the determination model, the timestamp and the geographic position are taken as the model input, the output information of the determination model is acquired, the timestamp output data set and the geographic position output data set are obtained; Based on the timestamp output data set and the geographic position output data set, information correlation is performed, the coincident information and the non-coincident information are acquired, the first information set and the second information set are obtained, the first information set is taken as the feedback data, and then the output feedback information is obtained. The abstract information item acquisition method includes:

7. The system of claim 1, wherein: Based on information comparison between the information output item and the text information item, the coincident data of the determination data and the output feedback information in the text information item is acquired, and the coincident data is taken as the abstract information item. The question and answer data generation includes label generation, and the question and answer information item acquisition method includes:

8. The system of claim 1, wherein: Based on data labeling of the archive text data and the target archive text, the question and answer label item is obtained, the question and answer data generation is taken as the question and answer label item, and then the question and answer information item is obtained. The question and answer data generation also includes evidence generation, and the question and answer information item acquisition method also includes:

9. The system of claim 8, wherein: A data proportion value is set, data elimination is performed on the metadata based on the data proportion value, the reserved data is acquired, and the reserved information item is obtained; The reserved data item is taken as the question and answer evidence label, data correlation is performed on the question and answer label item, and then the question and answer information item is obtained. ​

Images