Information extraction and dynamic updating method, system, equipment and medium

Abstract

The invention relates to the field of large models, in particular to an information extraction and dynamic updating method, system and device and a medium. The method comprises the following steps: performing cross-modal alignment on multi-modal data in an enterprise webpage to obtain standard data; inputting the standard data into a preset fine tuning and enlarging model, and extracting key information; encoding the key information by using double encoders to obtain a semantic feature vector, and carrying out context comparison on the semantic feature vector through a sliding window to judge an update boundary; and judging updating information according to the updating boundary and a preset time attenuation weight, and mapping the updating information into a knowledge graph with a timestamp. Standard data are obtained through cross-modal alignment, data of different modals can be unified to the same semantic framework, and semantic gaps and format differences among the data are eliminated. Standard data are input into a preset fine tuning model to extract key information, irrelevant noise information can be filtered out, and the method focuses on content valuable to enterprises.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of large model, and in particular, to an information extraction and dynamic updating method, system, device and medium. BACKGROUND

[0002] In the current era of digital information explosion, unstructured data such as text, images, audio, etc. grows at an astonishing rate. These data contain rich knowledge, but at the same time, they also pose great challenges to information extraction and processing. Currently, most mainstream information extraction technologies are based on rule templates or small-scale models. Rule template methods usually require manual pre-definition of a series of complex rules. These rules can only cover data in specific domains or specific formats, and their limitations are obvious when faced with unstructured data with diverse formats and complex semantics. For example, when processing news reports, different media may have very different ways of expressing the same event, and rule templates are difficult to fully adapt, resulting in the inability to effectively analyze the complex semantic associations in the data.

[0003] Small-scale models can achieve certain results on certain specific tasks, but due to their model size and training data limitations, they lack semantic understanding when dealing with large-scale, diverse unstructured data. They are difficult to capture deep semantic relationships in data and are powerless to some implicit semantic information, making it difficult to accurately extract valuable information and meet the needs of complex semantic analysis in practical applications.

[0004] Existing knowledge graph construction schemes also have many shortcomings. As a structured knowledge representation form, knowledge graphs play an important role in information integration and knowledge reasoning. However, current knowledge graph construction schemes lack adaptive structuring capabilities for multi-source heterogeneous data. Multi-source heterogeneous data comes from a wide range of sources and has diverse formats, including structured data, semi-structured data, and unstructured data. Existing schemes are difficult to automatically adjust the structure of the knowledge graph according to the characteristics of different data sources, making it difficult to integrate multi-source data. In addition, the cost of updating and maintaining the knowledge graph is high. As data continues to update and knowledge continues to accumulate, the knowledge graph needs to be constantly updated and maintained to ensure its accuracy and timeliness. However, existing updating and maintenance methods often require a lot of manual intervention, which is not only inefficient but also prone to errors, increasing the cost of knowledge graph construction and use. SUMMARY

[0005] The present disclosure provides an information extraction and dynamic updating method, system, device and medium to improve the completeness and accuracy of information extraction and dynamic updating information.

[0006] According to a first aspect of the present disclosure, a method for information extraction and dynamic updating is provided, the method comprising:

[0007] aligning multi-modal data in enterprise webpages to obtain standard data;

[0008] inputting the standard data into a preset fine-tuned large model to extract key information;

[0009] encoding the key information using a double encoder to obtain a semantic feature vector, and determining an update boundary by comparing the semantic feature vector with context using a sliding window;

[0010] determining update information according to the update boundary and a preset time decay weight, and mapping the update information into a timestamped knowledge graph.

[0011] Optionally, the step of aligning multi-modal data in enterprise webpages to obtain standard data comprises:

[0012] scraping multi-modal data containing text, images or videos from enterprise webpages, and removing duplicate content to obtain a preliminarily screened multi-modal data set;

[0013] performing noise reduction and cleaning on the text part of the multi-modal data set to obtain cleaned text;

[0014] performing semantic alignment between images, videos and the cleaned text through image recognition and voice recognition techniques to obtain standard data.

[0015] Optionally, the step of constructing a preset fine-tuned large model comprises:

[0016] embedding a pluggable theme adapter layer in a pre-trained large model to obtain an adapter-containing model;

[0017] training the adapter-containing model through LoRA fine-tuning technology to obtain a fine-tuned large model.

[0018] Optionally, the step of inputting the standard data into a preset fine-tuned large model to extract key information comprises:

[0019] adjusting the format of the standard data into an adapted format data;

[0020] inputting the adapted format data into the fine-tuned large model for inference analysis to obtain structured information;

[0021] integrating and processing the structured information as key information.

[0022] Optionally, the step of encoding the key information by the dual encoder to obtain a semantic feature vector, and determining the update boundary by context comparison of the semantic feature vector by a sliding window comprises:

[0023] inputting the key information into one of the dual encoders, extracting overall semantic information of the key information, and generating an overall representation vector in the semantic feature vector;

[0024] encoding each paragraph of the key information by another encoder to obtain a paragraph-level feature vector in the semantic feature vector;

[0025] performing context comparison of the paragraph-level feature vector by a sliding window, and determining the update boundary according to a cosine similarity of context and a semantic coherence index.

[0026] Optionally, the step of performing context comparison of the paragraph-level feature vector by a sliding window, and determining the update boundary according to a cosine similarity of context and a semantic coherence index comprises:

[0027] obtaining parameters of the sliding window, placing a window at a starting position of the paragraph-level feature vector sequence to obtain an initial window state;

[0028] calculating a cosine similarity between adjacent vectors to obtain a set of cosine similarity values;

[0029] evaluating content semantic coherence in the window by semantic logic to obtain a coherence evaluation score as a semantic coherence index;

[0030] determining whether it is an update boundary based on the set of cosine similarity values and the semantic coherence index, and if it is not a boundary, sliding the window until the entire sequence is traversed to obtain the update boundary.

[0031] Optionally, the step of determining the update information according to the update boundary and a preset time decay weight, and mapping the update information into a timestamped knowledge graph comprises:

[0032] extracting data related to the update from the key information according to the update boundary, and outputting data corresponding to the update boundary;

[0033] judging new and old information of the data corresponding to the update boundary based on an incremental determination algorithm of the time decay weight to obtain an update information determination result;

[0034] extracting structured feature data of the update information according to the update information determination result;

[0035] inputting the structured feature data into a preset multi-hop relationship reasoning module, reasoning and mapping into a timestamped knowledge graph triple, and outputting the timestamped knowledge graph.

[0036] According to a second aspect of the present disclosure, there is provided an information extraction and dynamic updating system, comprising:

[0037] an alignment module configured to perform cross-modal alignment on multi-modal data in the enterprise webpage to obtain standard data;

[0038] an extraction module configured to input the standard data into a preset fine-tuning large model to extract key information;

[0039] an updating module configured to encode the key information using a double encoder to obtain a semantic feature vector, and determine an updating boundary by comparing the semantic feature vector using a sliding window;

[0040] a mapping module configured to determine updating information according to the updating boundary and a preset time decay weight, and map the updating information into a timestamped knowledge graph.

[0041] According to a third aspect of the present disclosure, there is provided an electronic device. The electronic device comprises a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the method described above.

[0042] According to a fourth aspect of the present disclosure, there is provided a computer-readable storage medium, which stores a computer program, and the program is executed by a processor to implement the method described above.

[0043] According to a fifth aspect of the present disclosure, there is provided a computer program product comprising a computer program, and the computer program is executed by a processor to implement the method described above.

[0044] The information extraction and dynamic updating method, system, device and medium provided by the embodiments of the present disclosure can unify data of different modalities into the same semantic framework by cross-modal alignment to eliminate semantic gaps and format differences between data. The key information can be extracted by inputting the standard data into a preset fine-tuning large model to filter out irrelevant noise information and focus on valuable content for the enterprise. The semantic feature vector can be obtained by encoding the key information using a double encoder to more deeply mine the semantic information of the data. The updating boundary can be determined by comparing the semantic feature vector using a sliding window to accurately identify changes in the semantic level of the data. The updating information can be determined according to the updating boundary and a preset time decay weight to comprehensively consider the semantic changes and time factors of the data. The time decay weight can make the newer updating information occupy a larger proportion in the determination, while the influence of the older updating information gradually decreases. The updating information can be mapped into a timestamped knowledge graph to record the updating time and change process of the knowledge. This makes the knowledge graph not only contain the current knowledge state, but also retain the historical evolution information of the knowledge. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Further details, features and advantages of the present disclosure are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which:

[0046] Figure 1 A flowchart of the overall technical solution of the information extraction and dynamic update method provided by an exemplary embodiment of the present disclosure;

[0047] Figure 2 A flowchart of an information extraction and dynamic update system provided by an exemplary embodiment of the present disclosure;

[0048] Figure 3 A structural block diagram of an electronic device provided as an exemplary embodiment of the present disclosure;

[0049] Figure 4 A structural block diagram of a computer system provided by an exemplary embodiment of the present disclosure. DETAILED DESCRIPTION

[0050] The following describes embodiments of the present disclosure in more detail with reference to the accompanying drawings. Although certain embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are for illustrative purposes only and are not intended to limit the scope of protection of the present disclosure.

[0051] It should be understood that the various steps described in the method embodiments of the present disclosure may be performed in different orders and / or in parallel. In addition, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present disclosure is not limited in this respect.

[0052] The term "including" and its variations used in this document are open inclusions, that is, "including but not limited to". The term "based on" means "based at least in part on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one other embodiment"; the term "some embodiments" means "at least some embodiments". The relevant definitions of other terms will be given in the description below. It should be noted that the concepts of "first", "second", etc. mentioned in this disclosure are only used to distinguish different devices, modules or units, and are not used to limit the order or interdependence of the functions performed by these devices, modules or units.

[0053] It should be noted that the modification of "one" and "multiple" mentioned in the present disclosure is illustrative but not restrictive, and those skilled in the art should understand that "one or more" should be understood unless otherwise explicitly indicated in the context.

[0054] The names of the messages or information exchanged between the plurality of devices in the embodiments of the present disclosure are only for illustrative purposes, and are not used to limit the scope of the messages or information.

[0055] It can be understood that, before using the technical solutions disclosed in the embodiments of the present disclosure, the type, use range, use scenario, etc. of the personal information involved in the present disclosure should be informed to the user and the authorization of the user should be obtained through appropriate means according to relevant laws and regulations.

[0056] For example, in response to receiving the active request of the user, the prompt information is sent to the user to explicitly prompt the user that the operation requested to be performed will require obtaining and using the personal information of the user. Thus, the user can voluntarily choose whether to provide the personal information to the software or hardware such as electronic device, application program, server or storage medium, etc. that performs the operation of the technical solutions of the present disclosure according to the prompt information.

[0057] As an optional but not limited implementation manner, in response to receiving the active request of the user, the prompt information can be sent to the user in the form of a pop-up window, and the prompt information can be presented in the form of text in the pop-up window. In addition, the pop-up window can also carry selection controls for the user to select "agree" or "disagree" to provide personal information to the electronic device. It can be understood that the above notification and obtaining of user authorization process is only illustrative, and does not limit the implementation manner of the present disclosure, and other ways that meet the relevant laws and regulations can also be applied to the implementation manner of the present disclosure.

[0058] Current information extraction technology is mostly based on rule templates or small-scale models, which is difficult to effectively analyze the complex semantic association of unstructured data. Traditional dynamic updating methods rely on fixed frequency scanning or hash comparison, which cannot accurately identify the incremental changes at the content semantic level, and are prone to produce redundant storage or information omission. Existing knowledge graph construction schemes lack adaptive structuring capability for multi-source heterogeneous data, and the update and maintenance cost is high. With the development of large model technology, it is urgent to explore its application potential in semantic understanding and dynamic perception.

[0059] The application aims to solve the problems of low efficiency and uncontrollable quality generated by conventional RAG methods for information retrieval, and meet the application requirements of specific fields. Multi-modal semantic understanding module: a preset theme adapter is constructed based on a large model, and precise structured extraction of unstructured data is realized through deep semantic analysis (improve information extraction accuracy) Dynamic incremental detection algorithm: semantic vector comparison and context association analysis are used to identify document update content in real time (reduce redundant storage) Knowledge graph self-maintenance system: a structured database with version traceability is constructed to support multi-dimensional association query (improve retrieval efficiency).

[0060] In the embodiments of the present disclosure, multi-modal data (such as text, images, videos, etc.) in enterprise web pages often have semantic associations but different formats and expression methods. By aligning across modalities to obtain standard data, the data of different modalities can be unified under the same semantic framework, eliminating the semantic gap and format difference between the data. For example, a product picture and the corresponding text description in a web page can be aligned across modalities to ensure that the product features displayed in the picture are consistent with the text description, making the data more accurate and standardized, and providing a high-quality data basis for subsequent processing. Inputting the standard data into the preset fine-tuning large model to extract key information can filter out irrelevant noise information and focus on valuable content for the enterprise. For example, when processing enterprise news web pages, the large model can extract the core points of news events, the products or services involved, key personnel, and other information, reducing the complexity and computational load of subsequent processing.

[0061] Encoding the key information using a double encoder to obtain semantic feature vectors can more deeply mine the semantic information of the data. The double encoder can encode the data from different angles to capture more rich semantic features. For example, one encoder focuses on the semantic structure of the text, and the other encoder focuses on the sentiment tendency of the text, and the combination of the two can more comprehensively understand the semantics of the data. Comparing the semantic feature vectors through a sliding window to determine the update boundary can accurately identify the semantic changes of the data. For example, in an enterprise product introduction web page, when the product functions or parameters are updated, the sliding window can compare the semantic feature vectors before and after the update to determine the start and end positions of the update, avoiding false positives due to minor format changes, and improving the accuracy of update determination.

[0062] Determining the update information according to the update boundary and the preset time decay weight can comprehensively consider the semantic changes and time factors of the data. The time decay weight can make the newer update information have a greater weight in the determination, while the influence of older update information gradually decreases. For example, in an enterprise market dynamics web page, recent market trend information is more important, and the time decay weight can highlight these new information while retaining some historical information as a reference, making the determination of update information more reasonable.

[0063] The traditional dynamic updating method is prone to redundant storage or information omission, while the present method can accurately identify the content that needs to be updated by precise cross-modal alignment, key information extraction and update boundary determination, and only updates the part that has really changed, avoiding unnecessary storage overhead. At the same time, since it can capture subtle changes at the semantic level, it also reduces the possibility of information omission. For example, in the enterprise product evaluation webpage, even if only the evaluation content of individual users has changed, the method can accurately identify and update it, ensuring that the information in the knowledge graph is always up-to-date and complete.

[0064] Mapping the updated information into a timestamped knowledge graph can record the update time and change process of knowledge. This makes the knowledge graph not only contain the current knowledge state, but also retain the historical evolution information of knowledge. For example, in the enterprise technology development knowledge graph, the R&D process of a certain technology, the achievements at different stages and the update time can be clearly seen, providing a more comprehensive reference for the enterprise's technology decision-making.

[0065] The timestamped knowledge graph contains rich semantic information that has been aligned across modalities and encoded semantically, which can support more complex knowledge reasoning and application scenarios. For example, in enterprise market analysis, multi-modal data and time series information in the knowledge graph can be used for market trend prediction, competitor analysis and other advanced applications, providing strong support for enterprise strategic decision-making.

[0066] The timestamped knowledge graph makes it more convenient to maintain and update the knowledge graph. When adding, modifying or deleting knowledge, you can quickly locate the relevant knowledge nodes according to the timestamp information and perform targeted operations. At the same time, the timestamp information can also be used to evaluate the update frequency and quality of the knowledge graph, providing a basis for optimizing the knowledge graph.

[0067] In order to clearly describe the scheme, the complete technical solution of the embodiments of the present disclosure will be described in detail below. The complete technical solution of the present disclosure includes the training process and application process of the information extraction and dynamic updating model.

[0068] As shown in Figure 1 The information extraction and dynamic updating method provided by the embodiments of the present disclosure can include the following steps:

[0069] S1, cross-modal alignment of multi-modal data in enterprise webpages to obtain standard data;

[0070] S2, inputting the standard data into a pre-set fine-tuning large model to extract key information;

[0071] S3, encoding the key information with a double encoder to obtain a semantic feature vector, and determining the update boundary by comparing the context of the semantic feature vector.

[0072] S4, determining update information according to the update boundary and a preset time decay weight, and mapping the update information into a timestamped knowledge graph.

[0073] The application constructs a multi-modal data preprocessing channel, realizes standardized input of documents / web pages through text cleaning and cross-modal alignment technology. A pluggable theme adapter layer is embedded in the pre-trained large model to improve the accuracy of structured information extraction in specific fields through LoRA fine-tuning. A double-encoder architecture is used to generate semantic vectors, and overall document representation and fine-grained paragraph-level features are extracted for update detection. A sliding window context comparison mechanism is designed, combined with cosine similarity and semantic coherence indicators to determine the content update boundary. A time decay weight-based incremental determination algorithm is developed to dynamically adjust the new and old version information coverage threshold. A multi-hop relationship reasoning module is built to map the extracted structured data into timestamped knowledge graph triples. A hybrid index storage engine is deployed to integrate graph databases and vector databases for multi-condition joint queries.

[0074] The multi-modal data in the enterprise web page is aligned across modalities to obtain standard data, eliminating differences and noise between different modal data, making the data format uniform and the semantics clear, laying a solid foundation for subsequent processing, greatly improving data usability and quality, and facilitating subsequent information extraction and analysis.

[0075] The standard data is input into the fine-tuned large model to extract key information. The large model has strong semantic understanding ability and can accurately capture the core points of the data, remove redundant information, improve the accuracy and efficiency of information extraction, and provide high-quality data support for subsequent knowledge graph construction.

[0076] The semantic feature vector is obtained by double-encoder encoding the key information, and the update boundary is determined by sliding window context comparison, which can accurately identify the semantic changes of the data and avoid information omission or redundant storage caused by the inability to perceive semantic changes in traditional methods, ensuring the accuracy and integrity of the update information.

[0077] According to the update boundary and the time decay weight, the update information is determined and mapped into a timestamped knowledge graph, so that the knowledge graph can dynamically reflect the changes of enterprise information. The time decay weight considers the timeliness of the information, and the timestamp record enables the knowledge graph to have historical tracing ability, providing more timely, accurate and comprehensive knowledge support for enterprise decision-making, and helping enterprises better cope with market changes.

[0078] In some embodiments, as shown in Figure 1 The step of aligning the multi-modal data in the enterprise web page across modalities to obtain standard data includes:

[0079] The multi-modal data containing text, images or videos is scraped from the enterprise webpage, and the repeated content is removed to obtain a preliminary screened multi-modal data set.

[0080] The text part in the multi-modal data set is denoised and cleaned to obtain cleaned text.

[0081] The images and videos are semantically aligned with the cleaned text through image recognition and speech recognition techniques to obtain standard data.

[0082] For text data in documents or webpages, various cleaning strategies are used. Remove irrelevant information such as HTML tags, special characters, redundant spaces, etc. Correct spelling errors, and standardize different formats of dates, numbers, etc. For example, "2024-10-01" is converted to "2024-10-01". At the same time, stop words such as "of", "of", "and" are identified and processed to reduce noise interference and improve the quality of subsequent processing.

[0083] For documents or webpages containing multi-modal data such as text, images, and videos, cross-modal alignment is performed. Use image recognition technology to extract key information from images, such as object names, scene descriptions, etc., and associate them with the corresponding text content. For video data, audio is converted to text through speech recognition, and then aligned with images and subtitles in the video to achieve unified representation of multi-modal data, preparing for subsequent standardized input.

[0084] In some embodiments, as shown in Figure 1 The steps of constructing the pre-set fine-tuning large model include:

[0085] Embedding a pluggable theme adapter layer in the pre-trained large model to obtain an adapter-containing model;

[0086] Training the adapter-containing model through LoRA fine-tuning technology to obtain a fine-tuned large model.

[0087] Embed a pluggable theme adapter layer in the pre-trained large model, which can be customized according to different specific fields. For example, in the financial field, the adapter layer can learn financial terminology, industry rules and business logic; in the medical field, it focuses on medical knowledge, disease diagnosis and treatment plans, etc. In this way, the model can better adapt to the data characteristics of different fields and improve the relevance of information extraction.

[0088] The pre-trained large model is fine-tuned using LoRA (Low-Rank Adaptation) fine-tuning technology. LoRA reduces the number of parameters that need to be trained by introducing a low-rank matrix into the model parameters, thereby reducing computational cost and time consumption. Fine-tuning on a specific domain dataset enables the model to learn structured information extraction patterns in that domain, improving extraction accuracy.

[0089] In some embodiments, as shown in Figure 1 The standard data is input into a pre-set fine-tuned large model, and the step of extracting key information includes:

[0090] The format of the standard data is adjusted to adaptive format data;

[0091] The adaptive format data is input into the fine-tuned large model for inference analysis to obtain structured information;

[0092] The structured information is integrated and processed as key information.

[0093] In some embodiments, as shown in Figure 1 The key information is encoded using a dual encoder to obtain a semantic feature vector, and the step of determining the update boundary by comparing the context of the semantic feature vector using a sliding window includes:

[0094] The key information is input into one of the dual encoders, the overall semantic information of the key information is extracted, and an overall representation vector in the semantic feature vector is generated;

[0095] Each paragraph of the key information is encoded by another encoder to obtain a paragraph-level feature vector in the semantic feature vector;

[0096] The sliding window compares the context of the paragraph-level feature vector, and determines the update boundary according to the cosine similarity of the context and the semantic coherence index.

[0097] One encoder is used to encode the document as a whole to extract the global semantic information of the document. The encoder can use advanced neural network structures such as Transformer, which can capture the semantic associations between different parts of the document through self-attention mechanisms. For example, for a news report, the overall representation can reflect the theme, main points and overall sentiment of the report.

[0098] Another encoder focuses on paragraph-level feature extraction of the document. The document is divided into multiple paragraphs, and each paragraph is encoded to obtain a semantic vector for each paragraph. This way, local detailed information in the document, such as specific event descriptions and argument support, can be captured.

[0099] A sliding window mechanism is designed to compare the context of different parts of the document. The size of the sliding window can be adjusted according to the actual situation, for example, sliding in paragraph units. During the sliding process, the semantic similarity between the current window and the adjacent window is calculated, while considering the semantic coherence index. Semantic similarity can be calculated by methods such as cosine similarity, and semantic coherence can be evaluated by analyzing the logical structure of the text, lexical cohesion, etc.

[0100] The dual encoder generates semantic vectors for the overall representation of the document and the fine-grained paragraph-level features. These semantic vectors are a high-level abstraction and summary of the document content, while the paragraph-level feature vectors are specific to the content of each paragraph discussing subfields such as artificial intelligence, quantum computing, etc. When the sliding window determines the content update boundary, it needs to determine the similarity and difference between different parts of the content based on these semantic vectors.

[0101] Semantic vectors convert text content into numerical form, making it possible to compare and calculate content. By calculating the similarity between vectors, the similarity between different parts of the content can be quantified. For example, the cosine similarity between two paragraph vectors is calculated by calculating the cosine value of the included angle. The closer the value is to 1, the more similar the semantics of the two paragraphs; the closer the value is to 0, the greater the semantic difference. This quantification provides a basis for the sliding window to determine the update boundary.

[0102] Using the features extracted by the dual encoder, the sliding window can more accurately locate the content update boundary. Since the feature vector can capture the semantic details of the document, the sliding window can determine the start and end positions of the update based on significant changes in semantics.

[0103] In some embodiments, as shown in Figure 1 The sliding window compares the context of the paragraph-level feature vectors, and the steps for determining the update boundary based on the cosine similarity of the context and the semantic coherence index include:

[0104] Obtain the parameters of the sliding window, place the window at the starting position of the sequence of paragraph-level feature vectors, and obtain the initial window state;

[0105] Calculate the cosine similarity between adjacent vectors to obtain a set of cosine similarity values;

[0106] Evaluate the semantic coherence of the content within the window through semantic logic to obtain a coherence evaluation score as a semantic coherence index;

[0107] Based on the set of cosine similarity values and the semantic coherence index, determine whether it is an update boundary. If it is not a boundary, slide the window until it is traversed to obtain the update boundary.

[0108] According to the changes of semantic similarity and semantic coherence indicators, the content update boundary is determined. When the semantic similarity suddenly decreases and the semantic coherence is destroyed, it may mean that there is a content update. In the specification of an enterprise user manual, when a new function is introduced, the semantic similarity with the previous content will obviously decrease, and the semantic coherence will also be affected. At this time, it can be determined that the new function starts, that is, the update boundary.

[0109] The sliding window mechanism compares the content in the window with the adjacent window or reference content when it slides on the document. This comparison is based on the semantic vectors extracted by the dual encoder. When the sliding window moves to a certain position in the document, the similarity between the paragraph vectors in the current window and the paragraph vectors in the previous window is calculated. If the similarity suddenly decreases, it means that there is a content update.

[0110] In some embodiments, as shown in Figure 1 The step of mapping the update information into a timestamped knowledge graph according to the update boundary and the preset time decay weight includes:

[0111] According to the update boundary, the data related to the update is extracted from the key information, and the data corresponding to the update boundary is output.

[0112] The old and new information of the data corresponding to the update boundary is determined based on the time decay weight incremental determination algorithm, and the update information determination result is obtained.

[0113] According to the update information determination result, the structured feature data of the update information is extracted.

[0114] The structured feature data is input into a preset multi-hop relationship reasoning module, and is reasoned and mapped into a timestamped knowledge graph triple, and a timestamped knowledge graph is output.

[0115] The time decay weight incremental determination algorithm is developed to assign different weights to information at different time points. As time goes by, the weight of old information gradually decreases, and the weight of new information is relatively high. An exponential decay function is used to calculate the time decay weight, so that the information updated recently has greater influence in the determination process.

[0116] According to the time decay weight and the update situation of the document, the new and old version information coverage threshold is dynamically adjusted. When a large number of updates are detected, the threshold is appropriately reduced to more sensitively capture changes; when the update is less, the threshold is increased to avoid misjudgment. In this way, it can more accurately determine which information is newly added and which is a continuation or modification of old information.

[0117] A multi-hop relationship reasoning module is built to conduct in-depth analysis on the extracted structured data. By analyzing the indirect relationships between entities, complex knowledge hidden in the data is mined.

[0118] The extracted structured data is mapped into timestamped knowledge graph triples. Each triple consists of two entities and a relationship, such as (Company A, holding, Company B), and records the time when the relationship occurred. The knowledge graph visually displays the relationships between entities, facilitating intuitive understanding and analysis of information. Meanwhile, the introduction of timestamps enables the knowledge graph to reflect dynamic changes in information, providing more comprehensive information for subsequent queries and analysis.

[0119] According to the present disclosure, an information extraction and dynamic updating system is provided, as shown in Figure 2 The system comprises:

[0120] An alignment module 100 is configured to perform cross-modal alignment on multi-modal data in enterprise webpages to obtain standard data.

[0121] An extraction module 200 is configured to input the standard data into a pre-set fine-tuning large model to extract key information.

[0122] An updating module 300 is configured to encode the key information using a double encoder to obtain semantic feature vectors, and determine the update boundary by comparing the context of the semantic feature vectors using a sliding window.

[0123] A mapping module 400 is configured to determine update information according to the update boundary and a pre-set time decay weight, and map the update information into a timestamped knowledge graph.

[0124] The system also performs database deployment, deploying a hybrid index storage engine, integrating a graph database and a vector database. The graph database is good at processing relationship queries between entities and can efficiently store and query complex relationships in the knowledge graph. The vector database is used to store semantic vectors of documents, supporting fast retrieval based on semantic similarity. By integrating the two, joint storage and query of multi-modal data can be achieved.

[0125] Optimization processing is performed on the data stored in the database to improve storage efficiency and query performance. For example, reasonable index design is performed on the nodes and edges in the graph database to reduce the search space during query; compression and dimensionality reduction processing is performed on the vectors in the vector database to reduce storage space and computational cost.

[0126] A friendly query interface is developed to support users to query in multiple ways. Users can input keywords, upload pictures or video clips, specify the time range of the query, etc. The system will perform joint query in the hybrid index storage engine according to the user input conditions.

[0127] The query results are integrated and displayed, combining the relationship information in the graph database and the semantic information in the vector database to provide comprehensive and accurate query results for users. For example, when a user queries the relevant information of a company, the system will not only return the basic information and equity structure of the company, but also display the semantic related documents such as news reports and research reports related to the company. At the same time, the query results are presented in a visual manner, such as the visualization of the knowledge graph, the sorting and filtering of the document list, etc., to improve the user's query experience.

[0128] The embodiments of the present disclosure also provide an electronic device, comprising: at least one processor; a memory for storing instructions executable by the at least one processor; wherein the at least one processor is configured to execute the instructions to implement the above-mentioned method disclosed by the embodiments of the present disclosure.

[0129] Figure 3 The structure schematic diagram of the electronic device provided by an exemplary embodiment of the present disclosure is provided. As shown in the figure, the electronic device 300 includes at least one processor 301 and a memory 302 coupled to the processor 301, and the processor 301 can execute the corresponding steps in the above-mentioned method disclosed by the embodiments of the present disclosure. Figure 3

[0130] ​The processor 301 can also be referred to as a central processing unit (CPU), which can be an integrated circuit chip that has the processing capability of signals. Each step in the method disclosed in the embodiments of the present disclosure can be completed by the integrated logic circuit of hardware or the instructions in the form of software in the processor 301. The processor 301 can be a general-purpose processor, a digital signal processor (DSP), an ASIC (Application Specific Integrated Circuit), a field-programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in conjunction with the embodiments of the present disclosure can be directly embodied as hardware code processing for execution, or executed by a combination of hardware and software modules in the code processing. The software module can be located in the memory 302, such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, and other mature storage media in the art. The processor 301 reads the information in the memory 302, and completes the steps of the above method in combination with the hardware thereof.

[0131] In addition, various operations / processes according to the present disclosure, when implemented by software and / or firmware, can be installed from a storage medium or a network to a computer system having a dedicated hardware structure, for example Figure 4 The computer system 1900 shown is intended to represent various forms of digital electronic computer equipment, such as a laptop computer, a desktop computer, a workstation, a personal digital assistant, a server, a blade server, a mainframe computer, and other suitable computer. The electronic equipment can also represent various forms of mobile devices, such as a personal digital processor, a cellular phone, a smart phone, a wearable device, and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are merely examples, and are not intended to limit the implementation of the present disclosure described and / or claimed herein. Figure 4 A structural block diagram of a computer system provided for an exemplary embodiment of the present disclosure.

[0132] The computer system 1900 is intended to represent various forms of digital electronic computer equipment, such as a laptop computer, a desktop computer, a workstation, a personal digital assistant, a server, a blade server, a mainframe computer, and other suitable computer. The electronic equipment can also represent various forms of mobile devices, such as a personal digital processor, a cellular phone, a smart phone, a wearable device, and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are merely examples, and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0133] As shown in FIG. 19, the computer system 1900 includes one or more processors, such as a processor 1902, a memory 1904, a storage 1906, an input / output (I / O) interface 1908, a communication interface 1910, and a bus 1912. Figure 4As shown, the computer system 1900 includes a computing unit 1901 that can perform various appropriate actions and processes in accordance with a computer program stored in a read-only memory (ROM) 1902 or a computer program loaded from the storage unit 1908 into a random access memory (RAM) 1903. Various programs and data required for the operation of the computer system 1900 can also be stored in the RAM 1903. The computing unit 1901, the ROM 1902, and the RAM 1903 are connected to each other through a bus 1904. An input / output (I / O) interface 1905 is also connected to the bus 1904.

[0134] A plurality of components in the computer system 1900 are connected to the I / O interface 1905, including an input unit 1906, an output unit 1907, a storage unit 1908, and a communication unit 1909. The input unit 1906 can be any type of device that can input information to the computer system 1900, and can receive inputted digital or character information, and generate key signal inputs related to user settings and / or function controls of the electronic device. The output unit 1907 can be any type of device that can present information, and can include, but is not limited to, a display, a speaker, a video / audio output terminal, a vibrator, and / or a printer. The storage unit 1908 can include, but is not limited to, a magnetic disk, an optical disk. The communication unit 1909 allows the computer system 1900 to exchange information / data with other devices through a network such as the Internet, and can include, but is not limited to, a modem, a network card, an infrared communication device, a wireless communication transceiver, and / or a chipset, such as a Bluetooth™ device, a WiFi device, a WiMax device, a cellular communication device, and / or the like.

[0135] The computing unit 1901 can be various general and / or special purpose processing components with processing and computing capabilities. Some examples of the computing unit 1901 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The computing unit 1901 performs various methods and processes described above. For example, in some embodiments, the above-described methods disclosed by embodiments of the present disclosure can be implemented as a computer software program tangibly embodied in a machine-readable medium, such as the storage unit 1908. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device via the ROM 1902 and / or the communication unit 1909. In some embodiments, the computing unit 1901 can be configured to perform the above-described methods disclosed by embodiments of the present disclosure by any other appropriate means (e.g., by means of firmware).

[0136] The embodiment of the present disclosure further provides a computer readable storage medium, wherein when instructions in the computer readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform the above method disclosed by the embodiment of the present disclosure.

[0137] The computer readable storage medium in the embodiment of the present disclosure can be a tangible medium, which can contain or store programs for use by or in connection with an instruction execution system, apparatus or device. The above computer readable storage medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus or device, or any suitable combination of the above. More specifically, the above computer readable storage medium can include one or more wire-based electrical connections, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the above.

[0138] The above computer readable medium can be included in the above electronic device; or can exist separately without being assembled into the electronic device.

[0139] The embodiment of the present disclosure further provides a computer program product, comprising a computer program, wherein the computer program is executed by a processor to implement the above method disclosed by the embodiment of the present disclosure.

[0140] In the embodiments of the present disclosure, computer program code for carrying out operations of the present disclosure can be written in one or more programming languages or combinations of the same, including object oriented programming languages, such as Java, Smalltalk, C++, and conventional procedural programming languages, such as "C" language or similar programming languages. The program code can be executed completely on a user computer, partially on the user computer and partially on a remote computer, as a separate software package, partially on the user computer and partially on a remote computer, or completely on a remote computer or server. In the case involving a remote computer, the remote computer can be connected to the user computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer.

[0141] The flow diagrams and the block diagrams in the drawings are illustrations of architectures, functionalities, and operations of possible implementations of systems, methods, and computer program products according to various embodiments of present disclosure. In this regard, each block in the flow diagrams or block diagrams can represent a module, a segment, or a portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that in some alternative implementations, the functions noted in the block can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and / or flow diagrams, and combinations thereof, can be implemented by special purpose hardware-based systems that perform the specified functions or operations, or combinations of special purpose hardware and computer instructions.

[0142] The modules, components or units described in the embodiments of the present disclosure can be implemented by software or by hardware. In some cases, the name of the module, component or unit does not constitute a limitation on the module, component or unit itself.

[0143] The functions described in this specification can be implemented in part or in whole by one or more hardware logic components. For example, and without limitation, illustrative types of hardware logic components that can be used include Field-programmable Gate Arrays (FPGAs), Program-specific Integrated Circuits (ASICs), Program-specific Standard Products (ASSPs), System-on-a-chip systems (SOCs), Complex Programmable Logic Devices (CPLDs), etc.

[0144] The above description is merely illustrative of the embodiments of the present disclosure and the principles of the technology involved. It is understood that the disclosure scope of the present disclosure is not limited to the technical solutions formed by the specific combinations of the above technical features, and should also cover other technical solutions formed by the combinations of the above technical features or equivalent features thereof without departing from the above disclosure concept. For example, the technical solutions formed by the mutual replacement of the above features and the technical features disclosed in the present disclosure (but not limited to) having similar functions.

[0145] Although some specific embodiments of the present disclosure have been described in detail by way of examples, one skilled in the art should understand that the above examples are only for illustration, and are not intended to limit the scope of the present disclosure. One skilled in the art should understand that the above embodiments can be modified without departing from the scope and spirit of the present disclosure. The scope of the present disclosure is defined by the appended claims.

Claims

1. Information extraction and dynamic updating method, characterized in that: The method comprises: Perform cross-modal alignment on multimodal data in corporate web pages to obtain standard data; Input the standard data into a preset fine-tuned large model to extract key information; Encoding the key information using a dual encoder to obtain a semantic feature vector, and performing context comparison on the semantic feature vector using a sliding window to determine an update boundary; The update information is determined according to the update boundary and the preset time decay weight, and the update information is mapped into a knowledge graph with a timestamp.

2. The method according to claim 1, characterized in that The step of performing cross-modal alignment on the multimodal data in the enterprise webpage to obtain standard data includes: Capture multimodal data containing text, images, or videos from corporate web pages, remove duplicate content, and obtain a preliminarily screened multimodal dataset; Performing noise reduction and cleaning on the text portion of the multimodal dataset to obtain a cleaned text; The images and videos are semantically aligned with the cleaned text through image recognition and speech recognition technologies to obtain standard data.

3. The method according to claim 1, characterized in that The steps of constructing the preset fine-tuning large model include: Embed a pluggable topic adapter layer in the pre-trained large model to obtain a model with adapters; The adapter-containing model is trained using LoRA fine-tuning technology to obtain a fine-tuned large model.

4. The method according to claim 1, wherein The steps of inputting the standard data into the preset fine-tuned large model and extracting key information include: Adjusting the format of the standard data to adapted format data; Inputting the adapted format data into the fine-tuned large model for reasoning analysis to obtain structured information; The structured information is integrated and processed as key information.

5. The method according to claim 1, wherein The step of encoding the key information with a dual encoder to obtain a semantic feature vector, and performing context comparison on the semantic feature vector through a sliding window to determine an update boundary includes: Inputting the key information into one of the dual encoders, extracting the overall semantic information of the key information, and generating an overall representation vector in the semantic feature vector; By using another encoder, each paragraph of the key information is encoded separately to obtain a paragraph-level feature vector in the semantic feature vector; The sliding window performs contextual comparison on paragraph-level feature vectors, and determines the update boundary based on the cosine similarity and semantic coherence index of the context.

6. The method according to claim 5, characterized in that The sliding window performs context comparison on the paragraph-level feature vector and determines the update boundary according to the cosine similarity and semantic coherence index of the context, including: Obtaining parameters of a sliding window, placing the window at the starting position of the paragraph-level feature vector sequence, and obtaining an initial window state; Calculate the cosine similarity between adjacent vectors to obtain a set of cosine similarity values; The semantic coherence of the content in the window is evaluated through semantic logic to obtain a coherence evaluation score as a semantic coherence indicator; Based on the cosine similarity set and the semantic coherence index, it is determined whether it is an update boundary. If it is not a boundary, the window is slid until the traversal is completed to obtain the update boundary.

7. The method according to claim 1, characterized in that The step of determining update information according to the update boundary and a preset time decay weight, and mapping the update information into a knowledge graph with a timestamp includes: extracting update-related data from the key information according to the update boundary, and outputting data corresponding to the update boundary; An incremental judgment algorithm based on time decay weight is used to judge the new and old information of the data corresponding to the update boundary and obtain the update information judgment result; Extracting structured feature data of the update information according to the update information determination result; The structured feature data is input into a preset multi-hop relational reasoning module, inferred and mapped into a knowledge graph triple with a timestamp, and a knowledge graph with a timestamp is output.

8. Information extraction and dynamic update system, characterized by: The system comprises: an alignment module configured to perform cross-modal alignment on multimodal data in an enterprise webpage to obtain standard data; an extraction module configured to input the standard data into a preset fine-tuning large model to extract key information; an updating module configured to encode the key information using a dual encoder to obtain a semantic feature vector, and perform context comparison on the semantic feature vector using a sliding window to determine an update boundary; A mapping module is configured to determine update information based on the update boundary and a preset time decay weight, and map the update information into a knowledge graph with a timestamp.

9. An electronic device, characterized in that: include: at least one processor; a memory for storing the at least one processor-executable instruction; The at least one processor is configured to execute the instructions to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform the method according to any one of claims 1 to 7.

Images