Scientific literature extraction task-oriented field tracing method and system
By constructing multi-level queries and using semantic similarity matching algorithms, the accuracy and speed issues of field tracing in scientific literature have been solved. This has enabled accurate and rapid tracing of hierarchical structures, thereby improving the user experience of scientific literature extraction tasks.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-08
- Publication Date
- 2026-05-12
Smart Images
Figure CN122019761A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of natural language processing and natural science, specifically relating to a field tracing method and system for scientific literature extraction tasks. Background Technology
[0002] Traditional field extraction algorithms are generally used for sequence labeling tasks, often employing rules or common sequence labeling algorithms, such as Hidden Markov Models (HMMs) and Conditional Random Fields (CRFs) in early traditional machine learning; and Recurrent Neural Networks (LSTMs) and Deep Learning Interpretation Models (BERTs) in traditional deep learning. These algorithms convert text into a sequence of tokens and predict the location of each token, making the position of the extracted field relatively easy to determine (i.e., the predicted token position). Since the advent of large-scale models, their ease of use and versatility have made them a superior choice for text extraction tasks. However, large-scale models treat the extraction of specific fields as a generation task, thus requiring additional rules or algorithms to determine the source of the extracted fields—that is, their position in the original text—to facilitate user verification and error correction. Therefore, tracing the source of extracted fields is necessary (tracing the source refers to finding the position of the extracted field in the original text).
[0003] Extraction from scientific literature often involves a hierarchical structure, such as a sample and its attributes. To ensure a good user experience and timely verification, it is necessary to ensure that fields at different levels do not deviate from each other, and the tracing location of a sample's sub-attribute should, as far as possible, include both the sample name and the name of the sub-attribute. However, current technology cannot complete tracing tasks in a timely and accurate manner. In addition, the inconsistent literal expressions of the same field (such as the name of a sample) within scientific literature, such as abbreviations, simplified expressions, expanded expressions, similar words, and synonyms, also bring certain difficulties to the tracing task. Therefore, there is an urgent need for a technological innovation that can complete field extraction tasks in a timely and accurate manner when dealing with both long and short fields. Summary of the Invention
[0004] To address the shortcomings of existing technologies, the purpose of this invention is to provide a field tracing method and system for scientific literature extraction tasks. For field content at different levels (e.g., a sample name and the name of a certain attribute of the sample), the method ensures that the sample name and the sample attribute name are not separated. For field content of different lengths (e.g., short sample names and long summary-type sample characteristics), short fields enable sentence-level tracing, while long fields enable paragraph-level tracing. For short fields, users can verify information promptly; for long fields, the method ensures that information is as complete as possible without loss.
[0005] The first aspect of the present invention: To achieve the above-mentioned objective, the present invention provides a field tracing method for scientific literature extraction tasks, comprising the following steps: (1) Perform page layout analysis on scientific documents to obtain structured data containing text content and its page position information; (2) Extract the fields to be traced by LLM. The fields are divided into primary keys and subkeys according to the hierarchy. The primary key is the identifier of the entity or object, and the subkey is the attribute attached to the primary key, which is used to describe the characteristics and status attribute information of the primary key. (3) Filter the fragments in the layout analysis file based on the characteristics of the primary key / subkey values; (4) Construct recall queries according to the hierarchy of primary key / subkey; (5) Construct a semantic similarity matching algorithm. The input is the query and several filtered segments, and the output is the segment with the highest similarity value to the query. (6) Output the coordinate region Bounding Box corresponding to the segment with the highest similarity value, and output the position information of the Bounding Box to locate and highlight the corresponding field content on the display interface.
[0006] Furthermore, step (3) of filtering fragments in the layout analysis file based on the characteristics of primary key / subkey values specifically includes the following steps: (3.1) Determine if there exists a fragment that fully contains the primary key / subkey value. If it exists, return all fragments that fully contain the primary key / subkey value and jump to step (3.4). If it does not exist, jump to step (3.2). (3.2) Extract the numeric fields [num1, num2, …] from the primary key / subkey values, and determine whether there is a fragment containing more than half of the numeric fields. If there is, return all fragments containing more than half of the numeric fields in the primary key / subkey values and jump to step (3.4). If there is no such fragment, jump to step (3.3). (3.3) Return to all segments and jump to step (3.4); (3.4) Output the filtered segments from steps (3.1), (3.2) and (3.3).
[0007] Specifically, in step (4), recall queries are constructed according to the hierarchy of primary key / subkey, including primary key recall query and subkey recall query; the constructed primary key recall query is: "[Primary key value]: The [primary key name] mentioned in the text contains [primary key value]"; the constructed subkey recall query is: "[Subkey value]: The [subkey name] of the [primary key value] in the text is [subkey value]".
[0008] Specifically, the primary key value is an entity identifier, which is the specific name / ID of the entity; the primary key name is the entity type; the subkey value is an attribute value, which is the specific content of the attribute; the subkey name is the attribute type.
[0009] Furthermore, the construction of the semantic similarity matching algorithm in step (5) specifically includes the following steps: (5.1) Obtain the input for the semantic similarity matching algorithm: query and filtered fragment chunks; (5.2) Determine the number of filtered fragment chunks Num. If Num > Q, where Q is in the range of [768, 1280], determine if there are any cached (fragment, embedding vector) pairs. If they exist, read the cache directly. If not, use the embedding model to calculate the embedding vectors of the chunks and cache the (fragment, embedding vector) pairs. Calculate the dot product between the vectorized query and the embedding vector of the fragment to obtain the semantic similarity between the query and the filtered fragments. Take the 256 most semantically similar fragments and jump to step (5.3) for execution. If Num ≤ Q, return all fragments and jump to step (5.3). (5.3) Use a sliding window mechanism to divide each initial fragment returned in step (5.2) into several sub-fragments, and establish a mapping table between each initial fragment and the sub-fragments; (5.4) Concatenate all sub-segments with the query, that is, concatenate the query in front of each sub-segment, and then input them into the reranker similarity calculation model in batches. According to the GPU memory capacity, the input items of the model are divided into several batches, and the similarity between all sub-segments and the query is output. According to step (5.3), establish a mapping table between the initial segment and the sub-segments and return the initial segment corresponding to the sub-segment with the highest similarity.
[0010] Specifically, in step (5.3), the window size of the sliding window is 64, the step size is 48, and the overlapping part of adjacent windows is 16.
[0011] The second aspect of the present invention provides a field tracing system for scientific literature extraction tasks, comprising a data storage unit, a fragment filtering unit, a semantic matching algorithm unit, and a result output and location unit; The data storage unit is used to store the layout parsing file of scientific literature, the fields to be traced, and cached file (fragment, embedding vector) pairs; The fragment filtering unit is used to filter redundant and invalid scientific literature fragments; The semantic matching algorithm unit is used to calculate the similarity between the query and the scientific literature fragment, and return the scientific literature fragment that is most similar to the query; The result output and positioning unit jumps to the position of the fragment content in the scientific literature based on the fragment output by the semantic matching algorithm unit and the corresponding Bounding Box information of the fragment, and highlights it.
[0012] A third aspect of the present invention provides a computing device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the field tracing method for scientific literature extraction tasks.
[0013] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the field tracing method for scientific literature extraction tasks.
[0014] Compared with the prior art, the beneficial effects of the present invention include at least the following: 1. By constructing multi-level queries, the problem of deviations in the extracted content (such as the name of a sample and the name of a certain attribute of the sample) at different levels can be solved during source tracing.
[0015] 2. By separating table text and paragraph text, as well as numerical text and non-numerical text, the matching method improves the accuracy of source tracing while shortening the source tracing time.
[0016] 3. To address the computational bottleneck of reranker under long text input (due to the length of the input text), a sliding window chunking strategy is introduced to significantly improve inference speed while ensuring source tracing accuracy. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 This is a flowchart of a field tracing method for scientific literature extraction tasks according to the present invention. Figure 2 This is a flowchart of the scientific literature fragment filtering and fragment matching process of this invention; Figure 3 This is a flowchart of the semantic similarity matching algorithm of the present invention. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of this invention.
[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of this invention.
[0021] The inventive concept of this invention is as follows: For the source tracing task after extracting fields from scientific literature, this invention addresses the deviation phenomenon that occurs between different levels of extracted content (e.g., a sample name and the name of a certain attribute of the sample) during source tracing by constructing multi-level queries. It improves source tracing accuracy while shortening tracing time by separating table text and paragraph text, and separating numerical text and non-numerical text. To address the computational bottleneck of reranker under long text input (O() value, which is the length of the input text), a sliding window chunk strategy is introduced, significantly improving inference speed while ensuring source tracing accuracy.
[0022] Based on the technical concept of this invention, such as Figure 1 As shown in the figure, the flowchart of a field tracing method for scientific literature extraction tasks provided in this embodiment includes the following steps: (1) Perform page layout parsing on scientific documents to obtain structured data containing text content and its page position information; specifically: obtain the page layout parsing file DocParse of the scientific documents, the specific format of which is: { 'sentences':[ {'content': 'sentence1', 'bbox': [float1 1 float1 2 float1 3 float1 4 float1 5 float1 6 float1 7 float1 8 ]}, {'content': 'sentence2', 'bbox': [float2 1 float2 2 float23 float2 4 float2 5 float2 6 float2 7 float2 8 ]}, {'content': 'sentence3', 'bbox': [float3 1 float3 2 float3 3 float3 4 float3 5 float3 6 float3 7 float3 8 ]}, … ], 'tables': [ {'content': 'table1', 'bbox': [float1 1 float1 2 float1 3 float1 4 float1 5 float1 6 float1 7 float1 8 ]}, {'content': 'table2', 'bbox': [float2 1 float2 2 float2 3 float2 4 float2 5 float2 6 float2 7 float2 8 ]}, … ] } sentence i and table i These represent sentence text and table text in the layout analysis, respectively. The bbox represents the position information of the corresponding text in the document, consisting of 8 values. Every 2 values represent the location of a point, and 4 points represent a rectangular coordinate area (Bounding Box). (2) Extract the fields to be traced (numeric or text string) through LLM. The fields are divided into primary keys and subkeys according to the hierarchy. The primary key is the identifier of the entity or object (e.g., the name of the mineral, the number of the experiment). The subkey is the attribute attached to the primary key, used to describe the characteristics, status and other attribute information of the primary key (e.g., the location of the mineral, the time of the experiment, etc.). Taking the second-level traceability field of geoscience literature as an example: {'keyField': {'mineral': 'xx gold mine'}, 'subfield': {'region':'xxx town', 'composition': 'xx gold, xx iron, …', …}}, where mineral is the primary key field name, region and composition are the subkey field names, and xx gold mine, xxx town, xx gold, xx iron are the field content to be traced. (3) Filter fragments in the layout analysis file based on the characteristics of primary key / subkey values, specifically as follows: Figure 2 As shown, the flowchart for scientific literature fragment filtering and fragment matching provided in the embodiment includes the following steps: (3.1) Determine if there exists a fragment that fully contains the primary key / subkey value. If it exists, return all fragments that fully contain the primary key / subkey value and jump to step (3.4). If it does not exist, jump to step (3.2). (3.2) Extract the numeric fields [num1, num2, …] from the primary key / subkey values, and determine whether there is a fragment containing more than half of the numeric fields. If there is, return all fragments containing more than half of the numeric fields in the primary key / subkey values and jump to step (3.4). If there is no such fragment, jump to step (3.3). (3.3) Return all segments and jump to step (3.4); This step is to ensure the recall rate of subsequent similarity calculation as much as possible. Only texts that fully contain or contain more than half of the numerical fields can be determined to contain the extracted text, thereby reducing the number of texts input to the downstream model and shortening the time. The immediacy of source tracing will bring a better user experience. If it is uncertain whether the content is contained, then return all chunks. Between speed and accuracy, accuracy is still the priority factor. (3.4) Output the filtered segments from steps (3.1), (3.2), and (3.3); (4) Construct a recall query for each field content according to the field hierarchy; construct recall queries according to the primary key / subkey hierarchy, including primary key recall query and subkey recall query; the constructed primary key recall query is: "[Primary key value]: The [primary key name] mentioned in the text contains [primary key value]"; the constructed subkey recall query is: "[Subkey value]: The [subkey name] of the [primary key value] in the text is [subkey value]"; the subkey value is an attribute value, which is the specific content of the attribute; the subkey name is the attribute type. Taking the geoscience literature level 2 source field as an example, the primary key recall query is: "xx gold mine [primary key value]: The mineral [primary key name] mentioned in the text contains xx gold mine [primary key value]"; the subkey recall query is: "xxx town [subkey value]: The region [subkey name] of the xx gold mine [primary key value] in the text is xxx town [subkey value]"; (5) Construct a semantic similarity matching algorithm. The input is a query and several segments, and the output is the segment with the highest similarity value to the query among the segments; specifically as follows: Figure 3 As shown, the semantic similarity matching algorithm flowchart provided in the embodiment includes the following steps: (5.1) Obtain the input of the semantic similarity matching algorithm: query and the fragments (chunks) filtered in step (3); (5.2) Determine the number of filtered chunks (Num). If Num > 1024, determine whether there are cached (chunk, embedding vector) pairs. If they exist, read the cache directly. If they do not exist, use the embedding model to calculate the embedded vector of the chunks, cache the (chunk, embedding vector) pairs, and use the vectorized query and the embedding vector of the chunk to calculate the dot product to obtain the semantic similarity between the query and the filtered chunks. Take the top 256 most semantically similar chunks and return them and jump to step (5.3). If Num ≤ 1024, return all chunks and jump to step (5.3). (5.3) The sliding window mechanism is used to divide each initial fragment returned in step (5.2) into several sub-fragments. The window size is 64 and the step size is 48 (i.e., the overlapping part of adjacent windows is 16). A mapping table between each initial fragment and the sub-fragments is established. (5.4) Concatenate all sub-segments with the query, that is, concatenate the query before each sub-segment, and then input them into the reranker similarity calculation model in batches. The batch size is set according to the GPU memory size (control the memory usage to around 90%). Output the similarity between all sub-segments and the query. Based on step (5.3), establish a mapping table between the initial segment and the sub-segments and return the initial segment corresponding to the sub-segment with the highest similarity (where the similarity value is the value directly output by the reranker model). (6) Output the Bounding Box corresponding to the segment with the highest similarity value, and output the position information of the Bounding Box to locate and highlight the corresponding field content on the display interface.
[0023] Based on the same inventive concept, the embodiment also provides a scientific literature field extraction system based on a large model, including a data storage unit, a fragment filtering unit, a semantic matching algorithm unit, and a result output and positioning unit; the data storage unit is used to store the layout parsing file of the scientific literature, the fields to be traced, and cached file (fragment, embedding vector) pairs; the fragment filtering unit is used to filter redundant and invalid scientific literature fragments; the semantic matching algorithm unit is used to calculate the similarity between the query and the scientific literature fragment, and return the scientific literature fragment most similar to the query; the result output and positioning unit jumps to the position of the fragment content in the scientific literature according to the fragment output by the semantic matching algorithm unit and the corresponding bounding box information of the fragment, and highlights it.
[0024] It should be noted that the field tracing method and system for scientific literature extraction tasks provided in the above embodiments are illustrated using the above-described division of functional modules. These functions can be assigned to different functional modules as needed, i.e., the internal structure of the terminal or server can be divided into different functional modules to complete all or part of the functions described above. Furthermore, the field tracing method and system embodiments for scientific literature extraction tasks provided in the above embodiments belong to the same concept. Their specific implementation process is detailed in the field tracing method embodiment for scientific literature extraction tasks, and will not be repeated here.
[0025] Based on the same inventive concept, the embodiment also provides a computing device. At the hardware level, in addition to a processor and memory, it also includes an internal bus, network interface, memory, and other hardware required for business operations. The memory is non-volatile memory. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to implement the field tracing method and system for scientific literature extraction tasks described above. Of course, besides software implementation, this invention does not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution entity of the following processing flow is not limited to individual logic units, but can also be hardware or logic devices.
[0026] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A field tracing method for scientific literature extraction tasks, characterized in that, Includes the following steps: (1) Perform page layout analysis on scientific documents to obtain structured data containing text content and its page position information; (2) Extract the fields to be traced by LLM. The fields are divided into primary keys and subkeys according to the hierarchy. The primary key is the identifier of the entity or object, and the subkey is the attribute attached to the primary key, which is used to describe the characteristics and status attribute information of the primary key. (3) Filter the fragments in the layout analysis file based on the characteristics of the primary key / subkey values; (4) Construct recall queries according to the hierarchy of primary key / subkey; (5) Construct a semantic similarity matching algorithm. The input is the query and several filtered segments, and the output is the segment with the highest similarity value to the query. (6) Output the coordinate region Bounding Box corresponding to the segment with the highest similarity value, and output the position information of the Bounding Box to locate and highlight the corresponding field content on the display interface.
2. The field tracing method for scientific literature extraction tasks according to claim 1, characterized in that, Step (3) involves filtering fragments in the layout analysis file based on the characteristics of primary key / subkey values, specifically including the following steps: (3.1) Determine if there exists a fragment that fully contains the primary key / subkey value. If it exists, return all fragments that fully contain the primary key / subkey value and jump to step (3.4). If it does not exist, jump to step (3.2). (3.2) Extract the numeric fields [num1, num2, …] from the primary key / subkey values, and determine whether there is a fragment containing more than half of the numeric fields. If there is, return all fragments containing more than half of the numeric fields in the primary key / subkey values and jump to step (3.4). If there is no such fragment, jump to step (3.3). (3.3) Return to all segments and jump to step (3.4); (3.4) Output the filtered segments from steps (3.1), (3.2) and (3.3).
3. The field tracing method for scientific literature extraction tasks according to claim 1, characterized in that, In step (4), recall queries are constructed according to the hierarchy of primary key / subkey, including primary key recall query and subkey recall query. The constructed primary key recall query is: "[Primary key value]: The [primary key name] mentioned in the text contains [primary key value]". The constructed subkey recall query is: "[Subkey value]: The [subkey name] of the [primary key value] in the text is [subkey value]".
4. The field tracing method for scientific literature extraction tasks according to claim 3, characterized in that, The primary key value is an entity identifier, which is the specific name / ID of the entity; the primary key name is the entity type; the subkey value is an attribute value, which is the specific content of the attribute; the subkey name is the attribute type.
5. The field tracing method for scientific literature extraction tasks according to claim 1, characterized in that, The semantic similarity matching algorithm constructed in step (5) specifically includes the following steps: (5.1) Obtain the input for the semantic similarity matching algorithm: query and filtered fragment chunks; (5.2) Determine the number of filtered fragment chunks Num. If Num > Q, where Q is in the range of [768, 1280], determine if there are any cached (fragment, embedding vector) pairs. If they exist, read the cache directly. If not, use the embedding model to calculate the embedding vectors of the chunks and cache the (fragment, embedding vector) pairs. Calculate the dot product between the vectorized query and the embedding vector of the fragment to obtain the semantic similarity between the query and the filtered fragments. Take the 256 most semantically similar fragments and jump to step (5.3) for execution. If Num ≤ Q, return all fragments and jump to step (5.3). (5.3) Use a sliding window mechanism to divide each initial fragment returned in step (5.2) into several sub-fragments, and establish a mapping table between each initial fragment and the sub-fragments; (5.4) Concatenate all sub-segments with the query, that is, concatenate the query in front of each sub-segment, and then input them into the reranker similarity calculation model in batches. According to the GPU memory capacity, the input items of the model are divided into several batches, and the similarity between all sub-segments and the query is output. According to step (5.3), establish a mapping table between the initial segment and the sub-segments and return the initial segment corresponding to the sub-segment with the highest similarity.
6. The field tracing method for scientific literature extraction tasks according to claim 5, characterized in that, In step (5.3), the sliding window has a window size of 64 and a step size of 48, meaning that the overlapping portion of adjacent windows is 16.
7. A field tracing system for scientific literature extraction tasks, characterized in that, It includes a data storage unit, a fragment filtering unit, a semantic matching algorithm unit, and a result output and location unit; The data storage unit is used to store the layout parsing file of scientific literature, the fields to be traced, and cached file (fragment, embedding vector) pairs; The fragment filtering unit is used to filter redundant and invalid scientific literature fragments; The semantic matching algorithm unit is used to calculate the similarity between the query and the scientific literature fragment, and return the scientific literature fragment that is most similar to the query; The result output and positioning unit jumps to the position of the fragment content in the scientific literature based on the fragment output by the semantic matching algorithm unit and the corresponding BoundingBox information of the fragment, and highlights it.
8. A computing device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the field tracing method for scientific literature extraction tasks as described in any one of claims 1-6.
9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the field tracing method for scientific literature extraction tasks as described in any one of claims 1 to 6.