Article parsing method and device

Abstract

The present invention discloses an article parsing method and device, wherein the article parsing method comprises the following steps: (1) obtaining user information; (2) obtaining the original text and parsing the original text and metadata; (3) preprocessing the text segment; (4) analyzing the original text dependency; (5) generating expandable groups; (6) encapsulating all original data and generated data. The article is parsed using the present invention, and the parsed results can be used in classroom teaching, after-school reading, paper reading, leisure, office and other scenarios, and have the effects of reducing reading cognitive load, improving reading effect, improving reading efficiency, and improving reading sensory experience, thereby improving physical and mental experience and actual effect.

Description

Technical Field

[0001] The present invention belongs to the field of text analysis, and in particular relates to an article parsing method and device. Background Art

[0002] In order to improve the reading experience and effect in classroom teaching, after-class reading, paper reading, leisure, office and other scenarios, people will choose to use various tools including document editing tools and document readers for reading. Traditionally, document editing tools and document reading tools will obtain raw data, parse it, and then render the media.

[0003] However, this simple analysis method fails to take into account the user's reading speed requirements and the need to grasp the key points, which leads to widespread inconsistent reading time efficiency, repeated reading, and other situations that reduce the experience.

[0004] In recent years, with the development of natural language technology, some technologies have emerged that use large language models to analyze text during parsing, thereby attempting to solve this problem.

[0005] A Chinese patent document with publication number CN116797417A discloses an intelligent assistance system based on a large language model, including: a vocabulary generation module, an input sequence generation module, and an output result display module. It uses advanced natural language processing technology and machine learning algorithms to achieve in-depth understanding of course content and personalized tutoring and learning.

[0006] A Chinese patent document with publication number CN117112779A discloses a local document summarization method based on a large model; the method includes: uploading documents to be summarized, selecting a langchain loading tool to load documents, determining the size of documents, performing document summarization in a stuff mode, and performing document summarization in a map-reduce mode.

[0007] However, the parsing granularity of existing methods is relatively fixed. They are either based on a fixed granularity such as the full text, or they are based on user requests to parse text segments of a specified length. During this process, users are required to input or select the text to be parsed, which results in the possibility that the parsed content is useless to readers, or the reading process is interrupted by manual operations. In addition, existing solutions cannot simultaneously consider the diversity of reading needs and comprehension abilities of multiple groups of people, resulting in the parsed content not being able to balance the needs of fast reading and detailed reading at the same time. Finally, existing solutions fail to simultaneously consider and parse the two dimensions of text granularity and text structure, resulting in the parsing results not being able to well present the logical relationship between sentences, paragraphs, sections, etc., thereby reducing the learning experience of learning text segments and corresponding logic.

[0008] Therefore, the existing article parsing methods, data formats and systems are not always sufficient to meet the complex needs of users in classroom teaching, after-class reading, paper reading, leisure, office and other scenarios. Summary of the invention

[0009] The present invention provides an article analysis method and device, the analysis results of which are suitable for classroom teaching, after-class reading, paper reading, leisure, office and other scenarios, and have the effects of reducing reading cognitive load, improving reading effect, reading efficiency, and reading sensory experience, thereby improving physical and mental experience and actual effect.

[0010] An article parsing method comprises the following steps:

[0011] (1) Obtain user data, obtain reference reading goals from a reading goal database, and calculate actual reading goals based on the obtained user data and the reference reading goals;

[0012] (2) Obtain multiple articles related to the user, filter the obtained articles based on the actual reading goals, and disassemble the original manuscript information and metadata information of the remaining articles for subsequent processing;

[0013] (3) For the articles filtered in step (2), based on the actual reading target, use the natural language model to merge or split the article paragraphs whose length does not meet the granularity threshold; then, for the processed article paragraphs, generate a summary of the paragraph itself based on the given window size and related metadata;

[0014] (4) Identify the sentence splitting points in the paragraph and divide the text into sentence lists; according to the sentence lists, group the consecutive sentences based on the inter-sentence logic and analyze the dependencies between the groups;

[0015] (5) For groups with high information density, set a higher granularity threshold and repeat steps (3) and (4); based on the given window size, extract the group and the context of adjacent groups, and combine the group-related metadata to generate summaries for the sentences in the group to form logical blocks. At the same time, use the natural language model proxy to vote on whether to add lubrication blocks to ensure logical coherence;

[0016] (6) Encapsulating all the original data and generated data of the article after filtering in step (2), the data format includes the original manuscript information and metadata information of the article itself, as well as the parsed logical blocks, paragraph summaries and lubrication block information;

[0017] When the user needs to display an article, the parsed results are displayed on the user's device based on the encapsulated data, and the system responds to the user's interaction.

[0018] In step (1), the user data includes reading preferences, reading experience, knowledge reserves, etc. The user data is obtained as follows:

[0019] Having a user complete one or more readings of an article during which user data is measured or calculated using a portable device containing sensors; and / or,

[0020] Collect or calculate user data through inquiries, questionnaires, or data mining.

[0021] In step (1), the reading goal database includes at least one reference reading goal, each reference reading goal includes at least one indicator, and each indicator includes at least one value related to a time range;

[0022] According to the obtained user data and the reference reading goal, the actual reading goal is calculated using the large language model LLM. The actual reading goal includes at least one indicator, and each indicator includes at least one value related to a certain time range.

[0023] In step (2), manuscript information includes text content, images, formulas, endnotes, footnotes, codes and tables, etc.; metadata information includes the title, author, abstract, date, publication location and publication of the article, etc.

[0024] In step (3), the window size refers to the number of paragraphs processed by the large language model LLM each time.

[0025] In step (4), a segmentation algorithm is used to identify the sentence segmentation points in the paragraph and segment the text into a sentence list. The specific process is as follows:

[0026] S4.1 Initialize all parameters of the partitioning algorithm and an initial partitioning method;

[0027] S4.2 Analyze the possible defects of this division method and provide several new division methods;

[0028] S4.3 For each newly generated division method, evaluate the rationality of the generated new division method and give a score that includes a diversity reward;

[0029] S4.4 Calculate the partition value for each newly generated partition;

[0030] S4.5 iteratively, sequentially update the values ​​of all solutions of the previous analysis level of each newly generated sentence partition list;

[0031] S4.6 selecting a next partitioning method to be optimized;

[0032] S4.7 update partitioning algorithm parameters;

[0033] S4.8 Repeat steps S4.2 to S4.8 until the algorithm ends and the optimal division method is obtained.

[0034] In step (4), consecutive sentences are grouped based on inter-sentence logic using a top-down analysis method, a bottom-up analysis method, or a direct analysis method.

[0035] The present invention also provides an article parsing device, comprising a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the above-mentioned article parsing method.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] 1. The method of the present invention does not parse the article based on a fixed granularity such as the full text, but dynamically creates possible parsing results based on the content and logic of the article, which makes it more likely that the article presented by the software will adapt to the user's cognitive level, thereby providing more reading flexibility and efficiency.

[0038] 2. The analysis results generated by the method of the present invention have hierarchical and logical grouping. Therefore, in the relevant software design process, users do not need to select boxes to obtain further explanations of the text, which makes the user's reading experience more continuous.

[0039] 3. The method of the present invention can simultaneously consider and analyze the two dimensions of text granularity and text structure, so that the analysis results can well present the logical relationship between sentences, paragraphs, sections, etc., thereby improving the learning experience of learning text paragraphs and corresponding logic.

[0040] 4. Experiments show that the article analysis results generated by the method of the present invention can be used in reading-related software. Compared with traditional methods, this method has the advantages of reducing reading cognitive load, improving reading effect, improving reading efficiency, improving reading sensory experience, and thus improving physical and mental experience and actual effect.

[0041] 5. The method of the present invention is easy to deploy on different software and hardware platforms, and is applicable not only to academic scenarios such as classroom teaching, after-class reading, and paper reading, but also to other leisure, office and other scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 The present invention is a flowchart of an article parsing method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0043] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be pointed out that the embodiments described below are intended to facilitate the understanding of the present invention and do not have any limiting effect on the present invention.

[0044] like Figure 1 As shown, an article parsing method comprises the following steps:

[0045] Step S1: Obtain user information and reading goals.

[0046] Get one or more reference reading goals from the reading goal database.

[0047] The reading goal database contains one or more reference reading goals for a certain reading activity. Each reference reading goal contains at least one measurable or calculable indicator, and each indicator contains at least one value related to reading. For example: 250 words per minute, 80% correct rate for multiple-choice questions. The actual reading goal is calculated based on user data (such as reading habits, knowledge background, past performance, etc.) and the obtained reference reading goals. The actual reading goal contains at least one measurable or calculable indicator, and each indicator contains at least one value related to a certain time range.

[0048] There are at least three ways to initialize data acquisition. One possible implementation method is to let the user complete one or more reading activities without intervention, during which a portable device or reader containing sensors is used to measure physiological, psychological, reading characteristics or other information. When this method is used during the first reading, the most accurate data of the user without intervention can be obtained, which is helpful for calculating reading goals and comparing intervention effects without interference.

[0049] Another way is to use a portable device or reader containing sensors to measure physiological, psychological, reading characteristics or other information while the user is reading, with the intervention of the reader and the algorithm. When this method is used during reading, it will be possible to obtain more real-time data of the user during the reading process, which will help improve the matching degree between the parsed content and the user.

[0050] Another way is to collect user data by inquiry, questionnaire, data mining, etc. The above three implementation methods can be used selectively or in combination, and the embodiment of the present invention does not make specific limitations here.

[0051] In addition, the reading goal database contains one or more reference reading goals, which are used to provide a benchmark for actual reading goals. For example, when a user is reading a certain paper, the database returns two indicators, namely, the measurable indicator of reading speed and the computable indicator of expansion level. In an embodiment of the present invention, due to the diversity of physical and mental exercise and fitness activities that can be performed by users, and for the purpose of optimizing the user experience, the reference reading goals can also be simply edited without fully adopting the results returned by the database. In this case, the reference reading goals obtained may be completely consistent with the results returned by the database, or they may be different, and the embodiment of the present invention does not make specific limitations here.

[0052] Since the reference reading goal does not match the actual physical and mental conditions of the user, it is necessary to calculate the actual reading goal with the help of the data and the reference reading goal, so as to enhance reading comprehension, improve reading efficiency, and enhance the sensory experience of reading, thereby achieving the purpose of improving physical and mental experience and effect. When data or reference reading goals are missing, simple editing operations such as swapping and interpolation can be used to complete the data based on the existing information to ensure the normal operation of the algorithm.

[0053] Step S2: Acquire and preprocess article data.

[0054] First, a list of multiple articles related to the user is obtained, and source files are obtained from the Internet, an external device or a local device, and the files are optionally filtered to finally obtain a series of articles.

[0055] Among them, file filtering refers to removing some inappropriate articles from the obtained articles based on the actual reading goal, so as to reduce the amount of calculation and improve the reading experience. For example, when the value of the computable indicator of the article field is only "healthy", the preprocessing process will remove articles that are not "healthy".

[0056] When processing the source manuscript of an article, the first thing to do is to systematically disassemble the original manuscript information and metadata information of the article. This process involves a comprehensive analysis of the original manuscript content and the collection and classification of relevant metadata.

[0057] The decomposition of manuscript information includes the detailed decomposition of various information such as text content, pictures, formulas, endnotes, footnotes, codes, tables, etc. For formats that can be directly parsed, such as HTML, direct parsing can be performed directly with the help of HTML XPATH expressions. For formats that cannot be directly parsed, such as PDF, optical character recognition (OCR) technology is used to identify and parse them and convert them into editable forms. During the parsing process, the title, part ID (if any) and text of each part are recursively extracted, and their text and sub-parts are extracted at the same time. The extraction of endnotes and footnotes depends on document structure analysis to ensure that these annotation information can be correctly identified and separated. The processing of heterogeneous data such as code and tables needs to be further converted into structured data to facilitate subsequent analysis and processing.

[0058] The metadata is then broken down, which includes but is not limited to the title, author, abstract, date, publication location, publication journal, and other information of the article. First, the title and abstract are directly extracted from the front of the document, and then the language is standardized. The author information is extracted from the beginning or end of the article using pattern matching and matched with the database to ensure the accuracy of the author information. The date, publication location, and publication journal information are usually contained in the header or footer of the article, and regular expression technology is used for accurate extraction and identification. In addition, other metadata information, such as keywords, references, etc., are also extracted through parsing rules and formatted for subsequent data storage and retrieval.

[0059] Step S3: article paragraph preprocessing.

[0060] Based on the actual reading goal, the article paragraphs are selectively re-cut and merged, and the natural language model is used to merge or split the paragraphs whose length does not meet the granularity threshold to ensure that the information content and length of each paragraph are moderate. Then, for the processed paragraphs, a summary of the paragraph itself is generated based on the given window size and related metadata.

[0061] In this embodiment, all existing paragraphs are first traversed in sequence. Greater than a paragraph length threshold When the paragraph and context are sent to the natural language model agent, the potential cutting points of the paragraph are located and cut based on the actual reading goals and paragraph-related indicators.

[0062] The context may include but is not limited to: previous and next paragraphs, chapter titles, article metadata, etc.

[0063] Among them, paragraph-related indicators may include but are not limited to paragraph reading difficulty score, paragraph length, number of new words, etc.

[0064] Optionally, in order to ensure the accuracy of the segmentation site, a method such as ToT voting method can be further used to calculate the adoption probability of the segmentation site. Natural language agents are used to estimate the objective reading difficulty difference before and after segmentation, the difference in reading difficulty judged by readers, and the difference in the probability of achieving the reading goal. When the scores of multiple natural language agents for a certain position exceed the score threshold, The number of scores exceeds the threshold , then the paragraph is segmented at the specified position. There are no restrictions on the number of, type of natural language model used, evaluation dimensions, prompts used, and thinking framework.

[0065] Then, adjacent article paragraphs are selectively merged. In this embodiment, all existing paragraphs are first traversed in sequence. Less than a paragraph length threshold When reading a paragraph, the context is sent to the natural language model agent, and based on the actual reading goal and paragraph-related indicators, it is analyzed whether the paragraph can be merged with the adjacent paragraph, and the merger is optional.

[0066] Optionally, in order to reduce the amount of calculation, the paragraphs generated by the previous cutting are excluded from the merging range.

[0067] Optionally, in order to ensure that the merging is conducive to reading, the merging probability can be further calculated by using methods such as ToT voting method. Natural language agents estimate the objective reading difficulty difference between the paragraphs before and after the merger, the difference in the reader's subjective reading difficulty, and the difference in the probability of achieving the reading goal. When the scores of multiple natural language agents exceed the score threshold for a certain merger point, The number of scores exceeds the threshold , then the two paragraphs are merged at the merging point. There are no restrictions on the number of, type of natural language model used, evaluation dimensions, prompts used, and thinking framework.

[0068] Subsequently, for the processed paragraphs, a summary is generated for the paragraphs themselves based on the given window size and related metadata.

[0069] In this embodiment, the paragraph and the context are sent to the natural language model agent, and the paragraph is summarized based on the actual reading goal and paragraph-related indicators.

[0070] Optionally, an iterative approach is used to optimize the summary, or the output results of multiple natural language models are aggregated.

[0071] Step S4: Analyze article text dependencies.

[0072] Identify the sentence splitting points in the paragraph and split the text into sentence lists. Then, according to the sentence lists, group the consecutive sentences based on the inter-sentence logic and analyze the dependencies between the groups.

[0073] S4.1 is first initialized. For a given paragraph, the paragraph is sent to the natural language agent, which is required to output an initial segmented sentence list represented by a computer-recognizable string list. , where the first 0 is the analysis level and the second 0 is the improvement level. The corresponding value Set to initial value . Set the maximum number of iterations to The target value is and maximum level depth . Select As the current best solution. Initialization parameters , , is the initial parameter value.

[0074] S4.2 Then, this sentence list and the segmentation metadata are sent to another natural language agent, which is asked to analyze the possible defects of this segmentation method and provide n new methods.

[0075] Among them, the divided metadata includes complete sentence lists, actual reading goals, paragraph summaries, hierarchical requirements, metadata and other data.

[0076] When the given sentence segmentation list is identical to an existing sentence segmentation list, the newly generated solution is deleted.

[0077] To prevent over-subdivision, the maximum number of sub-schemes for each scheme is set. .

[0078] S4.3 For each newly generated sentence, the partition list , ask its other natural language agent to evaluate the rationality of the new division method and give a score, record . Introducing diversity rewards , based on the average difference between this scheme and other schemes:

[0079] ;

[0080] in, is the weight coefficient of the diversity reward.

[0081] In order to prevent the agent from giving too high or too low scores to interfere with subsequent score calculations, When it is greater than the specified value, the part of the score greater than the specified value will be scaled proportionally.

[0082] Optionally, to ensure that the score given by the agent is accurate, repeated sampling is performed using different parameters (such as temperature) and the average is taken as the score.

[0083] S4.4 Divide each newly generated sentence into a list Calculating Value , the calculation method is as follows:

[0084] ;

[0085] in yes The set of all scoring samples of is the sample size, is a weight parameter between 0 and 1 for the current iteration round t, which is used to balance the influence of the minimum and average values. When choosing a value, if you want to consider the potential of the current solution during the analysis process, you will increase If you focus more on improving past average performance, you will reduce The value of .

[0086] S4.5 Iteratively, for each newly generated sentence, divide the list The values ​​of all solutions at the previous analysis level are updated in sequence, and the calculation method is as follows:

[0087] ;

[0088] where i decreases from x-1 to 0, It is a weight parameter between 0 and 1 for the current iteration, which is used to balance the influence of the node's own V value and the maximum V value of its child nodes. If you want to improve the value of the previous solution to rely more on the evaluation of the solution itself, you will increase the value of β; if you want to consider the best potential for improvement more, you will decrease The value of .

[0089] S4.6 Select the next optimized solution;

[0090] ;

[0091] Where C is the candidate solution set, N( ) is the number of times the solution is evaluated, c is the exploration coefficient, and ε is a constant close to 0. ParentPlan(a) represents the parent solution that generates solution a.

[0092] S4.7 Update parameters, where:

[0093] ;

[0094] ;

[0095] ;

[0096] in , , is the adjustment step size, is the average V value of the tth iteration, is the number of unexplored solutions in round t, is the total number of solutions in round t.

[0097] S4.8 Repeat S4.2 to S4.8 until the maximum number of iterations is reached or maximum level depth , or no new optimization results are generated. The solution with the largest V value is taken as the partitioning solution.

[0098] The logical grouping between sentences can adopt the top-down analysis method, the bottom-up analysis method, and the direct analysis method. In this embodiment, the logical grouping between sentences adopts the top-down analysis method and iterates. For a given sentence list, the natural language model agent will receive data such as the complete sentence list, actual reading goals, paragraph summaries, hierarchical requirements, metadata, etc. Subsequently, the natural language model agent will group sentences with the same logical hierarchy into the same group. The agent repeats the above process until the agent outputs a mark to stop processing and outputs the hierarchical processing results in a computer-recognizable data format.

[0099] Step S5: Expandable group generation.

[0100] For groups with high information density, the previous processing steps are repeated at a higher granularity to obtain more detailed information. Subsequently, the group and the adjacent group context are intercepted based on the given window size, and combined with the group-related metadata, the sentences in the group are summarized, and the natural language model proxy is used to vote on whether to add lubrication blocks to ensure logical coherence.

[0101] First, traverse all groups. For each group, if the number of words or information density exceeds the threshold , the sentences in the group are treated as a paragraph, and S3 and S4 are repeated with a higher granularity. It should be noted that the content before and after the repetition will be stored separately. The processed data can be used for group expansion.

[0102] Then, we traverse all the groups, and for each group, we intercept the given group and the adjacent group context based on the given window size, and combine the group-related metadata to generate a summary of the sentences in the group. In order to ensure the logical coherence between groups, we adopt the strategy of "processing as many groups as possible at one time" for the groups.

[0103] Then, optionally, in order to ensure that the grouped summary is easy to read, a method such as ToT can be further used to add lubrication blocks (for example: first, but, for example...), or to improve the coherence between blocks. A natural language agent determines whether lubrication blocks need to be added between groups in a traversal manner, and whether the summaries of two adjacent groups need to be rewritten at the same time. The embodiment of the present invention does not limit the traversal method, the number of traversals, the type of natural language model used, the evaluation dimension, the prompts used, and the thinking framework.

[0104] Step S6: Data packaging and output processing results.

[0105] All raw data and generated data are encapsulated. The data format contains the article itself and the parsed structure, such as logical blocks, paragraph summaries, and lubrication block information. Each level includes title, text content, related metadata, sub-level links, and corresponding semantic tags. The data format supports dynamic content expansion, such as interactive annotations and navigation elements. The underlying format of data encapsulation can be structured storage or unstructured storage, which is not limited in the embodiments of the present invention. Finally, when the user needs to display the article, the parsed results are displayed on the user's device based on the encapsulated data, and the user's interaction is responded to.

[0106] Based on the same inventive principle, an embodiment of the present invention further provides an article parsing device, including a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement the article parsing method of the above embodiment.

[0107] To verify the effect of the present invention, the present invention conducted the following experiment using the analysis of academic papers as an example. Throughout the study, participants were asked to complete two independent reading tasks using two different reading tools. One of the tools was a common article reader with a chapter summary function (software A), and the other was a customized article reader (software B) that visualized the data analyzed by the present invention.

[0108] Twenty-four participants were recruited for this study, ranging in age from 20 to 23 years old (mean age = 21.71 years, standard deviation = 1.00). This group consisted of 13 undergraduates and 11 graduate students. Before the start of the study, participants submitted informed consent and completed a pretest questionnaire of basic information. All personal information collected was processed anonymously, strictly confidential, and used only for research purposes. During the study, participants performed two reading tasks using two instruments. To reduce possible sequence effects, the pairing of papers and conditions was balanced. The study used a mixture of quantitative and qualitative methods. Quantitative methods included multiple-choice questions, mid-test questionnaires, and post-test questionnaires. Qualitative methods included eye tracking and semi-structured interviews. After each reading task, participants were required to complete a questionnaire designed to assess their reading strategies, comprehension, and emotional experience in the given experimental condition. After completing both reading tasks, participants were required to complete a satisfaction scale.

[0109] As shown in Table 1, through data analysis, we obtained the following insights into the potential of the parsing method in improving the reading experience:

[0110] Scale analysis of self-reported reading strategies showed that participants' mean scores were significantly higher when using Software B (M=67.88, SD=23.59) than when using Software A (M=43.75, SD=23.95) (P=0.001). This significant difference suggests that the parsing method has a significant advantage in improving reading strategies. In addition, participants rated the features of Software B highly in the post-test questionnaire.

[0111] The study revealed significant benefits of using software B in improving reading coherence through eye tracking and coding analysis of video recordings. Taking P2 as an example, when P2 used software B, the eye movement path showed an F-type reading pattern. The F-type reading pattern focuses on reading the summary of main ideas and logical keywords vertically, and occasionally reading supplementary content horizontally. This pattern is characterized by more coherent and stable eye movement paths, with less frequent gaze point jumps. In contrast, when P2 used software A, the eye movement trajectory showed a clear scattered pattern. This pattern involved frequent focus shifts and often skipping large sections of text. Similar trends were observed in other participants, indicating the general applicability of the described parsing method across different users.

[0112] In the semi-structured interviews about reading strategies, improvements in reading strategies, clearer reading logic, improved reading orientation, and better reading comprehension were observed. Many participants reported that their reading strategies changed when using Software B. For example, some participants adopted the following approach: “I read the title and subtitle first, and then look at the key sentences in the title. I may not read the subsequent content in depth. However, if some sentences are not clear enough, I will choose to read the subsequent detailed description.” (P2). As described by P9, this approach not only saved time but also helped them quickly grasp the core content of the article. Some participants pointed out that Software A could not clearly and quickly understand the logical structure of the article. “It just completed the overview of each section, and the logical relationship needs to be sorted out slowly.” (P7). Eleven participants emphasized the significant effect of Software B in helping them understand the logical structure of the article. Some participants praised the role of the system parsing results in enhancing their reading orientation experience. They pointed out that with Software B, they could clearly identify the current reading section. For example, “If I forget the previous content, I can quickly recall what I have read by looking back at the subtitle or navigation bar.” (P7). The participants who used software B for reading had an average accuracy rate of 77.88% (SD = 20.64%) in multiple-choice questions, which was significantly higher than the participants who used software A (mean = 59.77%, SD = 26.60%), with a P value of 0.017. This shows the effectiveness of the parsing method in improving reading comprehension.

[0113] In addition, the results of the self-reported comprehension scale further supported the advantages of the described parsing method. Participants' comprehension scores were significantly higher when using Software B (mean = 68.92, SD = 13.17) compared to Software A (mean = 52.08, SD = 22.01), with a P value of 0.005. This indicates that participants believed that they understood the content better when reading with Software B. In addition, in the post-test questionnaire, participants gave the paragraph semantic segmentation function of Software B a CSAT value of 95.83%, further confirming the significant benefits of Software B in enhancing reading comprehension.

[0114] In the semi-structured interviews about comprehension, we also analyzed participants’ performance in grasping the content and article structure. This included an improvement in the depth of content understanding, a clear grasp of the article structure, and an enhanced emotional experience. The above results show that users not only feel more satisfied and comfortable with the granularity, and have a variety of reading methods, but are also more inclined to recommend the software to others.

[0115] Table 1

[0116]

[0117] The embodiments described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for analyzing an article, characterized in that: The following steps are involved: (1) Obtain user data, obtain reference reading goals from a reading goal database, and calculate actual reading goals based on the obtained user data and the reference reading goals; (2) Obtain multiple articles related to the user, filter the obtained articles based on the actual reading goals, and disassemble the original manuscript information and metadata information of the remaining articles for subsequent processing; (3) For the articles filtered in step (2), based on the actual reading target, use the natural language model to merge or split the article paragraphs whose length does not meet the granularity threshold; Then, for the processed article paragraphs, a summary is generated for the paragraphs themselves based on the given window size and related metadata; (4) Identify the sentence splitting points in the paragraph and divide the text into sentence lists; according to the sentence lists, group the consecutive sentences based on the inter-sentence logic and analyze the dependencies between the groups; (5) For groups with high information density, set a higher granularity threshold and repeat steps (3) and (4); based on the given window size, extract the group and the context of adjacent groups, and combine the group-related metadata to generate summaries for the sentences in the group to form logical blocks. At the same time, use the natural language model proxy to vote on whether to add lubrication blocks to ensure logical coherence; (6) Encapsulating all the original data and generated data of the article after filtering in step (2), the data format includes the original manuscript information and metadata information of the article itself, as well as the parsed logical blocks, paragraph summaries and lubrication block information; When the user needs to display an article, the parsed results are displayed on the user's device based on the encapsulated data, and the system responds to the user's interaction.

2. The article parsing method according to claim 1, characterized in that: In step (1), the user data includes reading preferences, reading experience, and knowledge reserves. The user data is obtained as follows: Having a user complete one or more readings of an article during which user data is measured or calculated using a portable device containing sensors; and / or, Collect or calculate user data through inquiries, questionnaires, or data mining.

3. The article parsing method according to claim 1, characterized in that: In step (1), the reading goal database includes at least one reference reading goal, each reference reading goal includes at least one indicator, and each indicator includes at least one value related to a time range; According to the obtained user data and the reference reading goal, the actual reading goal is calculated using the large language model LLM. The actual reading goal includes at least one indicator, and each indicator includes at least one value related to a certain time range.

4. The article parsing method according to claim 1, characterized in that: In step (2), the manuscript information includes text content, pictures, formulas, endnotes, footnotes, codes and tables.

5. The article parsing method according to claim 1, characterized in that: In step (2), the metadata information includes the title, author, abstract, date, publication location and publication of the article.

6. The article parsing method according to claim 1, characterized in that: In step (4), a segmentation algorithm is used to identify the sentence segmentation points in the paragraph and segment the text into a sentence list. The specific process is as follows: S4.1 Initialize all parameters of the partitioning algorithm and an initial partitioning method; S4.2 Analyze the defects of this division method and provide several new division methods; S4.3 For each newly generated division method, evaluate the rationality of the generated new division method and give a score that includes a diversity reward; S4.4 Calculate the partition value for each newly generated partition; S4.5 iteratively, sequentially update the values ​​of all solutions of the previous analysis level of each newly generated sentence partition list; S4.6 selecting a next partitioning method to be optimized; S4.7 update partitioning algorithm parameters; S4.8 Repeat steps S4.2 to S4.8 until the algorithm ends and the optimal division method is obtained.

7. The article parsing method according to claim 1, characterized in that: In step (4), consecutive sentences are grouped based on inter-sentence logic using a top-down analysis method, a bottom-up analysis method, or a direct analysis method.

8. An article parsing device, characterized in that: The invention comprises a memory and one or more processors, wherein the memory stores executable codes, and when the one or more processors execute the executable codes, the article parsing method according to any one of claims 1 to 7 is implemented.

Images