Method and system for generating a description from an academic table and selected content thereof

Abstract

The application provides a method and system for generating description according to academic forms and selected contents thereof, comprising: analyzing and extracting forms in PDF academic literature, form background knowledge and description text related to form content, and performing matching; constructing a domain knowledge base based on the form background knowledge, and preprocessing the form background knowledge and the description text related to the form content, and performing serialization processing on the form to obtain serialized form data; splicing information in a preset form that a user is interested in and the form using a preset template, and constructing serialized form data that the user is interested in; obtaining key background information based on the serialized form data that the user is interested in and the domain knowledge base, and generating a serialized string based on the obtained key background information and the serialized form data that the user is interested in; training a text generation model using the serialized string and the description text related to the form content, and obtaining the trained text generation model; and predicting and generating form-related description using the trained text generation model.

Description

Technical Field

[0001] This invention relates to the field of natural language processing, specifically to a method and system for generating descriptions based on academic tables and their selected content. More specifically, it relates to a method and system for extracting paper tables based on the TSDAE sentence vector generation model, the BART-base text generation model, and the SCI-Spacy model, and for generating descriptions based on content of interest to the user, in order to better improve research efficiency and allow readers to quickly understand the content of the tables. Background Technology

[0002] Text sentence vector generation is a computational technique in natural language processing used to process and encode sentence text. First, word embedding is performed on each word in the sentence, embedding a high-dimensional space containing the number of words into a much lower-dimensional continuous vector space. Then, the sentence vector is calculated based on the word vector representations of the entire sentence, obtaining the vector representation of the whole sentence. After encoding the entire sentence into sentence vectors, it can be used to calculate the semantic similarity between sentences and improve the training speed of language models. This invention uses and trains a TSDAE model to encode text and calculate the semantic similarity between background knowledge and table content, obtaining the most similar background knowledge statement corresponding to the table (mainly based on a Transformer-based decoding-encoding model network).

[0003] In addition, the text generation task in this invention essentially involves building a BART-based Natural Language Generation (NLG) model. Natural Language Generation (NLG) is a part of Natural Language Processing, generating natural language from knowledge bases or logical forms and their representations. For the table description generation task in this invention, table content and relevant background knowledge are concatenated into a serialized statement using a template, which is then input into the natural language generation model to generate the relevant table description.

[0004] For text processing and word segmentation in this invention, the SCI-Spacy model from the field of natural language processing was used to split and reassemble the text according to rules. Sci-Spacy, released in 2019 by the Allen Institute for Artificial Intelligence, is a Python toolkit for natural language processing in the biomedical field based on Spacy. It provides various functions including word segmentation, sentence segmentation, proper noun tagging, and named entity recognition, as well as multiple corresponding models based on biomedical corpora such as CRAFT, JNLPBA, BC5CDR, and BIONLP13CG. Its models include lightweight, medium, and heavyweight models that implement the complete spaCy natural language processing pipeline, and corpus-specific named entity recognition models. It is one of the latest and most effective tools for natural language processing in the specific academic field of biomedicine. In this invention, the heavyweight pipeline model en_core_sci_lg from the Sci-Spacy tool is introduced to perform coarse-grained pre-segmentation of the text. Since the development of this model is based on biomedical corpora, secondary segmentation and reassembly of the segmented fragments are required after preprocessing the text. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for generating descriptions based on academic tables and their selected content.

[0006] A method for generating a description based on an academic table and its selected content, provided by the present invention, includes:

[0007] Step S1: Parse and extract tables, table background information, and descriptive text related to table content from PDF academic documents, and match the tables, table background information, and descriptive text related to table content.

[0008] Step S2: Construct a domain knowledge base based on the table background knowledge, preprocess the table background knowledge and descriptive text related to the table content, and serialize the table to obtain serialized table data;

[0009] Step S3: Use a preset template to combine the information and tables in the preset tables that the user is interested in, and construct the serialized table data that the user is interested in;

[0010] Step S4: Obtain key background information based on the serialized table data of interest and the domain knowledge base, and generate a serialized string based on the obtained key background information and the serialized table data of interest;

[0011] Step S5: Train the text generation model using the serialized string and descriptive text related to the table content to obtain the trained text generation model;

[0012] Step S6: Use the trained text generation model to predict and generate table-related descriptions.

[0013] Preferably, step S1 employs:

[0014] Step S1.1: Use PDFFigure2 to parse PDF academic documents and obtain tables, background information on tables, and descriptive text related to table content.

[0015] Step S1.2: Use a heuristic algorithm to filter and match the tables, background information of the tables, and descriptive text related to the table content in the parsed PDF academic documents.

[0016] Preferably, step S1.1 involves: acquiring PDF academic documents, using PDFFigure2 to parse the PDF, and obtaining JSON data of relevant tables and PDF text; parsing the JSON data to obtain descriptive text and background information related to the table content.

[0017] Preferably, step S2 employs:

[0018] Step S2.1: Use the SCI-Spacy model to segment the descriptive text related to the table content and the background knowledge of the table into sentences;

[0019] Step S2.2: Clean up data that is irrelevant to the table based on the text after sentence segmentation.

[0020] Preferably, step S4 employs:

[0021] Step S4.1: Construct the TSDAE model and train the TSDAE model using the serialized tabular data of interest and the domain knowledge base to obtain the trained TADAE model;

[0022] Step S4.2: Based on the trained TSDAE model, extract key background information from the serialized tabular data of interest and the domain knowledge base;

[0023] Step S4.3: Use a preset template to concatenate the serialized table data of interest with the key background information obtained to generate a serialized string.

[0024] Preferably, step S4 involves training a BART-based text generation model using serialized strings and descriptive text related to the table content to obtain the trained text generation model.

[0025] A system for generating descriptions based on an academic table and its selected content, according to the present invention, includes:

[0026] Module M1: Parses and extracts tables, table background information, and descriptive text related to table content from PDF academic documents, and matches the tables, table background information, and descriptive text related to table content.

[0027] Module M2: Constructs a domain knowledge base based on table background knowledge, preprocesses the table background knowledge and descriptive text related to table content, and serializes the table to obtain serialized table data;

[0028] Module M3: Uses preset templates to combine information and tables from preset tables that the user is interested in, and constructs serialized table data that the user is interested in;

[0029] Module M4: Obtains key background information based on the serialized table data of interest and the domain knowledge base, and generates a serialized string based on the obtained key background information and the serialized table data of interest;

[0030] Module M5: Trains a text generation model using serialized strings and descriptive text related to table content to obtain the trained text generation model;

[0031] Module M6: Uses the trained text generation model to predict and generate table-related descriptions.

[0032] Preferably, module M1 adopts:

[0033] Module M1.1: Uses PDFFigure2 to parse PDF academic documents and obtain tables, background information about the tables, and descriptive text related to the table content.

[0034] Module M1.2: Uses heuristic algorithms to filter and match data on tables, background information about tables, and descriptive text related to table content in the parsed PDF academic documents;

[0035] The module M1.1 employs the following steps: acquiring PDF academic documents, using PDFFigure2 to parse the PDF, and obtaining JSON data of relevant tables and PDF text; parsing the JSON data to obtain descriptive text and background information related to the table content.

[0036] Preferably, the module M2 adopts:

[0037] Module M2.1: Uses the SCI-Spacy model to segment the descriptive text related to the table content and the background knowledge of the table into separate sentences;

[0038] Module M2.2: Cleans data that is not related to tables based on the text after sentence segmentation.

[0039] Preferably, the module M4 adopts:

[0040] Module M4.1: Construct the TSDAE model and train the TSDAE model using the serialized tabular data of interest and the domain knowledge base to obtain the trained TADAE model;

[0041] Module M4.2: Extracts key background information from the serialized tabular data and domain knowledge base based on the trained TSDAE model;

[0042] Module M4.3: Uses a preset template to concatenate the serialized table data of interest with the key background information obtained, and generates a serialized string.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. This invention is the first to process academic table data through manual annotation and machine processing, and proposes to control the generation of relevant descriptions based on the table content and information of particular interest to the user based on the background knowledge provided by the background knowledge base;

[0045] 2. The related table descriptions generated by this invention have high fluency. Based on the current training set data and test set data, the BLEU Score of the BART-base text generation task reaches 16.90. The fluency of 50 randomly selected generated texts observed by human observation is about 4.14 (human scoring of 1-5 points based on the fluency of the generated sentences).

[0046] 3. The performance of this invention is significantly better than other academic table text generation methods, making it the best invention method for academic table text generation to date;

[0047] 4. This invention can be used to process and parse tabular data and related background information in papers, and generate corresponding descriptions of the tables based on this data, which can be used to quickly understand the information in the article and greatly improve research efficiency. Attached Figure Description

[0048] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0049] Figure 1 This is a flowchart of the present invention.

[0050] Figure 2 This is a schematic diagram illustrating the data parsing and preprocessing process of this invention.

[0051] Figure 3 This is a schematic diagram illustrating the specific operation of the present invention.

[0052] Figure 4 This is a comparison chart of various indicators of the results of this invention and other automated testing tools.

[0053] Figure 5 This invention provides a comparison chart of the results of manual detection of various indicators using other methods, based on 50 randomly selected prediction results.

[0054] Figure 6 This invention provides statistical data for manually annotated academic tables.

[0055] Figure 7 This is a comparison diagram of the prediction table description of the present invention and the standard data marked by manual annotation. Detailed Implementation

[0056] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0057] To address the needs of daily application tasks, the purpose of this invention is to provide a method and system for generating descriptions based on academic tables and their selected content. This provides researchers with a reliable and efficient method and system for organizing and analyzing tabular data, enabling them to quickly understand the content of tabular data descriptions in academic literature and thus improve the efficiency of literature reading.

[0058] Example 1

[0059] According to the present invention, a method for generating descriptions based on an academic table and its selected content is provided, such as... Figures 1 to 7 As shown, it includes:

[0060] Step S1: Use PDFFigure2 to parse the PDF file, extract the tables, background information of the tables, and descriptive text related to the table content in the academic article, and use the background information of the tables as a knowledge base for a specific domain.

[0061] Step S2: Use a heuristic algorithm to filter and match the academic tables, background information about the tables, and descriptive text related to the table content obtained from the parsing, and align the tables, background information about the tables, and descriptive text related to the table content.

[0062] Step S3: Clean the background information and descriptive text related to the table content after filtering, and serialize the table.

[0063] Step S4: If the table contains information that the user is particularly interested in, then use the template to combine the information that the user is particularly interested in with the table to construct serialized data of the table with special interest.

[0064] Step S5: Construct the TSDAE model training dataset, validation dataset, and test dataset using tabular serialized data and a domain-specific knowledge base. Extract key background information from the tabular serialized data and the domain-specific knowledge base based on the trained TSDAE model.

[0065] Step S6: Use the template to concatenate the serialized table data with the obtained key background information to generate a serialized string;

[0066] Step S7: Train a BART-based text generation model based on the manually labeled training dataset, validation dataset, and test dataset of the generated serialized strings;

[0067] Step S8: Input the processed text generation model into the data to predict and generate relevant table descriptions.

[0068] The background knowledge of the table refers to the description of the sub-cells in the table;

[0069] The descriptive text related to the table content describes the entire table.

[0070] Specifically, step S1 includes:

[0071] Step S1.1: Retrieve the PDF to be processed from the database;

[0072] Step S1.2: Use PDFFigure2 to parse the PDF and obtain the relevant table data and the PDF's JSON data;

[0073] Step S1.3: Parse the JSON data to obtain descriptive text and background information related to the table content.

[0074] Specifically, step S2 employs the following approach: using a heuristic algorithm to filter and match the academic tables, table background information, and descriptive text related to the table content obtained through parsing, following the principles that the text contains table content (table header or table data), the text contains table description information, and the table context text (3 sentences); the matched text data is then linked and aligned with the table data in sequence.

[0075] Specifically, step S3 includes:

[0076] Step S3.1: Use the SCI-Spacy model to preprocess the descriptive text related to the table content and the table background knowledge to perform text segmentation;

[0077] Step S3.2: Clean the data that is not related to the table based on the text after sentence segmentation, and remove stop words and punctuation marks from the text.

[0078] Step S3.3: Serialize the table according to the template. The table serialization template is: "" <c> <c> <c> <r>, including special symbols <c>"This represents a separator for table content." <r>" represents a table line break symbol.

[0079] Sci-Spacy, released in 2019 by the Allen Institute for Artificial Intelligence, is a Python toolkit for natural language processing in the biomedical field, based on Spacy. It provides various functions including word segmentation, sentence segmentation, and named entity recognition, as well as multiple corresponding models based on biomedical corpora such as JNLPBA and BC5CDR. The en_core_sci_lg model used in step S301, trained with over 785,000 biomedical words and 600,000 word vectors, is the largest and most accurate sentence segmentation model provided by the Sci-Spacy toolkit.

[0080] Specifically, step S4 includes:

[0081] Step S4.1: If there is content that users are particularly interested in, use a template to combine the user-interested content with the serialized table. The serialized combination template is "<User-interested content>". <c><User Special Attention Content> <c> <h>Serialized tables, including special characters <h>"" indicates a separator between user-focused content and serialized tables.

[0082] Specifically, step S5 includes:

[0083] Step S5.1: Organize the parsed table data and manually annotate each table, with the annotations being the standard description of the table.

[0084] Step S5.2: Organize the obtained table - standard description array of 8967 entries, and group the data into: training set of 5380 data entries, validation set of 1793 data entries and test set of 1794 data entries, following the grouping ratio of 6:2:2;

[0085] Step S5.3: Preprocess the table-standard description data pairs using a template; the serialization and splicing template is "Serialize Table Data". <bkg>"Standard data description", including special symbols <bkg>The quotation mark (") serves as a separator between the table content and the background knowledge. This generates the standard model reference data.

[0086] Step S5.4: Preprocess the table and all its corresponding background knowledge sentences sequentially using the template. The serialization and splicing template is "Serialize Table Data". <bkg>"Background knowledge sentences". The processed data is a table containing several concatenated background knowledge sentences, represented as a table-background knowledge text array.

[0087] Step S5.5: For the divided dataset, use the processed table-standard description data and table-background knowledge text array as model inputs, train and debug the TSDAE model, and obtain the sentence vector network model.

[0088] Step S5.6: Calculate the similarity of sentence vectors using a projection algorithm, and select the n most similar background table-background knowledge sentences based on the input table-standard description data (n is set to 3 in this invention). After processing, each table data corresponds to n background knowledge sentences and one standard description sentence.

[0089] The TSDAE (Transformer-based Sequential Denoising Auto-Encoder) model learns sentence embeddings in an unsupervised manner based on a pre-trained Transformer and a sequential denoising autoencoder to generate sentence vectors. Research shows that TSDAE significantly outperforms other traditional language models that generate sentence vectors.

[0090] The background knowledge sentence selected by TSDAE that is closest to the table data was input into the untrained BASRT-base model, and the results were BLEU=2.00, METEOR=0.09, Bert Score=0.78, and BLEURT=-1.00.

[0091] BLEU, METEOR, BERT Score, and BLEURT are automated testing tools used to measure the quality of generated text. BLEU and METEOR calculate the amount of information in the generated text. Meanwhile, BERT Score and BLEURT are automated evaluation tools based on BERT pre-training, used to measure the similarity between the output sentence and the labeled text.

[0092] Specifically, step S6 includes:

[0093] Step S6.1: For the n most relevant background knowledge sentences in the acquired table data, use a template to concatenate the processed serialized table. The serialization concatenation template is "serialized table data". <bkg>Background knowledge sentences.

[0094] Specifically, step S7 includes: inputting the processed table-background knowledge-standard description data into the BART-base text generation model, and training and debugging it to obtain the text generation network model.

[0095] Specifically, step S8 includes: inputting the table for which descriptions need to be generated into the trained BART-base model after obtaining relevant description text through TSDAE and performing serialization processing, and returning the final result.

[0096] A system for generating descriptions based on an academic table and its selected content, according to the present invention, includes:

[0097] Module M1: Uses PDFFigure2 to parse PDF files, extracting tables, background information about the tables, and descriptive text related to the table content from academic articles, and using the background information about the tables as a knowledge base for a specific domain.

[0098] Module M2: Uses heuristic algorithms to filter and match the academic tables, background information about the tables, and descriptive text related to the table content obtained through parsing, and aligns the tables, background information about the tables, and descriptive text related to the table content.

[0099] Module M3: Cleans the background information and descriptive text related to the content of the filtered tables, and performs serialization processing on the tables.

[0100] Module M4: If the table contains information that users are particularly interested in, then use the template to combine the information that users are particularly interested in with the table to construct serialized data of the table with special interest.

[0101] Module M5: Utilizes tabular serialized data and a domain-specific knowledge base to construct the TSDAE model training dataset, validation dataset, and test dataset. Based on the trained TSDAE model, it extracts key background information from the tabular serialized data and the domain-specific knowledge base.

[0102] Module M6: Uses a template to concatenate serialized tabular data with key background information to generate a serialized string;

[0103] Module M7: Trains a BART-based text generation model using manually labeled training, validation, and test datasets of generated serialized strings;

[0104] Module M8: Inputs the processed data into the text generation model, predicts and generates table-related descriptions.

[0105] The background knowledge of the table refers to the description of the sub-cells in the table;

[0106] The descriptive text related to the table content describes the entire table.

[0107] Specifically, module M1 includes:

[0108] Module M1.1: Retrieves the PDF to be processed from the database;

[0109] Module M1.2: Uses PDFFigure2 to parse PDFs and obtain relevant tabular data and JSON data from the PDF;

[0110] Module M1.3: Parses JSON data to obtain descriptive text and background information related to the table content.

[0111] Specifically, module M2 employs a heuristic algorithm to filter and match the academic tables, table background information, and descriptive text related to the table content obtained through parsing, following the principles of: the text contains table content (table header or table data), the text contains table description information, and the table context text (3 sentences); the matched text data is then linked and aligned with the table data in sequence.

[0112] Specifically, module M3 includes:

[0113] Module M3.1: Uses the SCI-Spacy model to preprocess descriptive text related to table content and background knowledge for text segmentation;

[0114] Module M3.2: Cleans data unrelated to tables from the text after sentence segmentation, and removes stop words and punctuation marks from the text.

[0115] Module M3.3: Serializes tables based on a template. The table serialization template is: "" <c> <c> <c> <r>, including special symbols <c>"This represents a separator for table content." <r>" represents a table line break symbol.

[0116] Sci-Spacy, released in 2019 by the Allen Institute for Artificial Intelligence, is a Python toolkit for natural language processing in the biomedical field, based on Spacy. It provides various functions including word segmentation, sentence segmentation, and named entity recognition, as well as multiple corresponding models based on biomedical corpora such as JNLPBA and BC5CDR. Module M301 uses the en_core_sci_lg model, trained with over 785,000 biomedical words and 600,000 word vectors, which is the largest and most accurate sentence segmentation model provided by the Sci-Spacy toolkit.

[0117] Specifically, module M4 includes:

[0118] Module M4.1: If there is content that users are particularly interested in, use a template to combine the user-interested content with a serialized table. The serialized combination template is "<User-interested content>". <c><Content of Special User Attention> <c> <h>Serialized tables, including special characters <h>"" indicates a separator between user-focused content and serialized tables.

[0119] Specifically, module M5 includes:

[0120] Module M5.1: Organizes the parsed table data and manually annotates each table with standard descriptions of the table.

[0121] Module M5.2: Organize the acquired table - standard description array of 8967 records, and group the data into: training set of 5380 records, validation set of 1793 records and test set of 1794 records, following a grouping ratio of 6:2:2;

[0122] Module M5.3: Use templates to preprocess tables - standard description data pairs, serialization and splicing template is "Serialize Table Data" <bkg>"Standard data description", including special symbols <bkg>The quotation mark (") serves as a separator between the table content and the background knowledge. This generates the standard model reference data.

[0123] Module M5.4: Preprocesses the table and all its corresponding background information sentences sequentially using a template. The serialization and concatenation template is "Serialize Table Data". <bkg>"Background knowledge sentences". The processed data is a table containing several concatenated background knowledge sentences, represented as a table-background knowledge text array.

[0124] Module M5.5: For the partitioned dataset, the processed table-standard description data and table-background knowledge text array are used as model inputs. The TSDAE model is used for training and debugging to obtain the sentence vector network model.

[0125] Module M5.6: Calculates the similarity of sentence vectors using a projection algorithm, and selects the n most similar background table-background knowledge sentences based on the input table-standard description data (n is set to 3 in this invention). After processing, each table data corresponds to n background knowledge sentences and one standard description sentence.

[0126] The TSDAE (Transformer-based Sequential Denoising Auto-Encoder) model learns sentence embeddings in an unsupervised manner based on a pre-trained Transformer and a sequential denoising autoencoder to generate sentence vectors. Research shows that TSDAE significantly outperforms other traditional language models that generate sentence vectors.

[0127] The background knowledge sentence selected by TSDAE that is closest to the table data was input into the untrained BASRT-base model, and the results were BLEU=2.00, METEOR=0.09, Bert Score=0.78, and BLEURT=-1.00.

[0128] BLEU, METEOR, BERT Score, and BLEURT are automated testing tools used to measure the quality of generated text. BLEU and METEOR calculate the amount of information in the generated text. Meanwhile, BERT Score and BLEURT are automated evaluation tools based on BERT pre-training, used to measure the similarity between the output sentence and the labeled text.

[0129] Specifically, module M6 includes:

[0130] Module M6.1: For the n most relevant background knowledge sentences from the acquired tabular data, use a template to concatenate the pre-processed serialized table. The serialization concatenation template is "serialized table data". <bkg>Background knowledge sentences.

[0131] Specifically, module M7 includes: inputting the processed table-background knowledge-standard description data into the BART-base text generation model, and training and debugging it to obtain the text generation network model.

[0132] Specifically, module M8 includes: inputting the table for which descriptions need to be generated into the trained BART-base model after obtaining relevant description text through TSDAE and performing serialization processing, and returning the final result.

[0133] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.

[0134] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.< / bkg> < / bkg> < / bkg> < / bkg> < / h> < / h> < / c> < / c> < / r> < / c> < / r> < / c> < / c> < / c> < / bkg> < / bkg> < / bkg> < / bkg> < / h> < / h> < / c> < / c> < / r> < / c> < / r> < / c> < / c> < / c>

Claims

1. A method for generating descriptions based on academic tables and their selected content, characterized in that, include: Step S1: Parse and extract tables, table background information, and descriptive text related to table content from PDF academic documents, and match the tables, table background information, and descriptive text related to table content. Step S2: Construct a domain knowledge base based on the table background knowledge, preprocess the table background knowledge and descriptive text related to the table content, and serialize the table to obtain serialized table data; Step S3: Use a preset template to combine the information and tables in the preset tables that the user is interested in, and construct the serialized table data that the user is interested in; Step S4: Obtain key background information based on the serialized table data of interest and the domain knowledge base, and generate a serialized string based on the obtained key background information and the serialized table data of interest; Step S5: Train the text generation model using the serialized string and descriptive text related to the table content to obtain the trained text generation model; Step S6: Use the trained text generation model to predict and generate table-related descriptions; Step S1 adopts the following: Step S1.1: Use PDFFigure2 to parse PDF academic documents and obtain tables, background information on tables, and descriptive text related to table content. Step S1.2: Use a heuristic algorithm to filter and match the tables, background information of the tables, and descriptive text related to the table content in the parsed PDF academic documents; Step S1.1 involves: acquiring PDF academic documents, using PDFFigure2 to parse the PDF, and obtaining JSON data of relevant tables and PDF text; parsing the JSON data to obtain descriptive text and background information related to the table content; Step S2 employs the following: Step S2.1: Use the SCI-Spacy model to segment the descriptive text related to the table content and the background knowledge of the table into sentences; Step S2.2: Clean up data unrelated to the table based on the segmented text; Step S4 employs the following: Step S4.1: Construct the TSDAE model and train the TSDAE model using the serialized tabular data of interest and the domain knowledge base to obtain the trained TSDAE model; Step S4.2: Based on the trained TSDAE model, extract key background information from the serialized tabular data of interest and the domain knowledge base; Step S4.3: Use a preset template to concatenate the serialized table data of interest with the key background information obtained to generate a serialized string; Step S5 involves training a BART-based text generation model using serialized strings and descriptive text related to the table content, resulting in the trained text generation model.

2. A system for generating descriptions based on academic tables and their selected content, characterized in that, include: Module M1: Parses and extracts tables, table background information, and descriptive text related to table content from PDF academic documents, and matches the tables, table background information, and descriptive text related to table content. Module M2: Constructs a domain knowledge base based on table background knowledge, preprocesses the table background knowledge and descriptive text related to table content, and serializes the table to obtain serialized table data; Module M3: Uses preset templates to combine information and tables from preset tables that the user is interested in, and constructs serialized table data that the user is interested in; Module M4: Obtains key background information based on the serialized table data of interest and the domain knowledge base, and generates a serialized string based on the obtained key background information and the serialized table data of interest; Module M5: Trains a text generation model using serialized strings and descriptive text related to table content to obtain the trained text generation model; Module M6: Uses the trained text generation model to predict and generate table-related descriptions; The module M1 adopts: Module M1.1: Uses PDFFigure2 to parse PDF academic documents and obtain tables, background information about the tables, and descriptive text related to the table content. Module M1.2: Uses heuristic algorithms to filter and match data on tables, background information about tables, and descriptive text related to table content in the parsed PDF academic documents; The module M1.1 employs the following steps: acquiring PDF academic documents, using PDFFigure2 to parse the PDF, and obtaining JSON data of relevant tables and PDF text; parsing the JSON data to obtain descriptive text and background information related to the table content; The module M2 adopts: Module M2.1: Uses the SCI-Spacy model to segment the descriptive text related to the table content and the background knowledge of the table into separate sentences; Module M2.2: Cleans data unrelated to tables based on the segmented text; The module M4 adopts: Module M4.1: Construct the TSDAE model and train the TSDAE model using the serialized tabular data of interest and the domain knowledge base to obtain the trained TSDAE model; Module M4.2: Extracts key background information from the serialized tabular data and domain knowledge base based on the trained TSDAE model; Module M4.3: Uses a preset template to concatenate the serialized table data of interest with the key background information obtained, and generates a serialized string.

Images