A thesis information extraction method

Abstract

The application provides a thesis information extraction method, and relates to the field of information processing. A specified vocabulary is extracted from a to-be-processed thesis to construct a thesis word set. Based on the thesis word set and the subject vocabulary table corresponding to each subject, the target subject corresponding to the thesis is determined from each subject. Further, based on the thesis word set and the contribution type vocabulary table corresponding to each contribution type under the target subject, the target contribution type corresponding to the thesis is determined from each contribution type. The pre-defined field name corresponding to the target subject and the pre-defined field name corresponding to the target contribution type are obtained to form a slot set. Based on the slot set, structured information extraction is performed on the thesis to obtain an extracted field value set. The slot set can adapt to the subject and contribution type of the to-be-processed thesis, fully covers the core knowledge of the thesis, and improves the accuracy of the fields extracted by the slot set.

Description

Technical Field

[0001] This application relates to the field of information processing, and in particular to a method for extracting information from academic papers. Background Technology

[0002] With the explosive growth in the number of scientific papers, how to efficiently extract structured knowledge from massive amounts of papers has become a core challenge in scientific research informatization. Currently, paper information extraction methods are mainly divided into two categories: one is rule-based extraction methods based on fixed templates, which predefine a set of fixed fields and extract the same information items uniformly from all papers; the other is free extraction methods based on large language models, which directly prompt the large model to extract key information from the paper, and the extracted fields are freely determined by the large model. These methods have achieved certain results in general knowledge domains, but when facing specialized domains, they still have the following shortcomings: First, fixed templates cannot adapt to papers with different contribution types. Within the same field, papers with different contribution types (such as model development, phenomenon analysis, and impact assessment) contain significantly different core knowledge. Rule extraction methods based on fixed templates use the same information fields to extract knowledge from all papers, which may result in the paper's knowledge not being covered by the information fields, leading to low information density and poor usability of the extraction results.

[0003] Second, free extraction based on large models lacks determinism and verifiability. Free extraction methods based on large language models completely delegate field selection to the large model, making the extraction process opaque and unverifiable. Furthermore, the extracted fields from different papers are inconsistent, making it impossible to directly store them in structured databases for cross-paper analysis and comparison. Simultaneously, large models may produce illusions, extracting information not present in the papers, resulting in a lack of quality assurance.

[0004] Therefore, there is a need to provide a method for extracting information from papers that can be applied to papers in various fields and of various types, thereby improving the accuracy of information extraction. Summary of the Invention

[0005] In view of this, this application provides a method for extracting information from a paper, the method comprising: Extract specified words from the paper to be processed to construct a word set for the paper; Based on the aforementioned set of keywords for the paper and a pre-constructed subject-specific vocabulary list, the target subject corresponding to the paper is determined from each subject; wherein, the subject-specific vocabulary list corresponds one-to-one with the subject. Based on the aforementioned set of paper terms and a pre-constructed vocabulary of contribution types, the target contribution type corresponding to the paper is determined from the various contribution types included in the target discipline; wherein, the vocabulary of contribution types corresponds one-to-one with the contribution type. Obtain the predefined field names corresponding to the target discipline and the predefined field names corresponding to the target contribution type to form a slot set; Based on the slot set, structured information extraction is performed on the paper to obtain a set of extracted field values.

[0006] Optionally, after performing structured information extraction on the paper based on the slot set to obtain the extracted field value set, the method further includes: Determine the total number of field values ​​contained in the extracted field value set to obtain the total number of fields to be filled; Determine the total number of predefined field names contained in the slot set to obtain the total number of field names; The ratio of the total number of fields to the total number of field names is taken as the total filling ratio; Based on the total filling ratio, a first quality assessment is performed on the extracted field value set; If the result of the first quality assessment is unqualified, then the unfilled target field name is determined from the slot set; the target field name is used to extract information from the paper to supplement the extracted field value set.

[0007] Optionally, based on the total fill ratio, a first quality assessment is performed on the extracted field value set, including: Determine the number of required predefined field names in the slot set to obtain the total number of required field names; Determine the proportion of the number of required field names to the total number of field names to obtain the required field ratio; and determine the product of the required field ratio and a preset leniency coefficient to obtain the adaptive fill ratio threshold. Determine whether the total fill ratio is less than the adaptive fill ratio threshold; If so, then the result of the first quality assessment is determined to be unqualified; If not, then the result of the first quality assessment is determined to be qualified.

[0008] Optionally, the method further includes: The fields in the extracted field value set that correspond to the predefined field names that are required are designated as required fields; the number of these required fields is designated as the number of required fields. The ratio of the number of required fields to the number of required field names is used as the required field filling ratio; The overall quality score is obtained by weighted summing of the total fill ratio and the fill ratio of the required fields. Based on the comprehensive quality score, a second quality assessment is performed on the extracted field value set; If the result of the second quality assessment is unqualified, then the unfilled target field names are determined from the slot set; the target field names are used to perform structured information extraction on the paper to supplement the extracted field value set.

[0009] Optionally, based on the set of keywords for the paper and a pre-constructed subject-specific vocabulary, the target subject corresponding to the paper is determined from various subjects, including: For each word in the term set of the paper, determine the word frequency in the paper and the inverse document frequency in the subject vocabulary database composed of all the subject vocabulary lists; and use the product of the word frequency and the inverse document frequency as the word score; The sum of the word scores for each of the aforementioned words is taken as the total word score; For each subject, the words common to the corresponding subject vocabulary and the paper vocabulary set are taken as the first common words; the sum of the word scores of the first common words is taken as the first score; the ratio of the first score to the total word score is taken as the subject score. The subject with the highest subject score among all the subjects is taken as the target subject.

[0010] Optionally, based on the paper terminology set and a pre-constructed contribution type vocabulary, the target contribution type corresponding to the paper is determined from the contribution types included in the target discipline, including: For each contribution type, perform the following steps: The words shared by the contribution type vocabulary and the paper term set are taken as the second shared words; The sum of the word scores of the second shared words is determined to obtain the second score; and the ratio of the second score to the total word score is determined to obtain the type score. Determine the total frequency of words in the contribution type vocabulary in the paper; and determine the ratio of the total frequency to the total number of words in the paper's word set to obtain a vocabulary density score; Based on the type score and vocabulary density score corresponding to each of the aforementioned contribution types, the target contribution type is determined from among the aforementioned contribution types.

[0011] Optionally, based on the type score and vocabulary density score corresponding to each of the contribution types, the target contribution type is determined from the contribution types, including: Based on each type score, determine the type weight corresponding to the type score; Based on each of the vocabulary density scores, determine the density weight corresponding to the vocabulary density scores; For each contribution type, the ratio of the corresponding word density score to the maximum word density score is determined to obtain the normalized density score; using the type weight and the density weight, the type score and the normalized density score are weighted and summed to obtain the type confidence. Determine the adaptive confidence threshold based on the type confidence of each contribution type; Among all the contribution types, the contribution type whose corresponding type confidence is greater than the adaptive confidence threshold is taken as the target contribution type.

[0012] Optionally, the method further includes: Construct a relational database; wherein the relational database is used to store the corresponding structured data with the unique identifier of the paper as the primary key; the structured data includes at least the paper's metadata, the target discipline corresponding to the paper, the target contribution type, and the set of extracted field values; Construct a vector database; wherein the vector database stores corresponding vector data with the unique identifier of the paper as the association key; the vector data includes at least a vectorized representation of the paper's word set; When a paper retrieval instruction is received, the paper retrieval is performed using the relational database and the vector database to determine the retrieval results.

[0013] Optionally, a paper retrieval is performed using the relational database and the vector database to determine the retrieval results, including: Determine the semantic similarity score between the paper retrieval command and the vector data; the semantic similarity score is used to characterize the semantic similarity between the paper retrieval command and the vector data; Based on the semantic similarity score, candidate identifiers are determined from the unique identifiers of each paper. Extract candidate structured data corresponding to the candidate identifier from the relational database; A matching score is determined between the paper retrieval command and the candidate structured data; the matching score is used to characterize the degree of matching between the paper retrieval command and the candidate structured data. Based on the semantic similarity score and matching score corresponding to the candidate identifiers, the target identifier is determined from each of the candidate identifiers; The data corresponding to the target identifier is extracted from the relational database and the vector database to obtain the retrieval results.

[0014] Optionally, based on the semantic similarity score and matching score corresponding to the candidate identifiers, the target identifier is determined from the candidate identifiers, including: Identify the predefined explicit constraint terms contained in the paper search command and determine the number of explicit constraint terms; Based on the explicit constraint word count and the total word count of the paper retrieval instruction after removing stop words, the semantic ambiguity of the paper retrieval instruction is determined; wherein, the semantic ambiguity is used to characterize the degree of semantic ambiguity of the paper retrieval instruction. The semantic ambiguity is used as the semantic weight corresponding to the semantic similarity score; and the difference obtained by subtracting the semantic weight is used as the matching weight corresponding to the matching score. For each candidate identifier, the semantic similarity score and matching score corresponding to the candidate identifier are weighted and summed using the semantic weight and the matching weight to obtain a comprehensive retrieval score. Based on the comprehensive search score, the target identifier is determined from the candidate identifiers.

[0015] In summary, this application provides a method for extracting information from academic papers. Specific terms are extracted from the paper to be processed to construct a term set. Based on the term set and the corresponding subject lexicons, the target subject of the paper is determined. Further, based on the term set and the contribution type lexicons of each contribution type under the target subject, the target contribution type of the paper is determined. Predefined field names corresponding to the target subject and the target contribution type are obtained to form a slot set. Based on the slot set, structured information extraction is performed on the paper to obtain a set of extracted field values. The slot set can adapt to the subject and contribution type of the paper to be processed, fully covering the core knowledge of the paper and improving the accuracy of the fields extracted using the slot set. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the first process of a paper information extraction method provided in this application.

[0017] Figure 2 This is a schematic diagram illustrating the first principle of a paper information extraction method provided in this application.

[0018] Figure 3 This is a schematic diagram of the second process of a paper information extraction method provided in this application.

[0019] Figure 4 This is a schematic diagram of the third process of a paper information extraction method provided in this application.

[0020] Figure 5 This is a schematic diagram illustrating the second principle of a paper information extraction method provided in this application. Detailed Implementation

[0021] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0022] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0023] Please refer to Figure 1 , Figure 1 A first flowchart illustrating a paper information extraction method provided in this application, the method comprising: S11. Extract specified words from the paper to be processed to construct a set of words for the paper.

[0024] The specified terms mentioned above include, but are not limited to, the title, abstract, and keywords of the paper; specific word extraction can be achieved through natural language processing technology, and this application does not impose any particular limitations on this.

[0025] S12. Based on the thesis terminology set and the pre-constructed subject vocabulary list, determine the target subject corresponding to the thesis from each subject; wherein, the subject vocabulary list corresponds one-to-one with the subject.

[0026] Considering that papers from different disciplines contain different core knowledge, this application first determines the target discipline corresponding to the paper to be processed from each discipline, so as to use predefined extraction fields set for the target discipline to extract information from the paper and improve the accuracy of information extraction.

[0027] Specifically, this application pre-constructs a corresponding subject-specific glossary for each discipline; the subject-specific glossary consists of representative words from each sample paper belonging to the corresponding discipline. By comparing the set of words in the papers with the corresponding subject-specific glossaries for each discipline, the target discipline corresponding to the paper is determined. The specific implementation method for determining the target discipline will be illustrated in subsequent embodiments and will not be elaborated here.

[0028] S13. Based on the paper term set and the pre-constructed contribution type vocabulary, determine the target contribution type corresponding to the paper from the contribution types included in the target discipline; wherein, the contribution type vocabulary corresponds one-to-one with the contribution type.

[0029] Considering that papers with different contribution types within the same discipline contain different core knowledge, this application, after determining the target discipline corresponding to the paper, further determines the target contribution type corresponding to the paper, so as to use predefined extraction fields set for the target contribution type to extract information from the paper, thereby further improving the coverage and accuracy of information extraction.

[0030] It should be noted that the contribution type is not categorized based on the specific subfield to which the paper belongs, but rather on the objective problem that the paper addresses. As an optional example, the aforementioned disciplines include hydrology, computer science, seismology, and geothermal science; please refer to Table 1 below for the corresponding contribution types for each discipline.

[0031] Table 1. Disciplinary Contribution Types

[0032] In this application, a corresponding contribution type vocabulary is pre-constructed for each contribution type under each discipline. The contribution type vocabulary consists of representative words from sample papers belonging to the corresponding contribution type. By comparing the set of words in the papers with the contribution type vocabulary, the target contribution type corresponding to the paper is determined. The specific implementation method for determining the target contribution type will be exemplified in subsequent embodiments and will not be elaborated here.

[0033] S14. Obtain the predefined field names corresponding to the target discipline and the predefined field names corresponding to the target contribution type to form a slot set.

[0034] Please refer to Figure 2 , Figure 2 This is a schematic diagram illustrating the first principle of a paper information extraction method provided in this application. In this application, different extraction field names are predefined for different disciplines, forming core slots corresponding to the disciplines; different extraction field names are also predefined for different contribution types under each discipline, forming detailed slots corresponding to the contribution types.

[0035] The aforementioned core slots apply to all papers in this discipline, including but not limited to paper titles, author lists, conferences / journals, publication years, research objectives and motivations, and core contribution types. The specific settings can be configured according to actual needs, and this application does not impose any special limitations on them.

[0036] The detailed slots mentioned above are set according to the corresponding disciplines and contribution types. This application does not specifically limit the specific field names they contain. Taking the model development and validation type in hydrology as an example, the corresponding detailed slots include, but are not limited to, the hydrological model name, the characteristics of the study area and watershed, the input data and driving factors, the calibration and validation methods, the performance evaluation indicators and values, and the comparison results with the benchmark model. Taking the phenomenon analysis and mechanism type in hydrology as an example, the corresponding detailed slots include, but are not limited to, the research phenomenon, historical characteristics and definition standards, the main driving mechanisms and processes, and the spatiotemporal patterns of change. Taking the impact and risk assessment type in hydrology as an example, the corresponding detailed slots include, but are not limited to, the assessment type, the disaster-causing factor or pressure source, the exposure unit / analysis object, the quantified impact / cost / benefit / gap category, and the scale and degree of the impact.

[0037] S15. Based on the slot set, perform structured information extraction on the paper to obtain the extracted field value set.

[0038] At this point, the extraction slots suitable for the paper's subject and contribution type have been determined, resulting in a slot set. Using this slot set, structured information extraction is performed on the paper, yielding a set of extracted field values.

[0039] Taking the model development and validation type under hydrology as an example, the set of extracted field values ​​is shown in Table 2 below.

[0040] Table 2. Extracted Fields Set Table

[0041] In summary, this application provides a method for extracting information from academic papers. Specific terms are extracted from the paper to be processed to construct a term set. Based on the term set and the corresponding subject lexicons, the target subject of the paper is determined. Further, based on the term set and the contribution type lexicons of each contribution type under the target subject, the target contribution type of the paper is determined. Predefined field names corresponding to the target subject and the target contribution type are obtained to form a slot set. Based on the slot set, structured information extraction is performed on the paper to obtain a set of extracted field values. The slot set can adapt to the subject and contribution type of the paper to be processed, fully covering the core knowledge of the paper and improving the accuracy of the fields extracted using the slot set.

[0042] Based on the above embodiments: like Figure 2As shown, this application also performs a two-level adaptive quality assessment on the extracted field set, including an adaptive field filling ratio assessment and a weighted assessment of the non-emptiness and confidence of key extracted fields. If the quality assessment fails, supplementary extraction is performed to update the extracted field set and improve the quality of the extracted field set, i.e., the extracted information from the paper.

[0043] The following section will provide a detailed explanation of the adaptive field fill ratio evaluation.

[0044] As an optional embodiment, after performing structured information extraction on the paper based on the slot set to obtain the extracted field value set, the method further includes: Determine the total number of field values ​​contained in the extracted field value set to obtain the total number of fields to fill; Determine the total number of predefined field names contained in the slot set to obtain the total number of field names; The ratio of the total number of fields to the total number of field names is used as the total fill ratio; Based on the total fill ratio, a first quality assessment is performed on the extracted field value set; If the result of the first quality assessment is unqualified, then the unfilled target field names are determined from the slot set; the target field names are used to extract information from the paper to supplement the extracted field value set.

[0045] First, based on the total number of field values ​​actually contained in the extracted field value set and the total number of predefined field names contained in the slot set, the fill rate of the extracted field value set relative to the slot set is determined, which is the total fill rate mentioned above. The total fill rate can be expressed as: fill_rate = filled_slots / total_slots; where filled_slots is the total number of fields filled, and total_slots is the total number of field names.

[0046] By comparing the total fill ratio with a fill ratio threshold, the suitability of the extracted field value set is evaluated. To improve the accuracy of the quality assessment, this embodiment determines an adaptive fill ratio threshold that is dynamically determined by the proportion of required field names in the slot set. By comparing the total fill ratio with the adaptive fill ratio threshold, the result of the first quality assessment is determined. The process of determining the adaptive fill ratio threshold is explained below.

[0047] As an optional implementation, a first quality assessment is performed on the extracted field value set based on the total fill ratio, including: Determine the number of required predefined field names in the slot set to obtain the total number of required field names; Determine the proportion of required field names to the total number of field names to obtain the required field ratio; and determine the product of the required field ratio and the preset leniency coefficient to obtain the adaptive fill ratio threshold. Determine if the total fill ratio is less than the adaptive fill ratio threshold; If so, the result of the first quality assessment is determined to be unqualified; If not, then the result of the first quality assessment is deemed satisfactory.

[0048] The adaptive fill ratio threshold mentioned above can be expressed as: theta_fill = required_slots / total_slots * alpha_fill; where required_slots is the number of required field names, total_slots is the total number of field names, and alpha_fill is a preset leniency coefficient (default is 0.9, which can be set according to actual needs). This achieves the goal of adaptively adjusting the fill ratio threshold according to the density of required field names in the slot set, rather than setting a fixed value.

[0049] Based on this, if the total fill rate (fill_rate) is less than the adaptive fill rate threshold (theta_fill), the result of the first quality assessment is determined to be unqualified, thus triggering a re-extraction process. This involves identifying the unfilled target field names from the slot set and using these target field names to extract information from the paper, supplementing the extracted field value set. This achieves the purpose of re-invoking the large model to supplement the extraction for targeted prompts with unfilled field names. Furthermore, a maximum number of re-extractions can be set (e.g., 2 times).

[0050] If the total fill rate (fill_rate) is not less than the adaptive fill rate threshold (theta_fill), then the result of the first quality assessment is determined to be qualified, and the current set of extracted field values ​​is directly used as the paper information extraction result.

[0051] The following section provides a detailed explanation of the non-nullability of key extraction fields and the confidence-weighted evaluation.

[0052] As an optional embodiment, the method further includes: The fields with the predefined required field names in the extracted field value set are designated as required fields; the number of required fields is designated as the number of required fields. The ratio of the number of required fields to the number of required field names is used as the required field filling ratio; The overall quality score is obtained by weighted summing of the total fill ratio and the fill ratio of required fields. A second quality assessment is performed on the extracted field value set based on the comprehensive quality score. If the result of the second quality assessment is unqualified, then the unfilled target field names are determined from the slot set; the target field names are used to perform structured information extraction on the paper to supplement the extracted field value set.

[0053] In this embodiment, each field in the extracted field value set that corresponds to a required predefined field is subjected to a non-empty validation; the number of required fields in the extracted field value set is determined as the number of required fields. The ratio of the number of required fields to the number of required field names is used as the required field filling ratio. The required field filling ratio can be expressed as: key_rate = filled_required / required_slots; where filled_required is the number of required fields and required_slots is the number of required field names; this is equivalent to calculating the key field completeness rate of the extracted field value set.

[0054] Based on this, a weighted sum of the total fill rate (fill_rate) and the required field fill rate (key_rate) is calculated to obtain the overall quality score. As an optional implementation, the weight corresponding to the total fill rate (fill_rate) is w_fill, and the weight corresponding to the required field fill rate (key_rate) is w_key. These two weights are dynamically determined by the structure of the current slot set; w_key = required_slots / total_slots, w_fill = 1 - w_key; where required_slots is the number of required field names, and total_slots is the total number of field names in the current slot set. Contribution types with a high required field name ratio (such as model development) are given a higher weight for their required field fill rate.

[0055] After obtaining the overall quality score, a second quality assessment is performed on the extracted field value set based on the overall quality score. As an optional implementation, if the overall quality score quality_score >= 0.8, the result of the second quality assessment is determined as: quality_flag = complete, indicating that the core field extraction is complete and the quality assessment is qualified; if 0.6 <= quality_score < 0.8, then quality_flag = partial, indicating that some core fields are missing and the quality assessment is unqualified; if quality_score < 0.6, then quality_flag = incomplete, indicating that the core field extraction is severely incomplete and the quality assessment is unqualified.

[0056] For cases where the quality assessment is unsatisfactory, the unfilled target field names are determined from the slot set; the target field names are then used to perform structured information extraction on the paper to supplement the extracted field value set.

[0057] The process of determining the target discipline for a paper is explained in detail below.

[0058] As an optional implementation, based on the paper's terminology set and a pre-built subject-specific vocabulary, the target subject corresponding to the paper is determined from various disciplines, including: For each word in the thesis term set, determine the word frequency in the thesis and the inverse document frequency in the subject lexicon consisting of all subject lexicons; and use the product of the word frequency and the inverse document frequency as the word score; The sum of the word scores for each word is taken as the total word score; For each subject, the words shared by the corresponding subject vocabulary and the thesis vocabulary set are taken as the first shared words; the sum of the word scores of the first shared words is taken as the first score; and the ratio of the first score to the total word score is taken as the subject score. The subject with the highest score among all subjects will be selected as the target subject.

[0059] As mentioned earlier, information such as titles, abstracts, and keywords were extracted from the papers from which information extraction was to be performed, forming a term set. For each term in the term set, its TF-IDF score was determined, which is the term score mentioned above. The term score can be expressed as tf(w) = tf(w) * log(N / (df(w)+1)), where tf(w) is the term frequency, log(N / (df(w)+1)) is the inverse document frequency, N is the total number of sample papers constituting the subject term lexicon, and df(w) is the total number of papers in the subject term lexicon containing the current term.

[0060] Based on this, for each discipline, a TF-IDF weighted coverage rate is determined as the discipline score. The discipline score can be expressed as: cov_d = sum(tfidf(w) for w in V_d∩W_paper) / sum(tfidf(w) for w in W_paper); where w in V_d∩W_paper represents words that exist in both the discipline vocabulary V_d and the paper term set W_paper, i.e., the first common words mentioned above; sum(tfidf(w) for w in V_d∩W_paper) is the sum of the word scores of the first common words, i.e., the first score mentioned above; winW_paper represents words in the paper term set, and sum(tfidf(w) for w in W_paper) is the sum of the word scores of each word in the paper term set, i.e., the total word score mentioned above.

[0061] Finally, the subject with the highest subject score (cov_d) is selected from all subjects as the target subject to improve the accuracy of subject attribution.

[0062] The process of determining the target contribution type of a paper is explained in detail below.

[0063] As an optional implementation, based on the paper terminology set and a pre-built contribution type vocabulary, the target contribution type corresponding to the paper is determined from the various contribution types included in the target discipline, including: For each contribution type, perform the following steps: Words shared by the contribution type vocabulary and the thesis term set are considered as the second shared words; Determine the sum of the word scores for the second shared words to obtain the second score; and determine the ratio of the second score to the total word score to obtain the type score. Determine the total frequency of words in the contribution type vocabulary in the paper; and determine the ratio of the total frequency to the total number of words in the paper's word set to obtain a vocabulary density score; Based on the type score and vocabulary density score corresponding to each contribution type, the target contribution type is determined from among the contribution types.

[0064] In this embodiment, for each contribution type, it is determined whether it is likely to be the target contribution type corresponding to the paper, rather than selecting a target contribution type from among the various contribution types. That is, a paper may correspond to multiple target contribution types, which are mixed contribution types, in order to improve the comprehensiveness of subsequent information extraction.

[0065] Specifically, for each contribution type, its type score is first determined. The type score can be expressed as cov_i = sum(tfidf(w) for w in V_i∩W_paper) / sum(tfidf(w) for w in W_paper); where w in V_i∩W_paper represents words that exist both in the contribution type vocabulary V_i corresponding to the current contribution type and in the paper word set W_paper, which is the second common word mentioned above; correspondingly, sum(tfidf(w) for w in V_i∩W_paper) represents the sum of the word scores of the second common words, which is the second score mentioned above; sum(tfidf(w) for w in W_paper) is the total word score mentioned above.

[0066] Building upon this, this embodiment also introduces a lexical density correction. The lexical density score can be expressed as: density_i = freq_i / L; where freq_i is the total frequency of words in the contribution type vocabulary V_i in the paper; and L is the total number of words in the paper's word set.

[0067] The following section provides a detailed explanation of how to determine whether a contribution type is the target contribution type based on the type score and lexical density score corresponding to each contribution type. As an optional embodiment, the target contribution type is determined from among the contribution types based on the type score and lexical density score corresponding to each contribution type, including: Based on the scores for each type, determine the type weight corresponding to the type score; Based on each word density score, determine the density weight corresponding to the word density score; For each contribution type, the ratio of the corresponding word density score to the maximum word density score is determined to obtain the normalized density score; using type weight and density weight, the type score and the normalized density score are weighted and summed to obtain the type confidence. Determine the adaptive confidence threshold based on the type confidence of each contribution type; Among all contribution types, the contribution type with a type confidence score greater than the adaptive confidence score threshold is selected as the target contribution type.

[0068] In this embodiment, the type weight corresponding to the type score and the density weight corresponding to the word density score are both dynamically determined by the word distribution characteristics of the current paper.

[0069] As an optional implementation, the standard deviation of the type score cov_i corresponding to all contribution types is determined to obtain the standard deviation sigma_cov; the standard deviation of the word density score density_i corresponding to all contribution types is determined to obtain the standard deviation sigma_den; if sigma_cov + sigma_den = 0 (i.e., the type score and word density score of all contribution types are the same, and both standard deviations are zero), then the type weight beta_cov and the density weight beta_den are both set to 0.5, degenerating into equal-weight fusion; otherwise, the type weight beta_cov = sigma_cov / (sigma_cov + sigma_den), and the density weight beta_den = 1 - beta_cov. In summary, under this calculation method, the larger the standard deviation, the higher the discrimination of that dimension to different contribution types. Dimensions with higher discrimination automatically receive greater weights, causing the confidence calculation to tilt towards more discriminative features.

[0070] For each contribution type, its type confidence can be expressed as: conf_i = beta_cov * cov_i + beta_den * (density_i / max_density); where max_density is the maximum value among all word density scores, i.e., the maximum word density score mentioned above; density_i / max_density is the normalized value of the word density score, i.e., the normalized density score mentioned above. Specifically, when max_density = 0 (the paper text is empty or the vocabulary has no hits), let density_i / max_density = 0.

[0071] In this embodiment, the adaptive confidence threshold theta_type is determined by the mean mu_conf and standard deviation sigma_conf of the confidence scores corresponding to all contribution types. The adaptive confidence threshold can be expressed as: theta_type = mu_conf + gamma * sigma_conf; where gamma is the sensitivity coefficient, which defaults to 0 (i.e., the mean is used as the threshold); when the themes of the papers are highly concentrated, sigma_conf is large, and theta_type is automatically adjusted upward to avoid too many types being activated.

[0072] Based on this, contribution types with a confidence level higher than the adaptive confidence threshold `theta_type` are identified as activated target contribution types. If multiple target contribution types are activated, the paper is marked as a hybrid type, and the type confidence levels corresponding to all target contribution types are retained. For papers corresponding to a hybrid type, i.e., covering multiple contribution types, all detailed slots corresponding to each contribution type are activated, and the detailed slots are prioritized in descending order of the type confidence level of the corresponding contribution type. Detailed slots corresponding to contribution types with lower type confidence levels have lower priority and are presented with lower priority in the large model's extraction of prompts, avoiding excessively long prompts that could distract the large model and affect the paper's information extraction performance. Furthermore, if the type confidence levels of all contribution types are lower than the adaptive confidence threshold `theta_type`, the type with the highest confidence level is selected as the target contribution type.

[0073] Considering that related technologies build paper information extraction and paper semantic retrieval as independent systems, users cannot directly obtain the structured knowledge of the paper when performing semantic retrieval, requiring additional manual review steps, which seriously affects research efficiency; at the same time, existing methods lack a query mechanism that dynamically integrates semantic similarity and structured matching degree, and cannot adaptively adjust the retrieval strategy according to the semantic ambiguity of the user's query.

[0074] To address the above issues, this application constructs a relational database with a dual-track associated storage module based on extracted field sets, providing a semantic-structural dynamic fusion query architecture that supports efficient joint retrieval of papers.

[0075] As an optional embodiment, the method further includes: Construct a relational database; the relational database is used to store the corresponding structured data with the unique identifier of the paper as the primary key; the structured data includes at least the paper's metadata, the target discipline corresponding to the paper, the target contribution type, and the set of extracted field values; Construct a vector database; wherein the vector database stores the corresponding vector data with the unique identifier of the paper as the association key; the vector data includes at least the vectorized representation of the paper's word set; When a paper retrieval instruction is received, a paper retrieval is performed using relational databases and vector databases to determine the retrieval results.

[0076] In this embodiment, the relational database adopts a structured storage scheme with the unique identifier of the paper as the primary key. Each paper corresponds to one piece of structured data, which includes at least the paper's metadata, the target discipline corresponding to the paper, the target contribution type, and a set of extracted field values. Currently, in the aforementioned embodiment, various parameters related to the paper can also be added to the structured data, and can be set according to actual needs.

[0077] The vector database stores corresponding vector data using the unique identifier of the paper as the association key; the vector data includes at least a vectorized representation of the paper's word set, which is equivalent to a vectorized identifier including the paper's title, abstract, and keywords.

[0078] The relational and vector databases mentioned above establish associated indexes using unique paper identifiers, supporting dynamic semantic-structural fusion queries. Please refer to... Figure 3 , Figure 3 This is a schematic diagram of the second process of a paper information extraction method provided in this application. The following is combined with... Figure 3 The query process will be explained in detail.

[0079] As an optional implementation, paper retrieval is performed using relational databases and vector databases to determine the retrieval results, including: S31. Determine the semantic similarity score between the paper retrieval command and the vector data; the semantic similarity score is used to characterize the semantic similarity between the paper retrieval command and the vector data. S32. Based on semantic similarity scoring, candidate identifiers are determined from the unique identifiers of each paper; candidate structured data corresponding to the candidate identifiers are extracted from the relational database. S33. Determine the matching score between the paper retrieval command and the candidate structured data; the matching score is used to characterize the degree of matching between the paper retrieval command and the candidate structured data. S34. Based on the semantic similarity score and matching score corresponding to the candidate identifiers, determine the target identifier from each candidate identifier; S35. Extract the data corresponding to the target identifier from the relational database and the vector database to obtain the retrieval results.

[0080] After receiving the paper search command input by the user, a semantic search is first performed based on a vector database. Specifically, for each vector data in the vector database, a semantic similarity score is determined between the paper search command and the vector data; the higher the semantic similarity between the paper search command and the vector data, the higher the semantic similarity score. Based on the semantic similarity scores of each vector data, candidate identifiers are determined from the corresponding unique identifiers of each paper. As an optional embodiment, the unique identifiers of the papers corresponding to a predetermined number of vector data with high semantic similarity scores are used as candidate identifiers.

[0081] From the relational database, extract candidate structured data corresponding to the candidate identifiers. For each candidate structured data, determine the matching score between the paper retrieval command and the candidate structured data (based on the degree of matching between fields such as contribution type, discipline, and year and the paper retrieval command). The matching score is used to characterize the degree of matching between the two; the higher the degree of matching, the higher the matching score.

[0082] Finally, the target identifier is determined from the candidate identifiers by combining their semantic similarity and matching scores. Data corresponding to the target identifier is extracted from relational and vector databases to obtain the search results, which serve as a response to the paper search command.

[0083] Please refer to Figure 4 , Figure 4 This is a schematic diagram of the third process of a paper information extraction method provided in this application. As an optional embodiment, the target identifier is determined from the candidate identifiers based on the semantic similarity score and matching score corresponding to the candidate identifiers, including: S41. Identify the predefined explicit constraint terms contained in the paper retrieval instructions and determine the number of explicit constraint terms; S42. Based on the explicit constraint word count and the total word count of the paper retrieval instruction after removing stop words, determine the semantic ambiguity of the paper retrieval instruction; wherein, the semantic ambiguity is used to characterize the degree of semantic ambiguity of the paper retrieval instruction. S43. Use semantic ambiguity as the semantic weight corresponding to the semantic similarity score; and use the difference between one and the semantic weight as the matching weight corresponding to the matching score. S44. For each candidate identifier, the semantic similarity score and matching score corresponding to the candidate identifier are weighted and summed using semantic weight and matching weight to obtain the comprehensive retrieval score. S45. Based on the comprehensive search score, determine the target identifier from the candidate identifiers.

[0084] The comprehensive search score can be expressed as: score(p) = lambda * sim_vec(p) + (1 - lambda) * sim_struct(p); where lambda is the semantic weight, sim_vec(p) is the semantic similarity score, (1 - lambda) is the matching weight, and sim_struct(p) is the matching score. The weight lambda is adaptively calculated based on the semantic ambiguity index of the user-input paper search command, i.e., the semantic ambiguity_score. The specific steps are as follows: (a) Structured Constraint Recognition: Perform structured constraint parsing on the paper retrieval command to identify the explicit constraint fields contained therein, including but not limited to: contribution type keywords (such as "model development", "impact assessment"), year range (such as "after 2020"), subject keywords (such as "hydrology", "seismology"), journal / conference name, etc. Let n_constraint be the number of explicit constraint words identified.

[0085] (b) Semantic ambiguity calculation: ambiguity_score = 1 - n_constraint / (n_constraint + n_free), where n_free is the number of non-explicitly constrained words in the paper search command (i.e., the number of remaining words after removing stop words and explicit constraint fields). ambiguity_score ∈ [0,1], and a larger value indicates that the search command is more ambiguous and more biased towards free semantic description.

[0086] (c) Lambda adaptive mapping: lambda = lambda_min + (lambda_max - lambda_min) * ambiguity_score; where lambda_min is the minimum fusion weight (default is 0.2, ensuring that semantic retrieval always makes a basic contribution), and lambda_max is the maximum fusion weight (default is 0.9, ensuring that structured filtering always makes a basic contribution); when ambiguity_score approaches 0 (query constraints are clear), lambda approaches lambda_min, biased towards structured filtering; when ambiguity_score approaches 1 (query is completely free), lambda approaches lambda_max, biased towards semantic retrieval.

[0087] The weighting mechanism described above is based entirely on the lexical statistical features of the paper retrieval instructions. Unlike the retrieval weight determination method based on reinforcement learning path planning, it requires no training process and can be calculated in real time.

[0088] In addition to the semantic-structural dynamic fusion query in the above embodiments, other filtering methods are also supported. For example, the structured filtering priority mode, that is, first accurately filtering in the relational database according to conditions such as contribution type, discipline, year, and key field values, and then re-ranking the candidate data based on semantic similarity. This embodiment will not be elaborated here.

[0089] This completes the extraction, storage, and retrieval of information from the paper. Please refer to... Figure 5 , Figure 5 This is a schematic diagram of the second principle of a paper information extraction method provided in this application. The steps of extracting, storing and retrieving the above paper information can be divided into subject determination, contribution type determination, slot activation, field extraction and dual-track associated storage and retrieval. Multiple adaptive dynamic algorithm mechanisms are introduced in the processing, which effectively solves the technical problem that the existing paper information extraction methods have fixed templates and cannot adapt to different contribution types, and significantly improves the structuring degree and retrieval accuracy of paper knowledge in the field of earth sciences.

[0090] While this specification contains numerous specific implementation details, these should not be construed as limiting the scope of any invention or the scope of the claims, but rather are primarily intended to describe features of specific embodiments of a particular invention. Certain features described in the various embodiments herein may also be implemented in combination in a single embodiment. Conversely, various features described in a single embodiment may also be implemented separately in various embodiments or in any suitable sub-combination. Furthermore, while features may function in certain combinations as described above and even initially claimed in this way, one or more features from a claimed combination may be removed from that combination in some cases, and a claimed combination may refer to a sub-combination or a variation thereof.

[0091] Similarly, although the operations are depicted in a specific order in the accompanying drawings, this should not be construed as requiring these operations to be performed in the specific order shown or sequentially, or requiring all illustrated operations to be performed to achieve the desired result. In some cases, multitasking and parallel processing may be advantageous. Furthermore, the separation of various system modules and components in the above embodiments should not be construed as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0092] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for extracting information from a thesis, characterized in that, The method includes: Extract specified words from the paper to be processed to construct a word set for the paper; Based on the aforementioned set of keywords for the paper and a pre-constructed subject-specific vocabulary list, the target subject corresponding to the paper is determined from each subject; wherein, the subject-specific vocabulary list corresponds one-to-one with the subject. Based on the aforementioned collection of paper terms and a pre-constructed vocabulary of contribution types, the target contribution type corresponding to the paper is determined from the various contribution types included in the target discipline; wherein, the vocabulary of contribution types corresponds one-to-one with the contribution types; the contribution type is a type classified according to the objective problem solved by the paper. Obtain the predefined field name corresponding to the target discipline to obtain the core slot, and obtain the predefined field name corresponding to the target contribution type to obtain the detailed slot, and form a slot set by the core slot and the detailed slot; Based on the slot set, structured information extraction is performed on the paper to obtain a set of extracted field values.

2. The paper information extraction method as described in claim 1, characterized in that, After performing structured information extraction on the paper based on the slot set to obtain the extracted field value set, the method further includes: Determine the total number of field values ​​contained in the extracted field value set to obtain the total number of fields to be filled; Determine the total number of predefined field names contained in the slot set to obtain the total number of field names; The ratio of the total number of fields to the total number of field names is taken as the total filling ratio; Based on the total filling ratio, a first quality assessment is performed on the extracted field value set; If the result of the first quality assessment is unqualified, then the unfilled target field name is determined from the slot set; the target field name is used to extract information from the paper to supplement the extracted field value set.

3. The paper information extraction method as described in claim 2, characterized in that, Based on the total fill ratio, a first quality assessment is performed on the extracted field value set, including: Determine the number of required predefined field names in the slot set to obtain the total number of required field names; Determine the proportion of the number of required field names to the total number of field names to obtain the required field ratio; and determine the product of the required field ratio and a preset leniency coefficient to obtain the adaptive fill ratio threshold. Determine whether the total fill ratio is less than the adaptive fill ratio threshold; If so, then the result of the first quality assessment is determined to be unqualified; If not, then the result of the first quality assessment is determined to be qualified.

4. The paper information extraction method as described in claim 3, characterized in that, The method further includes: The fields in the extracted field value set that correspond to the predefined field names that are required are designated as required fields; the number of these required fields is designated as the number of required fields. The ratio of the number of required fields to the number of required field names is used as the required field filling ratio; The overall quality score is obtained by weighted summing of the total fill ratio and the fill ratio of the required fields. Based on the comprehensive quality score, a second quality assessment is performed on the extracted field value set; If the result of the second quality assessment is unqualified, then the unfilled target field names are determined from the slot set; the target field names are used to perform structured information extraction on the paper to supplement the extracted field value set.

5. The paper information extraction method as described in claim 1, characterized in that, Based on the aforementioned set of keywords for the paper and a pre-constructed subject-specific vocabulary, the target subject corresponding to the paper is determined from various disciplines, including: For each word in the term set of the paper, determine the word frequency in the paper and the inverse document frequency in the subject vocabulary database composed of all the subject vocabulary lists; and use the product of the word frequency and the inverse document frequency as the word score; The sum of the word scores for each of the aforementioned words is taken as the total word score; For each subject, the words common to the corresponding subject vocabulary and the paper vocabulary set are taken as the first common words; the sum of the word scores of the first common words is taken as the first score; the ratio of the first score to the total word score is taken as the subject score. The subject with the highest subject score among all the subjects is taken as the target subject.

6. The paper information extraction method as described in claim 5, characterized in that, Based on the aforementioned set of paper terms and a pre-constructed vocabulary of contribution types, the target contribution type corresponding to the paper is determined from the various contribution types included in the target discipline, including: For each contribution type, perform the following steps: The words shared by the contribution type vocabulary and the paper term set are taken as the second shared words; The sum of the word scores of the second shared words is determined to obtain the second score; and the ratio of the second score to the total word score is determined to obtain the type score. Determine the total frequency of words in the contribution type vocabulary in the paper; and determine the ratio of the total frequency to the total number of words in the paper's word set to obtain a vocabulary density score; Based on the type score and vocabulary density score corresponding to each of the aforementioned contribution types, the target contribution type is determined from among the aforementioned contribution types.

7. The paper information extraction method as described in claim 6, characterized in that, Based on the type score and vocabulary density score corresponding to each of the aforementioned contribution types, the target contribution type is determined from the aforementioned contribution types, including: Based on each type score, determine the type weight corresponding to the type score; Based on each of the vocabulary density scores, determine the density weight corresponding to the vocabulary density scores; For each contribution type, the ratio of the corresponding word density score to the maximum word density score is determined to obtain the normalized density score; using the type weight and the density weight, the type score and the normalized density score are weighted and summed to obtain the type confidence. Determine the adaptive confidence threshold based on the type confidence of each contribution type; Among all the contribution types, the contribution type whose corresponding type confidence is greater than the adaptive confidence threshold is taken as the target contribution type.

8. The paper information extraction method as described in claim 1, characterized in that, The method further includes: Construct a relational database; wherein the relational database is used to store the corresponding structured data with the unique identifier of the paper as the primary key; the structured data includes at least the paper's metadata, the target discipline corresponding to the paper, the target contribution type, and the set of extracted field values; Construct a vector database; wherein the vector database stores corresponding vector data with the unique identifier of the paper as the association key; the vector data includes at least a vectorized representation of the paper's word set; When a paper retrieval instruction is received, the paper retrieval is performed using the relational database and the vector database to determine the retrieval results.

9. The paper information extraction method as described in claim 8, characterized in that, Performing paper retrieval using the relational database and the vector database to determine retrieval results includes: Determine the semantic similarity score between the paper retrieval command and the vector data; the semantic similarity score is used to characterize the semantic similarity between the paper retrieval command and the vector data; Based on the semantic similarity score, candidate identifiers are determined from the unique identifiers of each paper. Extract candidate structured data corresponding to the candidate identifier from the relational database; A matching score is determined between the paper retrieval command and the candidate structured data; the matching score is used to characterize the degree of matching between the paper retrieval command and the candidate structured data. Based on the semantic similarity score and matching score corresponding to the candidate identifiers, the target identifier is determined from each of the candidate identifiers; The data corresponding to the target identifier is extracted from the relational database and the vector database to obtain the retrieval results.

10. The paper information extraction method as described in claim 9, characterized in that, Based on the semantic similarity score and matching score corresponding to the candidate identifiers, the target identifier is determined from each of the candidate identifiers, including: Identify the predefined explicit constraint terms contained in the paper search command and determine the number of explicit constraint terms; Based on the explicit constraint word count and the total word count of the paper retrieval instruction after removing stop words, the semantic ambiguity of the paper retrieval instruction is determined; wherein, the semantic ambiguity is used to characterize the degree of semantic ambiguity of the paper retrieval instruction. The semantic ambiguity is used as the semantic weight corresponding to the semantic similarity score; and the difference obtained by subtracting the semantic weight is used as the matching weight corresponding to the matching score. For each candidate identifier, the semantic similarity score and matching score corresponding to the candidate identifier are weighted and summed using the semantic weight and the matching weight to obtain a comprehensive retrieval score. Based on the comprehensive search score, the target identifier is determined from the candidate identifiers.

Images