A method for generating differences in scientific literature based on citations, figures and structure

Abstract

The application discloses a scientific literature difference generation method based on citations, figures and structures, and steps are as follows: first, functionally marking scientific literature citations; second, collecting contrast citations and corresponding scientific literatures; third, calculating sentence relationship scores based on a contrast figure method; fourth, calculating sentence structure scores in combination with a structure theory; and fifth, sorting sentences and extracting difference contents formed by the sentences. The method is simple and efficient, does not depend on large-scale training data, and automatically generates difference contents of scientific literatures.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of automatic text summarization, natural language processing and artificial intelligence, and in particular to a scientific literature difference generation method based on citations, graphs and structures. BACKGROUND

[0002] With the large number of publications of scientific literature in the research field, it is increasingly difficult for researchers to quickly familiarize themselves with the development and related changes in the field from the vast literature base, which poses a great challenge to research breakthroughs. Prior to this, there have been related text generation work for scientific literature, such as summary generation, related work generation or review generation. Among them, the summary generation of scientific literature is only for a single document to generate a summary, or further supplements the summary content of an existing document based on the evaluation of other documents; the generation of related work and review is similar, aiming to summarize multiple documents to generate a summary, and the generated summary content is mainly the similar content among these multiple documents, which is also the highlight content in these documents. However, the summary generation of a single document or the generation of related work and review cannot effectively help researchers quickly understand the changes in the research field, such as technical differences and research trends.

[0003] The concept of scientific literature difference generation task has not been proposed in the prior art, and there is no related research. Difference generation is aimed at scientific multi-document summarization, which has two key differences compared to previous work. First, the target documents for scientific multi-document summarization, such as related work or review generation, are usually reference documents or field-related documents, and the content described in these documents has a large span and it is difficult to ensure that there are similarities and differences between them; second, the current method for multi-document summarization usually summarizes the similar content between documents, but does not include the difference content, and the generation of difference content in scientific literature needs to extract relevant information from more documents, and needs to consider the similarities and differences between documents, both capturing the same content between documents and mining the differences between documents.

[0004] The common application of citations in existing text summarization work is usually to find relevant citing documents through citation networks, and to further supplement the summary of the cited document through the description and evaluation of these documents. Recently, citation function classification has also been concerned in the summary of scientific literature, which aims to identify the function of citations in scientific literature to determine the purpose of the citing author citing the cited document; but how to apply it to the difference generation task to mine a set of related documents needs further research.

[0005] Structural theory has also received attention in abstract tasks in recent years, but there are currently few practices that apply structural theory to scientific literature abstract work and combine multiple structural theories to analyze between and within documents separately. Summary of the Invention

[0006] In order to solve the above technical problems, the present invention provides a method for generating differences in scientific and technological documents based on citations, graphs and structures. Based on citations, graphs and structures, "comparative citations" are used - summary content containing the citing author's own deep understanding of the cited documents, to find a set of comparable scientific and technological document collections; the present invention also introduces comparative graphs, cross-document structure theory and rhetorical structure theory respectively, and takes graph and structure theory into consideration in the difference generation work at the same time, so as to analyze the sentence structure within and between documents, capture relevant information more comprehensively, and provide scientific researchers with the difference content between scientific and technological documents.

[0007] The present invention provides a method for generating differences in scientific and technological literature based on citations, graphs, and structures, characterized in that the method comprises the following steps:

[0008] Step 1: Integrate citation function annotation methods, design a citation function annotation method that is sensitive to comparative citations, and perform functional annotation on citations in the literature;

[0009] Step 2: Select comparative citations from the marked citations and collect the corresponding literature;

[0010] Step 3: Construct a comparison chart of the document collection and calculate the sentence relationship scores in the documents;

[0011] Step 4: Calculate the sentence structure scores in the literature based on structural theory;

[0012] Step 5: Comprehensively consider the sentence relationship scores and sentence structure scores in the document to sort and extract the sentences, and finally form the target difference content.

[0013] Furthermore, in step 1, the classic citation function annotation method is integrated to reclassify, map and define the functional categories, and citations with a certain degree of contrast clues that originally belonged to other functional categories are reclassified into the contrast category, and then a citation function annotation method that is sensitive to capturing "contrasting citations" is proposed.

[0014] Furthermore, in step 2, only "comparative citations" are collected from the marked citation set, and the citing documents and cited documents corresponding to the comparative citations are collected to form a comparable document set.

[0015] Furthermore, in step 3, the specific steps of constructing a comparison graph of the document set and calculating the sentence relationship scores in the documents are as follows:

[0016] Step 3-1: Construct a comparison graph of the document collection. The comparison graph contains a subgraph of documents or citations, nodes of sentences, and edges of three types of inter-sentence relationships. The subgraph of documents or citations represents documents or citations, the nodes of sentences represent sentences in documents or citations, and the edges of three types of inter-sentence relationships represent three types of relationships between sentences: intra-document sentence relationships, inter-document sentence relationships, and citation-related sentence relationships. Intra-document sentence relationships consider the prominence of sentences within documents to capture the prominence within documents; inter-document sentence relationships consider the differences between documents to capture the differences between documents; and citation-related sentence relationships consider the similarity between document sentences and citations to capture the correlation between documents.

[0017] Step 3-2: Based on the comparison graph of the constructed document collection, a graph method is used to calculate the sentence relationship score; a sentence contains three relationship scores, and the sentence vector is obtained, and then similarity is calculated to measure different relationship scores;

[0018] Score within the document: Score the sentences in the document with S S The subscript s indicates the position in the document; the set of sentences preceding it is denoted as Pre, S pre Represents the sentences in the set Pre, and the subscript pre represents the position in Pre; the set of sentences following it is denoted as Aft, S aft Represents a sentence in the set Aft, and the subscript aft represents the position in Aft; the coefficients α and β are used to represent the coefficients when calculating the similarity with the previous and next sentences, and α, β∈[0,1]; Sentence S S The intra-document prominence score of is as follows;

[0019]

[0020] Inter-document score: The citing document is the document in which the citation appears, and the cited document is the document cited by the citation. The sentence in the citing document CP is scored with S cp The subscript cp indicates the position of the sentence in the citing document; the sentence in the cited document set RPs is represented by S rp The subscript rp indicates the position of the sentence in the cited literature set; the sentence S cp The difference score between the documents is as follows:

[0021]

[0022] Citation relevance score: Sentences in the literature are scored with S p Indicates that the subscript p indicates the position in the document; the sentence in the citation CIT is marked with S cit The subscript cit indicates the position in the quotation, and the sentence S p The citation-related commonality score is as follows;

[0023]

[0024] Further, in step 4, two structure theories of cross-document and rhetoric are introduced to expand the semantic structure of the sentence to multiple dimensions for analysis; the cross-document structure theory is aimed at the sentence structure relationship between documents, and the rhetoric structure theory is aimed at the sentence structure relationship within the document; first, preset the score for each structure relationship, and give high score to the sentence structure type which is the focus of attention of the cross-document or rhetoric structure relationship; then classify the cross-document structure type and the rhetoric structure type of the sentences in the document, and get the scores of the two structure theories of each sentence according to the obtained structure relationship category and the corresponding preset score.

[0025] Further, in step 5, the relationship score and the structure score of the sentences in each document are considered comprehensively, and the sentences are sorted in order from high to low according to the score, and then the sentences are selected from each document in order from top to bottom according to the preset number of sentences and score to finally form the difference content.

[0026] Further, in step 5, the sentences are sorted and selected according to the relationship score and the structure score.

[0027] The beneficial effects described in the present invention are: the present invention proposes a new text generation task - scientific and technological literature difference generation, which aims to generate differences between documents based on a given set of related scientific and technological documents with different and similar contents. The differences contain both similar and different contents. By comparing the different and similar contents, it can better provide researchers with research changes and help them understand the development trends in the field. The present invention is based on the method involved in this, adopts the citation function annotation method, and borrows the comparative clues in the citation to find a comparable document collection; previous related abstract work mainly borrowed the citation network to supplement the abstract content of an article, without considering whether the set of documents it links to is accurately related; the present invention aims at the difference generation task to be solved, ensures the simultaneous existence of similarities and differences, and specifically starts from the citation function of "comparative citation", which contains the summary content of the citing author's profound understanding of the cited document, annotates the citations in the scientific and technological literature, and uses the effective comparative information contained in the comparative citation to find comparable clues and corresponding documents; the present invention collects a set of comparable scientific and technological literature collections by identifying specific functional citations, so that the document collection obtained by the collection is strongly related, describing the same topic, and describing the differences between them from multiple angles, so that the document collection obtained is clearly comparable. The present invention constructs a comparison graph, classifies and discusses the sentence relationships between documents and citations, fully considers the different relationships between sentences, and calculates the sentence relationship scores through an unsupervised method based on the comparison graph without relying on large data set training. The present invention also introduces cross-document structure theory and rhetorical structure theory, and comprehensively considers the different structural characteristics and semantic expressions of sentences in documents from the perspective of multiple documents and within a single document. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Schematic diagram of the process of the present invention;

[0029] Figure 2 Schematic diagram of a comparison diagram of the method of the present invention;

[0030] Figure 3 Schematic diagram of a generation case of the method of the present invention. DETAILED DESCRIPTION

[0031] In order to make the contents of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments in conjunction with the accompanying drawings.

[0032] The present invention provides a method for generating differences in scientific and technological literature based on citations, graphs and structures, such as Figure 1 shown.

[0033] First, we comprehensively examine classic citation function classification methods (such as "Automatic classification of citation function," which focuses on practical clues, and "Purpose and Polarity of Citation: Towards NLP-based Bibliometrics," which focuses on intention and sentiment). We then reclassify, map, and define functional categories. Citations originally belonging to other functional categories that exhibit a certain degree of contrasting clues are reclassified into the contrasting category. For example, "Automatic classification of citation function" considers improvements based on cited literature methods to be "modified citations," meaning that the citing author improves on the cited literature's methods; and "basic citations," meaning that the citing author expands on the cited literature's theories. Our annotation method considers both "improvement" and "expansion" to mean "alterations" within the same topic, meaning that the citing and cited literature contain both commonalities and differences, thus being considered "contrastive citations." Consequently, we propose a citation function annotation method that is sensitive to "contrastive citations."

[0034] Unlike previous annotation methods, the method proposed in the present invention that is sensitive to capturing "contrast citations" is mainly targeted at "contrast citations" and has a functional bias. It can conduct a more comprehensive and in-depth search of such citations based on the needs of contrast clues, and dig out citations with weak contrast clues but still existing. As shown in Table 1 below, the annotated citation categories include: "neutral citations", "positive citations", "negative citations", "future citations" and "contrast citations". Among them, "neutral citations" refer to the citing author's ordinary description of the cited document; "positive citations" refer to the citing author's positive description of the cited document; "negative citations" refer to the citing author's negative description of the cited document; "future citations" refer to the citing author's establishment of a working hypothesis on the cited document; "contrast citations" refer to the citing author's comparative description of the cited document; this citation function annotation method is easier to capture "contrast citations" with contrast clues than previous methods.

[0035] Table 1

[0036]

[0037] by Figure 3 Taking the citing documents in as an example, this annotation method is used to perform functional annotation on citations of scientific and technological literature.

[0038] Secondly, collect comparative citations and corresponding literature from the marked citations. Figure 3Take the comparative citations shown in the upper part as an example. This citation comes from the citing article [1] (Pre-training of deep bidirectional transformers for language understanding), which cites the cited article [2] (Deepcontextualized word representations) and the cited article [3] (Improving language understanding with unsupervised learning). From the citation content, we can see that the citing article [1] and the cited articles [2][3] describe the relevant content of the topic of "language representation model" from different perspectives; and from the word "unlike", we can see that there are differences between the three articles. Therefore, these three articles constitute a comparable document set.

[0039] Afterwards, based on the obtained document collection, the corresponding Figure 2 The comparison graph shown in . The subgraph represents each document or citation, the node represents the sentence in the document or citation, and belongs to its corresponding subgraph. The edge between the nodes represents the relationship between the sentences, including the relationship between sentences within the document, the relationship between sentences between documents, and the relationship between sentences related to the citation. The prominence, difference, and commonality embodied by the sentences are measured through these three relationships. After that, the relationship scores of different sentences are calculated, and the vector representation of the sentence is obtained by TF-IDF. The sentence relationship score is measured by the dot product similarity value of the vector between the sentences, and then the sentence relationship score is further calculated using the similarity value. Specifically, for sentence A and sentence B, assuming that the vector of sentence A is [1 1 0 21] and the vector of sentence B is [0 12 1 1], the similarity value between sentence A and sentence B is recorded as:

[0040] Similar(A,B)=[1 0 0 2 1]*[0 12 1 1] T =3;

[0041] The prominence score within the document is divided into two parts. If a sentence has a higher similarity with the previous sentence, it indicates that there is redundant information, which will have a negative impact on its score. If it has a higher similarity with the following sentence, it means that it contains key and prominent content, which will have a positive impact on its score. For sentences in the document, the set of sentences before sentence s is denoted as Pre, and the set of sentences after s is denoted as Aft. The prominence score within the document of sentence s is as follows, where Similar(s, Pre) and Similar(s, Aft) represent the sum of the similarity values ​​between sentence s and all sentences in Pre, and the sum of the similarity values ​​between sentence s and all sentences in Aft, respectively. In addition, α+β=1, and α,β∈[0,1];

[0042] SAL(s)=αSimilar(s,Aft)-βSimilar(s,Pre);

[0043] Inter-document difference score: For a set of sentences in the citing document CP and a set of sentences in the cited document RP, the inter-document difference score of a sentence cp in the citing document is as follows: Similar(cp,RP) represents the sum of the similarity values ​​between sentence cp and all sentences in RP;

[0044] DIF(cp)=-Similar(cp,RP);

[0045] The inter-document difference score for a sentence rp in the cited document is as follows: Similar(rp,CP) represents the sum of the similarity values ​​between the sentence rp and all the sentences in CP;

[0046] DIF(rp)=-Similar(rp,CP);

[0047] Citation-related commonality score: For a sentence s in a document and a set of citation sentences CIT, the citation-related commonality score of sentence s is as follows: Similar(s, CIT) represents the sum of the similarity values ​​between sentence s and all sentences in CIT;

[0048] COM(s)=-Similar(s,CIT);

[0049] The three sentence relation scores are linearly summed to represent the comprehensive sentence relation score of a sentence;

[0050] RELATIONSHIP SCORE(s) =SAL(s)+COM(s)+DIF(s);

[0051] Next, we introduce the cross-document and rhetorical structure theory, as shown in Table 2 below, to calculate sentence structure scores, capturing the semantic meaning of sentences across and within documents. First, we assign scores to each category based on importance, such as 5 points for "contrast" cross-document structure and 5 points for "contrast" rhetorical structure. Different structural types convey different semantic meanings. We then label the sentences in the document with their categories, and finally calculate the structural score for each sentence as follows:

[0052] STRUCTURE_SCORE(s)=CST(s)+RST(s);

[0053] Table 2

[0054]

[0055] Finally, the comprehensive scores of the sentences are used to sort the sentences and extract the sentences to form different contents;

[0056] SCORE(s)=RELATIONSHIP_SCORE(s)+STRUCTURE_SCORE(s).

[0057] The difference content is as follows Figure 3 As shown in the lower part, the generated content contains both similar content, namely "language representation model", and explains the differences, namely the different methodological information about language representation model in the latter part extracted from various literatures.

[0058] A method for generating differential content for scientific literature based on citation function, contrast graphs, and structural theory introduces citation function into differential content generation for the first time. It uses comparative citations to identify comparable and reliable literature, measures the relationships between sentences based on a constructed contrast graph, and incorporates structural theory to analyze sentence semantics from multiple dimensions. This unsupervised approach does not rely on large datasets and has excellent practicality for generating differential content.

[0059] The above description is only a preferred embodiment of the present invention and is not intended to further limit the present invention. All equivalent changes made using the contents of the present invention description and drawings are within the scope of protection of the present invention.

Claims

1. A method for generating differences in scientific literature based on citations, graphs, and structures, characterized in that: The method steps are: Step 1: Integrate citation function annotation methods, design a citation function annotation method that is sensitive to comparative citations, and perform functional annotation on citations in the literature; Step 2: Select comparative citations from the marked citations and collect the corresponding literature; Step 3: Construct a comparison chart of the document collection and calculate the sentence relationship scores in the documents; Step 4: Calculate the sentence structure scores of the documents using structural theory. We introduce two structural theories, cross-document and rhetoric, to expand the semantic structure of sentences to multiple dimensions for analysis. The cross-document structural theory focuses on the sentence structure relationships between documents, while the rhetoric structural theory focuses on the sentence structure relationships within a document. Step 5: Sentences are sorted and extracted by comprehensively considering the sentence relationship scores and sentence structure scores in the literature to ultimately form the target difference content; Among them, step 3 is specifically as follows: Step 3-1: Construct a comparative graph of the document collection. This graph contains a document or citation subgraph, sentence nodes, and three types of inter-sentence relationship edges. The document or citation subgraph represents the document or citation content itself, the sentence nodes represent the sentences in the document or citation, and the three types of inter-sentence relationship edges represent three types of relationships between sentences: intra-document sentence relationships, inter-document sentence relationships, and citation-related sentence relationships. Intra-document sentence relationships consider sentences within a document to capture the prominence within the document; inter-document sentence relationships consider sentences between documents to capture the differences between documents; and citation-related sentence relationships consider the similarity between document sentences and citations to capture the relevance between documents. Step 3-2: Based on the comparison graph of the constructed document collection, a graph method is used to calculate the sentence relationship score; a sentence contains three relationship scores, and the sentence vector is obtained, and then similarity is calculated to measure different relationship scores; Score within the document: Score the sentences in the document with S S The subscript s indicates the position in the document; the set of sentences preceding it is denoted as Pre, S pre Represents the sentences in the set Pre, and the subscript pre represents the position in Pre; the set of sentences following it is denoted as Aft, S aft Represents a sentence in the set Aft, and the subscript aft represents the position in Aft; the coefficients α and β are used to represent the coefficients when calculating the similarity with the previous and next sentences, and α, β∈[0,1]; the sentence S S The intra-document prominence score of is as follows; Inter-document score: The citing document is the document in which the citation appears, and the cited document is the document cited by the citation. The sentence in the citing document CP is scored with S cp The subscript cp indicates the position of the sentence in the citing document; the sentence in the cited document set RPs is represented by S rp The subscript rp indicates the position of the sentence in the cited literature set; the sentence S cp The difference score between the documents is as follows: Citation relevance score: Sentences in the literature are scored with S p Indicates that the subscript p indicates the position in the document; the sentence in the citation CIT is marked with S cit The subscript cit indicates the position in the quotation; the sentence S p The citation-related commonality score is as follows; 2. The method for generating differences in scientific and technological literature based on citations, graphs, and structures according to claim 1, characterized in that: In step 1, the classic citation function annotation method is integrated to reclassify, map and define the functional categories, and citations with a certain degree of contrast clues that originally belonged to other functional categories are reclassified into the contrast category, and then a citation function annotation method that is sensitive to capturing contrasting citations is proposed.

3. The method for generating differences in scientific and technological literature based on citations, graphs, and structures according to claim 1, characterized in that: In step 2, only comparative citations are collected from the marked citation set, and citing documents and cited documents corresponding to the comparative citations are collected to form a comparable document set.

4. The method for generating differences in scientific and technological literature based on citations, graphs, and structures according to claim 1, characterized in that: Step 4 is as follows: First, a score is preset for each structural relationship. If the cross-document or rhetorical structural relationship being compared is a sentence structure type that the present invention focuses on, a high score is given. Then, the cross-document structure types and rhetorical structure types of sentences in the literature were classified respectively, and the two structural theory scores of each sentence were obtained according to the obtained structural relationship categories and referring to their corresponding preset scores.

5. The method for generating differences in scientific and technological literature based on citations, graphs, and structures according to claim 1, characterized in that: In step 5, the relationship scores and structure scores of the sentences in each document are comprehensively considered, and the sentences are sorted in descending order according to the scores. Then, sentences are selected from each document in descending order according to the preset number of sentences and scores to finally constitute the difference content.

Images