We can complete the acquisition, dissemination, and communication of knowledge through scientific papers. The exponential growth of electronic scientific papers has made finding and selecting them difficult. Modeling and annotating the rhetorical structure of scientific articles can improve the efficiency of searching and reading [10]. On one hand, it can help search engines to quickly retrieve insight into the core of scientific article [9]; on the other hand, it can help the reader quickly browse and understand those articles. The study of the rhetorical structure and annotation of discourse has long-standing traditions. Automatic annotation of full articles under the existing stateof-the-art is difficult to achieve [3]. The abstract of scientific articles is briefer than full articles, which makes the annotation possible.
There are numbers of well-known approaches to modeling the rhetorical structure of publications, such as Harmsze [8], ABCDE (Annotations,Background, Contribution, Discussion and Entities) [13], and SALT(Semantically Annotated L A T E X) [7]. Based on the previous approaches, we provide a new model for abstract semantic analysis that includes the background, research problem, solution, and result. The model includes a preliminary verification by means of data crawling from the Internet for testing purposes. According to the methods for machine annotation of scientific discourse, Argumentative zoning [11], XIP(Xerox incremental parser ) [1], and SemTag [4], this paper use linguistic clues and location information to annotate the scientific publication's abstract automatically.
The predecessors of semantic technology have made many contributions for semantic annotated structure, which can improve the efficiency of publications' searching and reading [10]. There are well-known approaches to modeling the rhetorical structure of publications. Harmsze proposed one of the first and the most comprehensive models for extracting the rhetoric and argumentation within scientific papers. This model focused on developing a modular representation for the creation and evaluation of scientific publications [8].De Waard and Tel introduced a different model for representation of discourse called ABCDE, which developed a L A T E X style sheet to identify five components in a discourse [13]. They proposed finer-grained annotation to complement these structures and relationship types [3]. A semantic authoring framework to enrich scientific publications with semantic metadata was called SALT, offering an improved coarse-grained rhetorical structure and a fine-grained semantic network [7].
After modeling the rhetorical structure of publications, we proceed to the automatic annotating of the scientific discourse. The first attempt to automatically annotate rhetorical expressions in research papers is called argumentative zoning [11]. XIP detect rhetorical expressions from language uses of the authors, targeting salient sentences within scientific articles [1]. Both models use clear linguistic clues to annotate the scientific articles. SemTag is a system offering an automatic ontology of semantic information, which identifies the candidate instance's keywords needed annotation. The system is based on TAP, a knowledge base from Stanford University which constructs two text vectors -context (before and after each 10 words) and candidate instance-calculating similarity and selecting best matches [4]. The automatic annotation of research papers should capture and represent the evolution of ideas and findings that authors described in the articles [12]. The main line of the above research aims at extracting factual information from the texts of the articles and transforming them into structured data [2] [6]. Semantic structure models and machine annotation models are used for full scientific publications, but the abstracts of papers contain more standardized semantic structures and rhetoric [5]. Some journals like Nature provide a constant structure for their papers, which makes automatic annotation of abstracts possible.
Based on the models for full publications mentioned in Section 2-Harmsze, ABCDE, SALT-we provide a new model for abstracts that includes background, research problem, solution, and result. We use linguistic clues extracted from abstract annotation and position information to annotate it automatically.
The flowchart for our system is in Figure 1.
The system is divided into the following seven main phases: Phase 1: Data Acquisition Module The four semantic tags were added manually to every sentence of scientific papers' abstracts, which includes background, research problem, solution, and result.
Phase 3: Data Preprocessing Module This module cleaned the text of publications' abstracts, such as removing numbers and punctuation, reducing word roots, converting all words to lower case, and so on.
Phase 4: Features Selection and Weighted Module In this module, TF-IDF algorithm was used to manage the numerous and complicated information of texts, which helped us discover key words that are more important and appeared more frequently in one text.
Phase 5: Sentences Segmentation and Preprocessing Module This module was used to segmented the abstract text by period, cleaned the sentences as in Phase 3, and stored them as feature vectors with an ID, which described the sentences' positions in the paper's abstract.
Phase 6: Similarity Calculation Module This module acquired the similarity level of two feature vector quantities by cosine similarity algorithm, which calculated the value of the included angle of two vectors.
The automatic annotation the semantics of the paper,by linguistic clues(They propose...) and position information(the position of the sentence).
The results of TF-IDF revealed the information of words as well as the texts in the form of a matrix, and every word had a TF-IDF weight value in each text, which presented the significance of the word in that text. The calculation of the weighted value became the core of the arithmetic.
TF-IDF algorithm was divided into the concepts TF and IDF. TF(Term frequency) reflects that the words with higher rates will attain a higher TF value. The IDF is the acronym for inverse document frequency. Compared with these words in common use, the ones with a high frequency acquired a higher IDF value, which did not often appear in other texts. As a result, the TF-IDF weight value was calculated with the formula TF*IDF. The key words with high values were chosen according to the integrated conditions of their TF and IDF.
The weighted feature words of the research problem are in Table 1.
We calculated the similarity using the cosine formula as follows:
(A=[A1,A2,...,An],B=[B1,B2,...,Bn])
Coupled with the growth of the value, the similarity level of two vectors was reduced. In this way, we were able to learn whether the two vectors were similar. In our system, every text in the TF-IDF matrix was considered a word vector, and the test word-frequency vector was another. We used this arithmetic to reveal the similarity level of the two texts. Table 2 revealed the similarity between the sentence and tags that we proposed in the system.
Similarity was calculated by featuring words in Phase 6. We found that the difficult point was how to select the sentences from the abstract, which should add the research problem or solution tag. We discovered that the label of the sentence was relevant to their position. We processed information of abstract location using 88 articles from DMKD. The first one to two sentences were the background, the next one to two sentences were research questions, and then the solution was discussed in the next two to six sentences, and the final one to two sentences were the result. Different tags were labeled according to the location of the sentence. For example, for the third sentence mentioned in Phase 6, the system annotated the solution label to it because it is the fifth sentence of the paper's abstract and the sentence in front of it has been marked as the research problem.
This paper studies how to annotate semantic structure to the abstract of scientific papers automatically, based on real data from DBLP DB. We developed a small experimental system in which the abstract text is the system's input and the abstract with four labels is the output. The system's interface is as shown in Figure 2. Eighty articles were used to test the model proposed, the test results were as Table 3: This paper established a model to annotate the semantic structure for scientific papers' abstracts automatically. The first step of the process is to build a rhetorical structure that includes background, research problem, solution, and result. Then, through feature extraction and calculating weight for every segment of an abstract, we attain the feature vector. Next, according to the cosine algorithm, the system computes the score of the similarity. Last, the abstract was annotated by similarity and sentence positions.
The correct rate the system proposed in this paper could only reach about 67.5% of the automatic annotation of the semantic structure, and the effect will be better if the ontology system is introduced in the future, and the rhetorical words of scientific articles in different journals and different areas may be different; therefore, the next focus is to improve the accuracy and create a cross-domain model.
This work is supported by the National Natural Science Foundation of China (No. 61300147), China Postdoctoral Science Foundation (No. 2014M551185), and Science and Technology Program of Changchun (No. 14GH014).