Approach to transforming training data for improving the title generation performance for scientific texts

Due to the significant increase in the availability of scientific resources and their expansion, the analysis and systematization of scientific documents become an important task of natural language processing. Scientific articles contain much significant diverse information. Besides, their amount is constantly increasing, and tracking actual scientific publications takes a lot of time. Reducing the number of viewed documents and their generalization is possible using special tools for automatic text processing, including text classification, information extraction, and text summarization. As regards the summarization of scientific documents, one of the particular problems is the generation of the title for the scientific paper. Taking into account the large volumes of scientific resources, the title is especially significant. The title accuracy affects the visibility of the paper by the scientific community and therefore the number of prospective readers. Moreover, some recent studies showed that the quality of the paper title influences the number of citations. Despite this, the authors often spend not enough time creating a good title, which makes it noninformative and non-reflecting the content of the article. To overcome this weakness, the methods of automatic title generation for scientific texts can be developed and used. In this work, we propose an approach to improving the quality of title generation for scientific texts. The proposed approach uses training data filtering and generates new training examples. We consider the following steps: 1) determining recall-oriented ROUGE-1 scores between titles and source texts from the training set. These scores show how many words from the title came from the text. Thus, we can conclude the content correspondence between the title and the source text; 2) ranking examples of the training sample by the recall-oriented scores; 3) filtering examples having scores less than the threshold value k (k E [0; 1)); 4) training model for title generation on the filtered training sample; 5) enriching the filtered training sample to the original size with the pseudo examples generated from the trained model. These examples are generated only for examples removed in the previous step. The approach was tested on two English corpora of scientific texts (SciTLDR and arXiv). We used scientific abstracts as a source for text summarization. We evaluated the values of k in the range from 0,3 to 0,9 in increments of 0,1. In most cases, the results showed that the use of a training sample consisting of filtered and pseudo examples increases the performance of the title generation in comparison with the generation using the original training sample. In our experiments, the most preferred values of the threshold k were 0,7 and 0,8. Experiments were conducted using the BART-base model. The author declares no conflicts of interests.

Keywords

Authors

References