The huge growth in scientific publications makes it difficult for researchers to keep track of new research even in narrow sub-fields. While an abstract is a traditional way to present a high level view of the paper, recently it is getting supplemented with research highlights that explicitly identify the important findings in the paper. In this poster, we aim to automatically construct research highlights given the abstract of a paper. We use deep neural network-based models for this purpose and achieve high ROUGE and METEOR scores on a large corpus of computer science papers.
The count of scientific publications doubles roughly every 9 years [10], making it hard for researchers to track even their own fields. One recent trend is to provide research highlights -a bulleted list of the main contributions of the paper -along with the abstract and the main text. They are potentially easier to read than abstracts, especially on mobile devices, and focus more on findings than on background. Additionally research highlights could be useful for other tasks like finding surrogates for access-restricted papers [5,7] and keyphrase extraction [6]. We use a pointer-generator network with coverage mechanism to automatically generate highlights given the abstract of a research paper. Distinct from a prior work [2] that classifies sentences in the full text as highlights or not, our focus is on generation of highlights.
We use a dataset released by Collins et al. [2] containing URLs of 10142 computer science publications from ScienceDirect 1 . Each 1 https://www.sciencedirect.com/ EEKE '21, September 30, 2021, Online example in the dataset is organized as (abstract, author-written research highlights): 8115 pairs are used for training, 1014 pairs for validation and 1013 pairs for testing. In this dataset, the average abstract size is 186 words while that of highlights is 52; for 98% of the papers, highlights are 1.5 times or more shorter than the abstract.
We have used three deep learning-based models to generate research highlights. Model 1 is the sequence-to-sequence (seq2seq) model with attention [3]. Each abstract is tokenized and the tokens are converted to 128-dimensional GloVe vectors [4] that are sequentially fed into the encoder which is a single-layer bidirectional Long Short-Term Memory (BiLSTM). The decoder is a single-layer unidirectional LSTM. The model uses neural attention [1] to attend to the words in the source document while generating the target words for the summary. Model 2 is a pointer-generator network [8], which augments the above seq2seq model with a special copying mechanism. When generating words, the decoder probabilistically decides between generating new words from the vocabulary (i.e. from the training corpus) and copying words from the input abstract (by sampling from the attention distribution). While the generator helps in novel paraphrasing, copying helps to tackle out-of-vocabulary (OOV) words. Model 3 augments the second model with coverage mechanism of Tu et al. [9] to avoid erroneously repeating the same words during decode. For all the models, we used the same vocabulary of around 50K tokens, beam search in the decoder with size 4, maximum input size of 400 tokens and maximum output size of 100 tokens.
Results are shown in Table 1
We applied three different deep neural models to generate research highlights from the abstract of a research paper. The pointer-generator network with coverage mechanism achieved the best performance.
But the predicted research highlights are not yet perfect. A simple post-processing operation could be to remove sentences that contain entities that are absent in the given abstract. We are currently exploring this and other techniques to improve the system.
This work is supported by research grant from Department of Science and Technology, Government of India at Indian Association for the Cultivation of Science, Kolkata and National Digital Library of India Project sponsored by the Ministry of Education, Government of India at IIT Kharagpur.