生成科學論文之條列式貢獻

Abstract

科學論文的貢獻側重於描述其原創之處和重要價值，對於每個科學研究來說這都可以被認為是其最核心的部分。一個能精確辨認論文貢獻並將其組織為結構化摘要的系統對於輔助自動化處理科學文本和幫助讀者理解等應用具有潛在價值。雖然近期的工作開始致力於與論文貢獻相關的任務的研究中，目前仍缺少高品質的大規模資料集來輔助深度學習模型的訓練。有鑑於此，我們收集並整理了一個資料集，其中包含大約兩萬四千篇計算機科學領域的論文及其作者條列之貢獻，根據我們提出的標記框架，這些科學貢獻又被進一步分為對應的不同類別。接著我們正式定義了生成科學論文之條列式貢獻這個任務。利用大量的無監督資料和原始論文中重要語句以及生成目標所包含的貢獻類別，我們提出了一個細粒度的訓練策略。實驗結果表明我們提出的方法優於具競爭力的基線模型和其他訓練策略，證明了其有效性。 我們也進行了詳細分析以研究我們所提出的資料集和任務的特性及其挑戰之處。

Contributions of scientific papers highlight their novelty and key values, which are essentially the core parts of every research work. Systems that are capable of identifying the contributions of the papers precisely and organizing them into well-structured summaries are valuable in aiding both automatic text processing and human comprehensions. Though recent works have focused more on tasks dealing with the contributions of the scientific documents, there is currently no large-scale dataset with high quality that can facilitate the training of modern deep learning based models. To this end, we curate a dataset consisting of 24K computer science papers with contributions explicitly listed by the authors, which are further classified into different contribution types based on our newly-introduced annotation scheme. Then we formally formulate the task of generating disentangled contributions for scientific documents. We present fine-grained post-training strategy leveraging abundant unsupervised data and the contribution types of both highlight sentences in the source documents and the generation targets. Experimental results show that the proposed method outperforms competitive baselines and other post-training strategies, demonstrating the effectiveness of our approach. Detailed analysis is also conducted to study the characteristics and challenges of our dataset as well as the newly-proposed task.