Focus on the Action: Learning to Highlight and Summarize Jointly for Email To-Do Items Summarization

Findings of the Association for Computational Linguistics: ACL 2022, pages 4095 - 4106 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Focus on the Action: Learning to Highlight and Summarize Jointly for Email To-Do Items Summarization Kexun Zhang1, Jiaao Chen2, Diyi Yang2 1Zhejiang University 2Georgia Institute of Technology kexunz@zju.edu.cn jiaaochen@gatech.edu diyi.yang@cc.gatech.edu Abstract Automatic email to-do item generation is the task of generating to-do items from a given email to help people overview emails and schedule daily work. Different from prior re- search on email summarization, to-do item gen- eration focuses on generating action mentions to provide more structured summaries of email text. Prior work either requires large amount of annotation for key sentences with potential ac- tions or fails to pay attention to nuanced actions from these unstructured emails, and thus often lead to unfaithful summaries. To fill these gaps, we propose a simple and effective learning to highlight and summarize framework (LHS) to learn to identify the most salient text and ac- tions, and incorporate these structured repre- sentations to generate more faithful to-do items. Experiments show that our LHS model outper- forms the baselines and achieves the state-of- the-art performance in terms of both quanti- tative evaluation and human judgement. We also discussed specific challenges that current models faced with email to-do summarization. 1 Introduction Automatic email to-do generation is the task of summarizing to-do items from given emails (Mukherjee et al., 2020) to help people overview overwhelming numbers of emails they receive ev- ery day (Radicati and Hoang, 2011) and schedule their daily work. Unlike prior research on emails such as generating email conversation summariza- tion (Muresan et al., 2001; Nenkova and Bagga, 2003; Rambow et al., 2004), keyword extraction (Turney, 2000; Lahiri et al., 2016; Lin et al., 2017) and subject line generation (Zhang and Tetreault, 2019; Xue et al., 2020), to-do items generation fo- cuses more on action mentions to provide more structured summaries from email communications, which requires identifying important tasks to be performed among all the action items and aligning them with the right users (Mukherjee et al., 2020). To tackle these challenges in to-do item genera- tion, a two-stage framework (Zhang and Tetreault, 2019; Mukherjee et al., 2020) is generally utilized to first extract the commitment sentence which is the most related to the to-do item and then gener- ate to-do items based on those selected sentences. Despite the effectiveness, several limitations ex- ist: (1) extra annotations are usually needed for the first stage (Mukherjee et al., 2020) to learn the classifiers which require extra cost/expertise and are often hard to obtain in low-resourced set- tings; (2) any information loss in the identification stage might lead to bigger noises in the to-do gen- eration stage; (3) directly extracting actions from less-structured text usually leads to unfaithful sum- maries (Chen and Yang, 2021) that mismatch the relations between users and actions. To fill in these gaps, in this work, we propose a learning to highlight and summarize model, where we learn the important …

Content truncated to 500 words