MulTiple: A Multi-answer Time-sensitive Complex Question Answering Dataset

Anonymous

Abstract: Time-sensitive question answering is to answer questions with specific timestamps from the given long document. Existing works mostly focus on only one of the high-quality answers, but it is common that multiple answers simultaneously satisfy the constraints of a specific timestamp in the time-sensitive question. For example, an individual may play two different roles during a specific timestamp. In this paper, we construct a Multi-answer Time-sensitive question answering dataset, MulTiple, consisting of 17,580 multi-answer instances. Each contains a question, a corresponding long document and multiple answers. To ensure that the generated questions have multiple answers, we propose a global iteration method to obtain time-evolving events with multiple objects for the same subject and relation. Moreover, the baseline model IterBird is proposed to progressively gain multiple answers by integrating an iterative mechanism with the single-answer model. We construct extensive experiments on MulTiple and results show that IterBird significantly outperforms other baselines with SEM scores of 25.65\% and 22.69\%. It demonstrates that existing models struggle to obtain the full answers, even as clue words are provided in the time-sensitive questions. The dataset and code are released in \url{http://github.com/multipledata/MTQA}.

Paper Type: long

Research Area: Question Answering

Contribution Types: Data resources

Languages Studied: English