Authors:
Kong Yuntao
;
Nguyen Minh Phuong
;
Teeradaj Racharak
;
Tung Le
and
Nguyen Le Minh
Affiliation:
School of Information Science, Japan Advanced Institute of Science and Technology, Nomi, Japan
Keyword(s):
Question Answering, Multi-hop QA, Two-step Tuning, Transfer Learning, Multi-task Learning.
Abstract:
Multi-hop question answering (QA) requires a model to aggregate information from multiple paragraphs to predict the answer. Recent research on multi-hop QA has attempted this task by utilizing graph neural networks (GNNs) with sophisticated graph structures. While such models can achieve good performance, their computation is rather expensive. In this paper, we explore an alternative method that leverages a single-hop QA model to deal with multi-hop questions. Our system called ‘Answer Multi-hop questions by Single-hop QA’ (AMS) consists of three main parts that first filter a document and then conduct prediction using the attention-based single-hop QA model with multi-task learning. Specifically, AMS is constructed based on the co-attention and self-attention architecture. Lastly, consider that BERT-based model is pre-trained in a general domain and the data distribution can be different from multi-hop QA task. We propose two-step tuning mechanism to enhance the model’s performance,
which is based on transfer learning from other QA datasets. To verify AMS effectiveness, we consider the previous state-of-the-art Hierarchical Graph Network (HGN) with the same document filter as our baseline. Experiments on HotpotQA show that AMS can outperform HGN by 1.78 points and 0.56 points for Joint EM and Joint F1, respectively. Meanwhile, it has smaller model’s size and uses less computational resource. We also experiment with other GNN-based models and achieve better results.
(More)