Explicit History Selection for Conversational Question Answering

Abstract

Topic shift is very common in multi-turn dialogues, making it a great challenge in the filed of conversational question answering. Existing methods usually select the most adjacent turns as history information, however, it is useless or even harmful in case of topic shift. This paper proposes two explicit history selection models: SHSM and DHSM, to address this issue. The former is a simple history selection model, which only selects <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\boldsymbol{k}$</tex> previous history turns; and the latter is a dependent history selection model, which selects the most relevant <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\boldsymbol{k}$</tex> history turns through a turn-dependent graph. The two models are then trained in a consistency framework. Experimental results on QuAC show that our model can cope with topic shift problem, and it outperforms existing state-of-the-art methods by 0.8 on <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\boldsymbol{F}_{\mathbf{1}}$</tex> score, 0.7 on HEQ-Q score, and 1.4 on HEQ-D score.