Integrating Hierarchical Semantic into Iterative Generation Model for Entailment Tree Explanation

TL;DR

This work proposes an architecture of integrating the Hierarchical Semantics of sentences under the framework of Controller-Generator (HiSCG) to explain answers and is the first to notice hierarchical semantics of sentences between the same layer and adjacent layers to yield improvements.

Abstract

Manifestly and logically displaying the line of reasoning from evidence to answer is significant to explainable question answering (QA). The entailment tree exhibits the lines structurally, which is different from the self-explanation principle in large-scale language models. Existing methods rarely consider the semantic association of sentences between and within hierarchies within the tree structure, which is prone to apparent mistakes in combinations. In this work, we propose an architecture of integrating the Hierarchical Semantics of sentences under the framework of Controller-Generator (HiSCG) to explain answers. The HiSCG designs a hierarchical mapping between hypotheses and facts, discriminates the facts involved in tree constructions, and optimizes single-step entailments. To the best of our knowledge, We are the first to notice hierarchical semantics of sentences between the same layer and adjacent layers to yield improvements. The proposed method achieves comparable performance on all three settings of the EntailmentBank dataset. The generalization results on two out-of-domain datasets also demonstrate the effectiveness of our method.

Figures (7)

• Figure 1: The task of explaining a hypothesis from premises with an entailment tree. A hypothesis is comprised of a question and an answer. Premises (blue) are obtained from the provided corpus (or given directly) with distractors (grey). Premises act as leaf nodes in the entailment tree which contains intermediate nodes (purple) and hypothesis as the root node (red). The values in parentheses are calculated by Sentence-BERTreimers-gurevych-2019-sentence using cosine similarity between facts and hypotheses.
• Figure 2: The framework of our proposed model HisCG.
• Figure 3: The training details of the Hierarchical Semantic Encoder.
• Figure 4: Use t-sne to visualize the example in Fig \ref{['question_describe']}. Red nodes represent hypotheses, blue nodes are relevant facts, and yellow nodes are distractors.The representation distributions are from w/o hierarchical semantics encoder (left) and w hierarchical semantics encoder (right).
• Figure 5: The result of Task1 test set in terms of the number of leaves. The value on bars indicates the proportion of entailment trees containing a certain number of leaves that are correctly predicted by HisCG.