基於自然語言解釋之可信賴式模型

Abstract

各種深度學習模型，例如預訓練語言模型，已大幅推進了自然語言處理領域的發展。這些模型雖然帶來了驚艷的進步，其複雜的結構往往犧牲了可解釋性，並影響了人們對它們的信賴程度。為了實現具可解釋性的自然語言處理研究，發展能提供自然語言解釋的可信賴式模型開始在近期受到關注。然而，目前大多數開發此種可信賴式模型的方法都需要大量人工標註之自然語言解釋。其過程成本昂貴、需要大量資源，特別是在涉及專業知識的領域上。 在本研究中，我們提出了兩個方向來應對上述挑戰：（1）提升可信賴式自然語言處模型在少樣本下之性能。（2）運用既有之現實世界資料構建具專業知識的自然語言解釋數據集。在前者，我們通過深入分析自然語言解釋與預測之終端任務標籤之間的關係，展示了如何利用無標註的「解釋-答案對」來建立偽平行數據，並基於其提出了一個名為「ZARA」的新框架，能透過自我訓練提升可信賴式模型在少樣本下的性能。實驗結果顯示，對於任務準確性和解釋品質上我們的方法在四個FEB基準資料集任務上都展現了顯著的進步。 對於後者，我們選擇需要高度專業知識的醫學領域進行研究，運用線上電子e院服務的問答資料，我們構建了一個大規模數據集，其任務目標為生成具自然語言解釋之相關建議。基於此數據集，我們提出了一個具語境感知機制來促進生成之建議和解釋之間對齊性的可信賴式模型。在定性評估及專業人員之人工評估中，我們的模型在建議生成和自然語言解釋生成上都展現了極具潛力的結果。

Deep neural models such as pre-trained language models (PLMs) have advanced the field of natural language processing (NLP) significantly. Yet, the complex structures of these models often come at the cost of interpretability. Towards trustworthy, explainble NLP, self-rationalization models which aim to provide free-text explanations (FTEs)---also known as rationales---are gaining attention recently. However, most existing approaches for developing self-rationalization models require abundant human-annotated FTEs, which are expensive and resource-intensive to collect, especially for domains that involve expert knowledge. In this work, we present two directions to address the above-mentioned challenges. (1) Improving few-shot self-rationalization. (2) Construction of expert-knowledge FTEs leveraging existing real-world data. For the former, we show how to leverage unlabeled rationale-answer pairs to build pseudo-parallel data by an in-depth analysis of the relationship between explanations and end-task labels, and propose ZARA, a novel framework which boosts the performance of few-shot self-rationalization via self-training. Experiments on the FEB benchmark show our approach demonstrates significant improvements on both the task accuracy and the explanation metric across four datasets. For the latter, we focus on the medical domain and present a large-scale dataset constructed from an online e-Hospital service, for the task of medical suggestion generation with FTEs. In addition, we propose a self-rationalization model with a discourse-aware mechanism promoting the alignment between the generated suggestions and explanations. Quantitative metrics and qualitative evaluation by physicians show our proposed model achieves promising performance on both suggestion and explanation generations.