自然語言推理之後設可解釋性建模

Abstract

本研究之主軸為利用注意力機制，在自然語言推理任務上，以自然語言形式之解釋賦予模型可解釋性，並進行人類信任度評估。近年來人工智慧系統的可解釋性逐漸受到重視，可解釋性使開發者、相關開發人員及終端使用者能夠了解人工智慧系統，進而得以設計更完備的系統、產品以及更符合使用者需求的應用產品。注意力機制做為系統原生的可解釋機制，能夠同時提供忠實且合理的解釋，目前於注意力機制可解釋性之研究，多以注意力權重進行視覺化的呈現來賦予模型決策可解釋性，然而在一般互動的情境中，解釋多是以自然語言的方式表達。而在可解釋性的評估部分，目前所採用的評估方式甚少加入終端使用者—人類進行評估；若有，其評估方式之完備性也難以為人工智慧系統之應用部署提供洞見。\n\n本研究利用 Transformer 架構模型之注意力機制，以自然語言之方式呈現其解釋，賦予模型可解釋性；同時探討在提供不同任務知識後，對於此方法之解釋有何影響；最後以不同模型之解釋進行人類信任度之評估，分析人類對於解釋之信任及偏好。實驗顯示，在自然語言推理任務上，模型之效能與注意力關注區間確實相關；在加入不同特性之任務知識後，模型的解釋能夠忠實地呈現其訓練任務之特性；最後在人類信任度上，人類對於解釋方式偏好不盡相同，但是長而資訊豐富的解釋方式，較短而精確的解釋方式來得有優勢。

The explainability of artificial intelligence (AI) model has recently attracted much interest from the researchers. Explainability provides developers, stakeholders and end users with a better understanding of how the model works and can assist in better interaction between human and machine. Attention mechanism, as an intrinsic explainable method, is considered more suitable for faithful and plausible explanations. The majority of research on attention mechanism, however, focuses on visualization of the attention weight as a way to make the model explainable. Yet in real-life interactions, explanations are more likely presented in natural language. Furthermore, while evaluating model explainability, little research has taken human responses into consideration or included measurement of human reactions. The void of human-related research led to absence of useful insights to develop and deploy AI applications.\n\nThis research employs natural language inference paradigm, using transformer-based attention weight to provide explanations of the task performance of the model. After the training, we also evaluate human trust and preference towards the explanation provided by different models. The results indicate that in natural language inference tasks, the model performance, and , long, contextual explanations are more advantageous than short, concise explanation in gaining human trust.