A Peek Into the Memory of T5: Investigating the Factual Knowledge Memory in a Closed-Book QA Setting and Finding Responsible Parts

抄録

Recent research shows that Transformer-based language models (LMs) store considerable factual knowledge from the unstructured text datasets on which they are pre-trained. The existence and amount of such knowledge have been investigated by probing pre-trained Transformers to answer questions without accessing any external context or knowledge (also called closed-book question answering (QA)). However, this factual knowledge is spread over the parameters inexplicably. The parts of the model most responsible for finding an answer only from a question are unclear. This study aims to understand which parts are responsible for the Transformer-based T5 reaching an answer in a closed-book QA setting. Furthermore, we introduce a head importance scoring method and compare it with other methods on three datasets. We investigate important parts by looking inside the attention heads in a novel manner. We also investigate why some heads are more critical than others and suggest a good identification approach. We demonstrate that some model parts are more important than others in retaining knowledge through a series of pruning experiments. We also investigate the roles of encoder and decoder in a closed-book setting.