|吉川 克正||Ｄ２||松本 裕治||関 浩之||新保 仁||浅原 正幸|
発表題目： Document Level Event and Argument Analysis with Probabilistic Logic
発表概要：Events and arguments create the main structure of a document and have important information for document understanding. I focus on the three types of relations between, event-event, event-argument, and argument-argument. My final goal is to propose a new method to extract the main structure considering these relations simultaneously. Probabilistic Logic enables us to use flexible constraints and easily implement human's linguistic knowledge coorporating with machine learning models. For previous work, I have been already engaged in temporal relations (event-event) and event-argument relations with Markov Logic, a popular probabilistic logic approach. Now I am tackling Japanese event-argument relation extraction. In Japanese, we have many zero-anaphoric cases and some event-arguments are crossing sentence boundaries often difficult to identify. I experiment on a basic setting with intra-sentential constraints and achieve a state-of-the-art results. I also try to identify cross-sentential event-arguments with zero-anaphora resolution.