The goal of this paper is to describe a method to automatically extract all basic attributes namely actor, action, object, time and location which belong to an activity, in each sentence retrieved from Twitter. Previous work had some limitations, such as inability of extracting infrequent activities, high setup cost, inability of extracting all attributes. To resolve these problems, this paper proposes a novel approach that treats the activity extraction as a sequence labeling problem, and automatically makes its own training data. This approach can extract infrequent activities, and has advantages such as domain-independence, scalability, and unnecessary hand-tagged data.%本論文の目的は,twitterから取得した文中に現れる行動の基本属性(行動主,動作,対象,時間,場所)を自動的に抽出することである.しかし,先行研究では,頻度が低い行動を獲得できない.そして,抽出する前に,動詞リストとカテゴリワード(対象を表すワード)を予め準備しておく必要がある.そこで本論文では,条件付確率場(ConditionalRandom Fields)と自己教師あり学習(Self-Supervised Learning)を用いて,行動属性の自動抽出手法を提案する.提案手法では,人手でラベル編集や行動のドメインの定義などの必要がなく,頻度が低い行動も獲得できる.
展开▼