本論文では,入力文と正解事例の類似度計算を行い,類似度最大の正解事例に対応する動詞語義及び名詞意味役割を付与する手法を提案する.これは正解事例が少ない場合,統計的学習モデルでは十分な精度が得られないためである.本システムは慣用句同定や複合名詞内解析システムと連携可能な枠組みで開発している.またCRFとの比較実験を行い,その結果について報告する.%In this paper, we propose a method to identify semantic role labels utilizing a few correctly label-annotated examples. The similarity between input sentence and correct labeled example sentences is evaluated by manually defined similarity function. This is because a statistical learning model doesn't work well without many correct examlpes. We develop Japanese semantic role label identification system based on the proposal method. Moreover, this system can be added deverbal compound analyzer and idiom identification system. We also report the comparison between the proposal method and CRFs in the experiment.
展开▼