Đang chuẩn bị nút TẢI XUỐNG, xin hãy chờ
Tải xuống
This paper describes a fully automatic twostage machine learning architecture that learns temporal relations between pairs of events. The first stage learns the temporal attributes of single event descriptions, such as tense, grammatical aspect, and aspectual class. These imperfect guesses, combined with other linguistic features, are then used in a second stage to classify the temporal relationship between two events. We present both an analysis of our new features and results on the TimeBank Corpus that is 3% higher than previous work that used perfect human tagged features. . | Classifying Temporal Relations Between Events Nathanael Chambers and Shan Wang and Dan Jurafsky Department of Computer Science Stanford University Stanford CA 94305 natec shanwang jurafsky @stanford.edu Abstract This paper describes a fully automatic two-stage machine learning architecture that learns temporal relations between pairs of events. The first stage learns the temporal attributes of single event descriptions such as tense grammatical aspect and aspectual class. These imperfect guesses combined with other linguistic features are then used in a second stage to classify the temporal relationship between two events. We present both an analysis of our new features and results on the TimeBank Corpus that is 3 higher than previous work that used perfect human tagged features. 1 Introduction Temporal information encoded in textual descriptions of events has been of interest since the early days of natural language processing. Lately it has seen renewed interest as Question Answering Information Extraction and Summarization domains find it critical in order to proceed beyond surface understanding. With the recent creation of the Timebank Corpus Pustejovsky et al. 2003 the utility of machine learning techniques can now be tested. Recent work with the Timebank Corpus has revealed that the six-class classification of temporal relations is very difficult even for human annotators. The highest score reported on Timebank achieved 62.5 accuracy when using gold-standard features as marked by humans Mani et al. 2006 . This paper describes an approach using features extracted 173 automatically from raw text that not only duplicates this performance but surpasses its accuracy by 3 . We do so through advanced linguistic features and a surprising finding that using automatic rather than hand-labeled tense and aspect knowledge causes only a slight performance degradation. We briefly describe current work on temporal ordering in section 2. Section 4 describes the first stage of