{"id":75,"date":"2023-10-31T20:22:39","date_gmt":"2023-10-31T20:22:39","guid":{"rendered":"https:\/\/www.isi.edu\/people-hobbs\/?page_id=75"},"modified":"2023-11-13T05:53:58","modified_gmt":"2023-11-13T05:53:58","slug":"an-annotated-corpus-of-typical-durations-of-events","status":"publish","type":"page","link":"https:\/\/www.isi.edu\/people-hobbs\/research-areas\/temporal-representation\/an-annotated-corpus-of-typical-durations-of-events\/","title":{"rendered":"An Annotated Corpus of Typical Durations of Events"},"content":{"rendered":"\n\n\t

(Annotations Integrated into\u00a0TimeBank)<\/h2>\n

Suppose we read the sentence, “George W. Bush met with Vladimir Putin in Moscow.<\/em>” We don’t know exactly how long that meeting lasted, but we do get\u00a0some<\/em>\u00a0temporal information from the sentence.\u00a0 We know the meeting lasted more than ten seconds and less than one year. As we guess narrower and narrower bounds, our chances of being correct go down. Just how accurately can we make duration judgments like this? How much agreement can we expect among people? Will it be possible to extract this kind of information from text automatically?<\/p>\n

The goal of the project is to extract such implicit and vague typical durations of events from text automatically. This research is potentially very important in applications in which the time course of events is to be extracted from news. For example, whether two events overlap or are in sequence often depends very much on their durations.\u00a0 If a war started yesterday, we can be pretty sure it is still going on today.\u00a0 If a hurricane started last year, we can be sure it is over by now.<\/p>\n

We have developed annotation guidelines to minimize discrepant judgments and annotated all the 48 non-Wall-Street-Journal (non-WSJ) news articles (2132 events), as well as 10 WSJ articles (156 events) from the TimeBank corpus; we have developed a method for measuring inter-annotator agreement when the judgments are intervals on a scale; and we have showed that machine learning techniques applied to this data yield\u00a0coarse-grained<\/em>\u00a0event duration information, considerably outperforming a baseline and approaching human performance.<\/p>\n

Documents<\/u><\/h3>\n