Unsupervised Timeline Generation for Wikipedia History Articles

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2343–2349, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Unsupervised Timeline Generation for Wikipedia History Articles Sandro Bauer and Simone Teufel Computer Laboratory University of Cambridge Cambridge, United Kingdom {sandro.bauer, simone.teufel}@cl.cam.ac.uk Abstract This paper presents a generic approach to con- tent selection for creating timelines from in- dividual history articles for which no exter- nal information about the same topic is avail- able. This scenario is in contrast to existing works on timeline generation, which require the presence of a large corpus of news arti- cles. To identify salient events in a given his- tory article, we exploit lexical cues about the article’s subject area, as well as time expres- sions that are syntactically attached to an event word. We also test different methods of en- suring timeline coverage of the entire histori- cal time span described. Our best-performing method outperforms a new unsupervised base- line and an improved version of an existing supervised approach. We see our work as a step towards more semantically motivated ap- proaches to single-document summarisation. 1 Introduction While there has been much work on generating his- tory timelines automatically, these approaches are commonly evaluated on events that took place in re- cent decades, as they depend on the availability of large numbers of articles describing the same his- torical period. If such a rich data source is avail- able, it is possible to exploit document creation times, redundancy across documents, as well as back-references to earlier events in order to identify salient events. For instance, the start of the Iraq War in 2003 is mentioned frequently in a general news corpus, including in articles published years after the event took place. The high number of mentions sug- gests that the beginning of the Iraq War was an im- portant historical event. However, for most historical periods covered in history articles (e.g., Antiquity or the Middle Ages), such cues are not commonly available, as no news articles from these eras exist. Generating event time- lines for arbitrary historical periods is therefore a much harder problem, which requires methods that rely less heavily on the types of rich, parallel and dense information contained in news clusters. To investigate this problem, we approach timeline generation as a special single-document summarisa- tion task. In other words, we assume that the in- formation to be summarised is contained in a single history article, and that no further mentions of spe- cific events exist externally. This is a realistic sce- nario, for instance, for a specialist article describing the history of music in Ancient China. We introduce a method for selecting salient con- tent in history articles of any subject area, as long as the events in the text are roughly ordered chronolog- ically. The hypothesis is that knowledge of an text’s subject area can help decide which content should be selected. Another intuition is that certain com- binations of events should be avoided in …

Content truncated to 500 words