Unsupervised Biographical Event Extraction Using Wikipedia Traffic

Unsupervised Biographical Event Extraction Using Wikipedia Traffic Alexander Hogue Joel Nothman e -lab, School of Information Technologies University of Sydney NSW 2006, Australia {ahog5691@uni.,joel.nothman@,james.r.curran@}sydney.edu.au James R. Curran Abstract Biographical summarisation can provide succinct and meaningful answers to the question “Who is X?”. Current super- vised summarisation approaches extract sentences from documents using features from textual context. In this paper, we explore a novel approach to biographical summarisation, by extract- ing important sentences from an entity’s Wikipedia page based on internet traffic to the page over time. Using a pilot data set, we found that it is feasible to extract key sentences about people’s notability with- out the need for a large annotated corpus. 1 Introduction “What is Julian Assange known for?” is a question which can be answered in many ways. Previous computational approaches to answering questions like these have focused on summarisation: select- ing a subset of sentences from a group of docu- ments relating to a person and ordering them (Bi- adsy et al., 2008; Zhou et al., 2004). Full text sum- maries do provide some insight into the notability of their subject, but can also contain superfluous information. To pinpoint the notoriety of individ- uals, we aim to extract the sentences from a doc- ument which show how the document’s subject is notable. We provide an alternate, unsupervised approach to the broader task of biography abstraction, which exploits external information about text, rather then extracting textual features directly. In this paper, we respond to “What is Julian As- sange known for?” with sentences mentioning im- portant events which have occurred in his life. It is the breadth of possible reasons for which one could be notable that make supervised approaches to this task difficult — the creation of a corpus large enough to cover the range of notable events would require a prohibitive amount of annotation. Furthermore, this task would be tedious for an- notators, since sentences expressing notability are sparse among documents, and some high-level un- derstanding of the notability of the page subject is required to judge each sentence’s notability. We hypothesise that when a notable event hap- pens to a person, traffic to their Wikipedia page peaks abruptly, and an edit is made to their page describing the event. To explore this hypothesis, a simple outlier- based method is applied to extract peaks (short pe- riods of sudden activity) from Wikipedia page traf- fic data, which are used to locate page edits which align to sentences contributing to the notability of the page subject. Event reference identification is a difficult task (Nothman, 2014), and errors in event extraction may mask the performance of our system, so in our initial approach we choose the sentence as our unit of event description. We evaluate by creating a corpus of Wikipedia pages about people. Each sentence annotated with its human-judged significance to the person’s no- tability. We then measure how reliably page traf- fic data can be used to identify these most notable events. Our initial investigation …

Content truncated to 500 words