Identifying Cross-Document Relations between Sentences

Identifying Cross-Document Relations between Sentences Yasunari Miyabe †§ Hiroya Takamura ‡ Manabu Okumura ‡ †Interdisciplinary Graduate School of Science and Engineering, Tokyo Institute of Technology, Japan ‡Precision and Intelligence Laboratory, Tokyo Institute of Technology, Japan miyabe@lr.pi.titech.ac.jp, {takamura,oku}@pi.titech.ac.jp Abstract A pair of sentences in different newspaper articles on an event can have one of sev- eral relations. Of these, we have focused on two, i.e., equivalence and transition. Equiv- alence is the relation between two sentences that have the same information on an event. Transition is the relation between two sen- tences that have the same information except for values of numeric attributes. We pro- pose methods of identifying these relations. We first split a dataset consisting of pairs of sentences into clusters according to their similarities, and then construct a classifier for each cluster to identify equivalence re- lations. We also adopt a “coarse-to-fine” ap- proach. We further propose using the identi- fied equivalence relations to address the task of identifying transition relations. 1 Introduction A document generally consists of semantic units called sentences and various relations hold between them. The analysis of the structure of a document by identifying the relations between sentences is called discourse analysis. The discourse structure of one document has been the target of the traditional discourse anal- ysis (Marcu, 2000; Marcu and Echihabi, 2002; Yokoyama et al., 2003), based on rhetorical struc- ture theory (RST) (Mann and Thompson, 1987). §Yasunari Miyabe currently works at Toshiba Solutions Cor- poration. Inspired by RST, Radev (2000) proposed the cross-document structure theory (CST) for multi- document analysis, such as multi-document summa- rization, and topic detection and tracking. CST takes the structure of a set of related documents into ac- count. Radev defined relations that hold between sentences across the documents on an event (e.g., an earthquake or a traffic accident). Radev presented a taxonomy of cross-document relations, consisting of 24 types. In Japanese, Etoh et al. (2005) redefined 14 CST types based on Radev’s taxonomy. For example, a pair of sentences with an “equivalence relation” (EQ) has the same information on an event. EQ can be considered to correspond to the identity and equivalence relations in Radev’s taxonomy. A sentence pair with a “tran- sition relation” (TR) contains the same numeric at- tributes with different values. TR roughly corre- sponds to the follow-up and fulfilment relations in Radev’s taxonomy. We will provide examples of CST relations: 1. ABC telephone company announced on the 9th that the number of users of its mobile-phone service had reached one million. Users can ac- cess the Internet, reserve train tickets, as well as make phone calls through this service. 2. ABC said on the 18th that the number of users of its mobile-phone service had reached 1,500,000. This service includes Internet ac- cess, and enables train-ticket reservations and telephone calls. The pair of the first sentence in 1 and the first sen- tence in 2 is in TR, because the number of users 141 has changed from one million to 1.5 millions, …

Content truncated to 500 words