Annotating Statement Pairs Extracted from Web Documents with Semantic Relations

摘要

Recognizing semantic relation is indispensable for summarizing multi-documents. In this paper, we discuss what types of semantic relations are needed to identify a variety of viewpoints on a given topic on the Web and to present these relations to users together with supporting evidence in a way that makes it clear how they are related. In order to construct a Japanese corpus, containing this information, we define semantic relations dealing with facts and opinions, and discuss how to efficiently collect valid examples and their associated semantic relations from Web documents by splitting complex sentences into fundamental units of meaning called "statements" and annotating relations at the statement level. We present an annotation scheme and examine its reliability by annotating around 1,500 pairs of statements. We are preparing the corpus for public release this summer.