Generation of coherent gene summary

Abstract

Typical approaches to automatic summarization make efforts to generate a coherent document by arranging the order of sentences according to certain criteria such as the publication date of the text in which they appear. These methods can be applied to the domains where the facts contained in the sentences are related to one another in an implicit way, which enables the ordering of the facts. However, when there exists no implicit relation between facts, the presented facts would appear fragmented to the user, and in such cases, an additional sentence that functions to weave the facts together by explicitly stating the relations that exist between them should be inserted to make the text more natural and comprehensible. This quality of a text is called coherence. When generating a gene summary, coherence is an important matter because implicit pieces of knowledge about the relations among the gene-related facts are not available in every reader``s mind, and without mentioning them explicitly, the comprehension of the material would not be efficient. Such ``linking`` sentences should be in the form of the main verb connecting each main concept of the facts to be correlated. We utilize combinatory categorial grammar to find the syntactic dependency tree of a sentence, and use the dependecy relation to identify sentences with such a structure. Also, we constrain the types of linking sentence such that the two main concepts should not be much different from the concepts used in the original sentences. We tackle the issue with the notion of concept generality, which we determine by examining the syntactic dependency structure within a noun phrase to which a concept belongs. We have shown that such forms of linking sentence actually enhance the smoothness of the transition between the original sentences. Also, we have shown that such forms of a sentence help readers efficiently understand the text, based on the cognitive psychological ground. We evaluated our ap...