Inferring the semantic properties of sentences by mining syntactic parse trees

Abstract

We extend the mechanism of logical generalization toward syntactic parse trees and attempt to detect semantic signals unobservable in the level of keywords. Generalization from a syntactic parse tree as a measure of syntactic similarity is defined by the obtained set of maximum common sub-trees and is performed at the level of paragraphs, sentences, phrases and individual words. We analyze the semantic features of this similarity measure and compare it with the semantics of traditional anti-unification of terms. Nearest-Neighbor machine learning is then applied to relate the sentence to a semantic class. By using a syntactic parse tree-based similarity measure instead of the bag-of-words and keyword frequency approaches, we expect to detect a subtle difference between semantic classes that is otherwise unobservable. The proposed approach is evaluated in three distinct domains in which a lack of semantic information makes the classification of sentences rather difficult. We conclude that implicit indications of semantic classes can be extracted from syntactic structures

We are grateful to our colleagues SO Kuznetsov, B Kovalerchuk and others for valuable discussions and to our anonymous reviewers for their suggestions. This research is partially funded by the EU Project No. 238887, a unique European Citizens' attention service (iSAC6+) IST-PSP. This research is also funded by the Spanish MICINN (Ministerio de Ciencia e Innovacion) IPT-430000-2010-13 project Social powered Agents for Knowledge search Engine (SAKE), TIN2010-17903 Comparative approaches to the implementation of intelligent agents in digital preservation from a perspective of the automation of social networks, and the AGAUR 2011 Fl_B00927 research grant awarded to Gabor Dobrocsi and the grup de recerca consolidat CSI-ref.2009SGR-1202