Inferring Knowledge from a Large Semantic Network

Inferring knowledge from a large semantic network Dominique Dutoit Thierry Poibeau Memodata et CRISCO 17, rue Dumont d’Urville F-14000 Caen Thales and LIPN Domaine de Corbeville F-91404 Orsay memodata@wanadoo.fr thierry.poibeau@thalesgroup.com Abstract In this paper, we present a rich semantic network based on a differential analysis. We then detail implemented measures that take into account common and differential features between words. In a last section, we describe some industrial applications. 1 Introduction: textual and differential semantics In textual analysis, each lexical item from a text is broken down in a list of semantic features. Features are intended to differentiate one word from another: a naive example would be a feature back that could express the difference between a chair and a stool. Of course, most of the time, features are not so easy to define. Some feature typologies have been provided, but there are still much discussions about the nature of a feature in a text. Most of the studies concerning differential semantics are based on a human approach to texts (this can lead to different problems, see below). Textual Semantics, also called differential semantics, is revisiting the concepts of continental structuralism like decomponential semantics (Cavazza, 1998). The problem is then to have a lexical formalism that allows, for a lexical item, a simple description and some other features which could be dynamically inferred from the text. For example, the dictionary should mention that a “door” is an aperture, but it is more questionable to mention in the dictionary that “one can walk through a door”. However, it can be an important point for the interpretation of a sentence in context. That is the reason why Pustejovsky introduced in the nineties the notion of “generative lexicon” (Pustejovsky, 1991) (Pustejovsky, 1995). His analysis has to deal with the notion of context: he proposes to associate to a word a core semantic description (the fact that a “door” is an “aperture”) and to add some additional features, which can be activated in context (“walk-through” is the telic role of a “door”). However, Pustejovsky does not take into account important notions such as lexical chains and text coherence. He proposes an abstract model distant from real texts. Semantic features can be used to check out text coherence through the notion of “isotopy”. This notion is “the recurrence within a given text section (regardless of sentence boundaries) of the same semantic feature through different words” (Cavazza, 1998). The recurrences of these features throughout a text allows to extract the topic of interest and some other points which are marginally tackled in the text. It provides interesting ways to glance at the text without a full reading of it; it also helps the interpretation. In this paper, we present a rich semantic network based on a differential analysis. We then detail implemented measures that take into account common and differential features between words. In a last section, we describe some industrial applications. 2 The semantic network The semantic network used in this experiment is a multilingual …

Content truncated to 500 words