Cross-language Ontology Learning

Introduction

To indicate why we think integrating information from different languages is a good idea, we turn to the field of lexical typology. It is concerned with the ways in which different languages “dissect” semantics, or meaning, and form it into words (Koch, 2001). E.g., where English has ‘sibling’ as a unifying word for ‘brother’ and ‘sister’, French only has ‘frère’ and ‘sœur’ but no unifying word; how kinship relations are expressed varies greatly in the languages of the world (Koch, 2001). Another example is that English ‘go’ corresponds to both German ‘gehen’ (go by foot) and ‘fahren’ (go by some means of transportation) (Goddard, 2001). Figure 1 also depicts how a particular semantic field is subject to different categorizations in different languages. We believe that this type of diversity will prove an asset in an ontology learning system, providing different “views” that each contributes useful information. We describe how such differences influence features in our machine learning experiments later in the chapter.

Figure 1.

Words related to trees/wood in different languages. Ewe is a language in the Niger-Congo family and is spoken in Ghana, Togo and Benin. Note how English and French are closer to each other in this example than to Swedish, even though English and Swedish are both Germanic languages. (Figure from Mikael Parkvall, Stockholm University, personal communication)

Exploiting Cross-Language Data

Our major hypothesis is that combining information from different languages will improve the results of the ontology learning task, which is traditionally approached in a single language framework. We describe the process of going from a domain-specific parallel or comparable corpus and language-specific terms to a hierarchical is-a ordering of cross-language term sets.

Regarding cross-language data, especially data taken from a parallel corpus, there are two extreme standpoints possible. One could argue that, by adding texts translated into another language, we are in fact doubling the amount of data, basing the argument on crude numbers such as bytes used for storage or the like. The other extreme would be to say that we do not add anything at all, but are merely repeating the exact same information, just using a different ”encoding”. We show that the use of an additional language reveals certain pieces of information that were hidden in the initial language.

According to Sager (1994), the notion of equivalence is central to the field of translation. Equivalence relations, in turn, have the properties of being reflexive (any word is a translation of itself), symmetrical (if A is a translation of B, then B is a translation of A) and transitive (if A is a translation of B, and B of C, then A is also a translation of C) (Boolos & Jeffrey, 1989). In practice, the notion of equivalence could be modified to a notion of relative equivalence (see Figure 1), just as the synonymy relation is commonly relativized.

In addition to lexico-typological aspects, there are other sources of discrepancy, when dealing with translations. For a given word in the source language, it is not always possible to come up with a single translation to that word. This can have different causes, apart from the typological differences already discussed:

• •

  The source word is polysemic or homonymic and the different senses of the word give rise to different translations in the target language.

• •

  The target language has more than one word with more or less the same meaning, used interchangeably as translations of the source word (the target language words are synonyms).

We expect cases as the ones mentioned to occur less frequently in domain-specific text than they would in general, non-technical texts. Though problematic at times, we actually expect that these discrepancies will be a source for added information when trying to learn an ontology over a domain, rather than pose a problem.