Morphological Tags in Parallel Corpora

Multilingual parallel corpora can be annotated with morphosyntactic tags by monolingual tools, freely available for a number of different languages. However, each of the tools is typically bundled with a specific tagset and assumes a specific way of tokenization. The variety of tagging schemes and tag formats may be a problem for the user: a relatively simple tag query in a multilingual setting often means spending a while with tagset manuals. The aim of the present contribution is to suggest a solution that would delegate the task of dealing with multiple tagsets to the system. The core component of the proposal can be viewed as an abstract interlingual tagset. It is actually a hierarchy of linguistic categories, partially ordered by their specificity, mapped to tags in languagespecific tagsets. In order to capture different views of word classes, as seen by the tagsets, the common tagset takes three different perspectives of word class: lexical, inflectional and syntactic, each potentially coupled with its own set of morphological categories. Thus, the tag for the Czech relative pronoun který ‘which’ is decoded as a category with the properties of lexical pronoun, inflectional adjective and syntactic noun, each with its appropriate morphological characteristics. The common tagset is formalised as a tangled hierarchy of types, each of the types corresponding to a linguistic category and some of the types to one or more languagespecific tags. Tags in all tagsets can be described as objects with properties such as lexical, inflectional and syntactic word class, and the relevant morphological categories. Then the standard methods of Formal Concept Analysis (Ganter & Wille, 1999) can be used to construct the hierarchy automatically as a concept lattice and to (partially) resolve tag queries that do not quite match the tags used for the specific language. in a way similar to that used by Janssen (2004) for dealing with lexical gaps in the multilingual lexical database. Language-specific subsets of the abstract common tagset can be extracted using the links to tags in language-specific tagsets. Abstract language-specific tagsets can be used to generate or interpret tags in a format of the user’s or a tool’s preference. In addition, the modular setup allows for underspecified tag queries and for mappings between tagsets