Towards a Bootstrapping Framework for Corpus Semantic Tagging

Towards a Bootstrapping Framework for Corpus Semantic Tagging Roberto Basili, Michelangelo Della Rocca, Maria Teresa Pazienza Dipartimento di Informatica, Sistemi e Produzione, Universita' di Roma, Tor Vergata (ITALY) {bas ili, dellaroc, pazienza}@inf o. ut ovrm. it Abstract Availability of source information for se- mantic tagging (or disambignating) words in corpora is problematic. A framework to produce a semantically tagged corpus in a domain specific perspective using as source a general purpose taxonomy (i.e. Word- Net) is here proposed. The tag set is de- rived from higher level Wordnet synsets. A methodology aiming to support semantic bootstrapping in a NLP application is de- fined. Results from large scale experiments are reported 1. 1 Introduction Lexical Acquisition (LA) processes strongly rely on basic assumptions embodied by the source informa- tion and training examples. Several approaches to LA rely on some forms of declarative descriptions of source data: bracketed or POS tagged corpora are just examples. Many authors claim that class-based methods are more robust against data sparseness problems (Dagan,1994), (Pereira, 1993), (Brown et al.,1992). Other works (Basili et al.,1993a, 1996) demonstrated that a variety of lexical acquisition methods over small size corpora are viable when- ever a domain specific semantic bias is available: using high level semantic classes (rather than sim- ple words) increase the robustness of the proba- bility driven methods, usually affected by coverage and data sparseness problems. Furthermore, domain specific semantic classes add expressivity to the un- derlying statistical acquisition model (Basili et al., 1996): this saves the knowledge engineer from hav- ing to deal with mysterious scores with no linguistic flavor. Semantic data possess an explanatory power that is truly required in specific knowledge domains. 1 This work has been partially supported by the Esprit LRE project n. 2110 - ECRAN Modeling semantic information is much more cor- pus and domain dependent than POS or syntactic tagging. Bracketed corpora are core components of an underlying grammatical knowledge to which re- suits of different inductive methods equivalently re- fer. Such equivalence is no longer valid for seman- tic tagging when corpora (as well as underlying do- mains) change. In order to design tagging capabili- ties at a semantic level, it is more important to de- sign adaptation capabilities to process a given corpus in a domain driven fashion. Tagging is a dynamic process that aims to produce a core semantic infor- mation to support several induction processes over the same domain. Availability of source information to support any tagging activity is problematic: general purpose sources (e.g. MRDs and static Lexical Knowledge Bases) may be too generic and worsen the induc- tion quality, while specific domain sources are usu- ally absent. Although semantic information is cru- cial to the induction of most lexical knowledge, ac- cessing it is often impossible. As gold standards are fairly questionable, it is necessary to rely on sources that are as much systematic as possible and adapt- ing their description to the underlying corpus. The widespread diffusion of WordNet, and its large scale as well, …

Content truncated to 500 words