Sentence Classification and Clause Detection for Croatian Kristina

Sentence Classification and Clause Detection for Croatian Kristina Vučković, Željko Agić, Marko Tadić Department of Information Sciences, Department of Linguistics Faculty of Humanities and Social Sceinces, University of Zagreb {kvuckovi, zagic, marko. tadic}@ffzg. hr FASSBL 7 Conference Dubrovnik, Croatia 2010 -10 -05

Overview § What? § classifying Croatian sentences by structure § detecting independent and dependent clauses § How? § implemented a prototype system in Noo. J § linked it with a morphosyntactic tagger § evaluated on a sample from Croatian corpora § Why? § rule-based chunking and shallow parsing

Classification and detection § sentence segmentation is easy when considering sentence boundaries only § here, we: § detect boundaries of clauses in complex sentences § assign type to sentences § sentence classification § purpose: declarative, interrogative, etc. § structure: simple and complex § complex sentences § independent complex, i. e. compound sentences § dependent complex sentences

Classification and detection § independent complex sentences § independent clause connected to the main clause by using a conjunction § type defined by the choice of conjunction § e. g. constituent clause, conjunctions {i, pa, te, niti} § disjunctive, opposite, exclusive, conclusive and explanatory clause § Svi su spavali, jedino sam ja bio budan. (exclusive) § dependent complex sentences § main clause is independent, all the others depend on it and cannot stand alone in a sentence § Predicative, subjective, objective, attributive, appositional and adverbial clause § Ispričat ću ti što mi se dogodilo. (objective)

The system § prototype implemented in Noo. J § finite state transducer cascades (local grammars) § Croatian lexical resources § each cascade detects and annotates a different type of clause § built on top of a chunker for Croatian § the top-level grammar § two types of subgraphs: main clauses and independent clauses

The system § Main clause grammar § presence of a VP and possibly any other phrase § independent clauses recognized just by using the conjunctions § implementation of dependent clause detection varies across clause types

Experiment setup § used the CW 100 corpus § XCES-encoded to word level § sentence delimited, tokenized, manually lemmatized and MSD-annotated § 200 randomly selected sentences § 100 for the development and 100 for testing § utilized the Cro. Tag tagger § Noo. J input format allows external annotation § created three systems § no preprocessing § tagging input sentences with Cro. Tag (~85% accuracy) § using the manually assigned tags from CW 100 § recall, precision, F 1 -measure

Results § scores for the three systems § “perfect” tagging system is the top-performer § benefits of automatic tagging? Recall Precision F 1 -measure No tagging 0. 730 0. 890 0. 802 Cro. Tag 0. 657 0. 899 0. 759 Manual tagging 0. 813 0. 928 0. 867 § distribution of assigned types § main, objective, opposite, adverbial, attribute, . . . § misclassifications § attributive and objective most commonly misclassified § data sparseness

Conclusions and future work § the system scores good in terms of F 1 -measure § open issues § § verb coordination dislocated nominal predicates attribute classes starting with a PP complex insertion of dependent clauses § no real benefit from automatic MSD-tagging § future work § resolving the issues § re-evaluation on a larger test set? § integration with a rule-based shallow parser

Thank you for your attention. www. accurat-project. eu The research within the project ACCURAT leading to these results has received funding from the European Union Seventh Framework Programme (FP 7/20072013), grant agreement no 248347.