transparent gif

 

Ej inloggad.

Göteborgs universitets publikationer

Multilingual Grammars and Universal Dependencies

Författare och institution:
Prasanth Kolachina (Institutionen för data- och informationsteknik (GU))
Publikationstyp:
Licentiatavhandling
Förlag:
University of Gothenburg
Förlagsort:
Göteborg
Publiceringsår:
2016
Språk:
engelska
Datum för examination:
2016-09-15
Tidpunkt för examination:
10:00
Opponent:
Filip Ginter
Sammanfattning (abstract):
Abstract syntax trees are an alternative representation to syntactic structures commonly found in NLP systems. This representation allows for sharing of structures across languages, making it well suited to serve as a translation interlingua. Grammatical Framework is a grammar formalism that captures cross-linguistic generalizations through the use of abstract syntax. The Resource Grammar Library (GF-RGL) in GF implements multilingual grammars for over 30 languages. Universal Dependencies (UDs) is a parallel effort to use shared structures to analyse sentences in different languages. The set of part-of-speech tags and functions are shared across languages. The linguistic data available from this project is annotated data i.e. sentences annotated with UD structures in over 40 languages. The main contribution of this thesis is to bridge these two representations: despite the similar motivation behind these two efforts, the representations used vary significantly. Hence, we propose a conversion method to convert the abstract syntax trees in GF to the structures used in UD.We find that the correspondence between GF-RGL and UD is significant, and the differences between the two raise interesting questions about the level of abstraction. We also present practical applications to our method: (1) the use of GF parser as a dependency parser and (2) to bootstrap UD treebanks from GF treebanks. Another topic addressed in this thesis is the problem of out-of-vocabulary words that comes up in symbolic systems. We address this problem in the context of part-of-speech tagging and statistical dependency parsing. We propose a simple method to use a distributional thesaurus to replace unknown words and show through empirical evaluation that our method improves both overall accuracies and accuracies for unknown words. Our method is generic and can be adapted to fit other NLP systems.
Ämne (baseras på Högskoleverkets indelning av forskningsämnen):
TEKNIK OCH TEKNOLOGIER ->
Elektroteknik och elektronik ->
Datorsystem
Nyckelord:
NLP, Grammatical Framework, Universal Dependencies, Abstract Syntax
Postens nummer:
241087
Posten skapad:
2016-09-01 08:25

Visa i Endnote-format

Göteborgs universitet • Tel. 031-786 0000
© Göteborgs universitet 2007