lunduniversity.lu.se

Language Technology at LTH

Computer Science | Faculty of Engineering, LTH

Denna sida på svenska This page in English This page in English

Semantic Technologies

Semantics

Our group has developed a line of universal, high-performance semantic parsers based on the Framenet or PropBank/NomBank paradigms and syntactic dependencies.

New! You can enter your own sentences and try the parser online here: http://barbar.cs.lth.se:8081/.

In 2007, our semantic parser obtained the best results at the SemEval evaluation on semantic structure extraction in English. In 2008, the parser won the very competitive CoNLL evaluation on the joint analysis of syntactic and semantic dependencies in English. In 2009, CoNLL organized a new evaluation on seven languages. We obtained the second best average semantic score, all tasks, with a labeled semantic F1 of 80.31 and with the best F1 score for the Chinese and German data and the second best one for English.

For English, the parser consisted originally of two components: A converter that transformed constituent parse trees into dependency graphs and the parser itself that constructed semantic dependencies. The constituent-to-dependency converter, as a standalone application, has become a standard component used by many researchers, including the yearly competition at the Conference on Computational Natural Language Learning (CoNLL 2007, 2008, and 2009). As dependency parsers are becoming mainstream for English, our semantic parsers are more likely to use the output of a native dependency analysis now.

Apart from semantic parsing, we also developed systems for the ordering of temporal relations and coreference solving. Both have been tested with texts written in Swedish, but are extensible to any language for which annotated data are available.

All our semantic tools use supervised classification techniques.

Page Manager: