Learning and Representing Natural Language Phrases in a Hybrid Symbolic/Connectionist Approach

Our general research interests include the representation of natural language using connectionist and symbolic methods. Our approach aims at evaluating and integrating properties of symbolic and connectionist architectures. Primarily, we concentrate on syntactic and semantic representations focusing on structural disambiguation and semantic classification. As a general task we chose the analysis of phrases. Phrasal analysis often can not rely on as much predictive top-down knowledge as complete sentence analysis and therefore more bottom-up analysis is needed. In this context, connectionist networks appear to be a particularly useful method for learning and representing necessary knowledge for a bottom-up analysis. Using online available corpora and library classifications we designed several hybrid symbolic/ connectionist architectures. As examples for structural disambiguation we focused on prepositional phrase attachment and coordination using localist relaxation networks, distributed plausibility networks, and a symbolic chart parser. As examples for semantic classification we designed a combination of a preprocessing chart parser with a connectionist autoassociator as well as a connectionist architecture using recurrent sequential classification networks. These architectures allow the combination of predefined symbolic knowledge with learned connectionist knowledge for natural language processing.

@InProceedings{Wer91b,
 	 author =  {Wermter, Stefan},
 	 title = {Learning and Representing Natural Language Phrases in a Hybrid Symbolic/Connectionist Approach},
 	 booktitle = {Proceedings of the AAAI Spring Symposium on Machine Learning of Natural Language and Ontology},
 	 journal = {None},
 	 editors = {}
 	 number = {}
 	 volume = {}
 	 pages = {191--195},
 	 year = {1991},
 	 month = {}
 	 publisher = {None},
 	 doi = {}
 	 url = {None},
 }