Semantics and Parsing

The Plink parser is a human-like parser. It is deterministic, linear, and generates one interpretation. The current version of the system has a very weak semantics module.

Semantics is used in making parsing decisions. Two simple examples show this
I saw the girl with the boy.
I saw the girl with the telescope.
In this case the boy is with the girl, where the telescope is the instrument of saw.

It is easy to encode these examples so that Plink can generate the correct interpretations. However, it is quite difficult to do this for all sentences. A common mechanism to 'learn' the semantics is to work with large corpora, and find how things attach. This is effective for the examples when a corpus contains several sentences of the form I-see-girl-telescope, but most sentences are unique.

The obvious solution is to take advantage of hierarchy. While there may be no examples of I-see-girl-telescope, there may be an example of human-see-X-telescope or human-see-human-telescope, or even I-see-human-telescope. This leads to several questions.

• Which pattern is important? In this case see and telescope are important, where the object is less important, and the actor even less important.
• Which hierarchy should be used? Common examples are those from the Longman's Dictionary of Contemporary English, and WordNet. Parsing correctness (as measured by Parseval or another measurement) can show which hierarchy, or ontology, is better for parsing.
• What happens when evidence conflicts?
• What learning algorithms are useful?
• What about extra-sentential context?

Exploration of these questions will aid Natural Language parsing. Additionally, it will provide an evaluation for ontologies.

Extra Links

• Natural Language Parsing with Cell Assemblies. A draft for this years AISB conference in postscript in tex
• Natural Language Parsing with Cell Assemblies. A talk I'm planning on giving at U. of Sheffield in November 2000.
• This paper was presentend to the European Conference on AI http://www.ecai2000.hu-berlin.de/ in August 2000
• The paper can be found in postscript and in latex. The slides from the talk at the conference are also here.
• GATE Homepage
• Rob Gaizauskas' home page
• Keh Kok Yong's home page


This page was developed by Chris Huyck
This page last modified: