Week 6

I wrote up a description of my Methods:

The Link Grammar Parser, written by Temperley, Sleator, and Lafferty, provides a set of links connecting pairs of words for a given input sentence. In its latest version, it supports 107 different linkage types. Our goal is to use this software to generate additional features for the BANNER system to use to identify named entities. This will require modifying BANNER to send each sentence to Link Grammar and then accept the resulting linkage parse tree and process it into features.

To maximise the effectiveness of the additional features, we will first use these features separately from the rest. The hypothesis is that the other features may overpower the parse features and not result in the most effective use of them. Should this prove successful, we will also evaluate the use of the parse features simply alongside the others to determine if this extra measure is useful.

In 2006, Pyysalo, Salakoski, et al., adapted Link Grammar for the biomedical domain. Their work produced a 10% decrease in error when using the parser on biomedical texts. The authors of the original Link Grammar Parser have since made updates to their software, so the effectiveness of using an outdated adapted parser is an open question.