WeSearch_LexicalFiltering

Background

Working with a lattice of lexical hypotheses and an (über)tagger, we seek to develop a filtering function that discards unlikely hypotheses. The formalisation of the lexical filtering process may be found [http://dl.dropbox.com/u/680530/WeSearch/Lexical%20Filtering/formalisation.pdf here].

TNT output for filtering of LE types

One such filter function maps PTB tags output from the TNT tagger onto LE Types. Mappings may be derived intuitively from inspection of a [http://dl.dropbox.com/u/680530/WeSearch/Lexical%20Filtering/tnt.le.confusion.pdf confusion matrix] detailing the choices of TNT with respect to LE types.

An alternative approach is to programmatically find mappings based on the preferred outcomes of lexical filtering, (i.e. gains in parser efficiency versus losses in parser accuracy and coverage). These outcomes may be approximated by examining the relations between TNT precision, TNT recall and the lexical ambiguity of LE types.

Frequency of LE types, cross-validated across subsets of the WeScience corpus:

type	frequency	std. dev.
n	3830	450
v	1712	237
p	1401	132
d	1119	124
aj	1073	126
av	411	58
c	381	43
cm	129	24
pp	61	11
pt	15	9
x	1	1

ROC plots of the TNT performance on the most frequent LE types [http://dl.dropbox.com/u/680530/WeSearch/Lexical%20Filtering/roc.png]

• Rebecca Driden (2009), [http://www.dridan.com/research/papers/dridan-phdthesis.pdf Using Lexical Statistics to Improve HPSG Parsing], PhD Thesis, Saarland University