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 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 confusion matrix detailing the choices of TNT with respect to LE types.

An alternative approach is to 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 ambiguity of LE types.

Frequency of LE types in JH0:

type	frequency
n	1,134,661
p	498,443
v	454,513
d	335,667
aj	332,243
c	182,759
av	145,290
cm	33,618
pp	34,496
pt	3,864

Plots of the TNT performance on the most frequent LE types

http://dl.dropbox.com/u/680530/WeSearch/Lexical%20Filtering/roc.png

A plot of the precision vs. lexical items filtered for each handled LE type:

http://dl.dropbox.com/u/680530/WeSearch/Lexical%20Filtering/filtering.png

Effective threshold ranges for LE Types:

type	min	min-precision	min-filtering	max	max-precision	max-filtering
n	0.35	0.93639	493184	1.00	0.94253	210876
v	0.47	0.94183	217656	1.00	0.96008	130873
p	0.35	0.89550	1196773	1.00	0.94866	708213
d	0.59	0.92464	408522	1.00	0.93566	318760
aj	0.36	0.73078	629913	1.00	0.90287	276468
av	0.30	0.63747	430465	1.00	0.74478	115081

Related Work

• Driden, R. (2009), Using Lexical Statistics to Improve HPSG Parsing, PhD Thesis, Saarland University

• Ninomiya, T., Matsuzaki, T., Miyao, Y., Tsuruoka, Y and Tsujii, J. (2010). HPSG Parsing with a Supertagger. In Bunt, H., Merlo, P., and Nivre, J. (Eds.): Trends in Parsing Technology: Dependency Parsing, Domain Adaptation, and Deep Parsing. Springer, pp. 243-256.