Proceedings of the Ninth International Workshop on Parsing Technologies (IWPT), pages 194–195,
Vancouver, October 2005. c©2005 Association for Computational Linguistics
TFLEX: Speeding up Deep Parsing with Strategic Pruning
Myroslava O. Dzikovska
Human Communication Research Centre
University of Edinburgh
Edinburgh, EH8 9LW, UK
mdzikovs@inf.ed.ac.uk
Carolyn P. Rose
Carnegie Mellon University
Language Technologies Institute
Pittsburgh PA 15213, USA
cprose@cs.cmu.edu
1 Introduction
This paper presents a method for speeding up a
deep parser through backbone extraction and prun-
ing based on CFG ambiguity packing.1 The TRIPS
grammar is a wide-coverage grammar for deep nat-
ural language understanding in dialogue, utilized in
6 different application domains, and with high cov-
erage and sentence-level accuracy on human-human
task-oriented dialogue corpora (Dzikovska, 2004).
The TRIPS parser uses a best- rst beam search al-
gorithm and a chart size limit, both of which are a
form of pruning focused on  nding an n-best list of
interpretations. However, for longer sentences lim-
iting the chart size results in failed parses, while in-
creasing the chart size limits signi cantly impacts
the parsing speed.
It is possible to speed up parsing by implement-
ing faster uni cation algorithms, but this requires
considerable implementation effort. Instead, we de-
veloped a new parser, TFLEX, which uses a sim-
pler technique to address ef ciency issues. TFLEX
combines the TRIPS grammar with the fast parsing
technologies implemented in the LCFLEX parser
(Ros·e and Lavie, 2001). LCFLEX is an all-paths
parser which uses left-corner prediction and ambi-
guity packing, and which was shown to be ef cient
on other uni cation augmented context-free gram-
mars. We describe a way to transfer the TRIPS
grammar to LCFLEX, and a pruning method which
achieves signi cant improvements in both speed and
coverage compared to the original TRIPS parser.
1This material is based on work supported by grants from
the Of ce of Naval Research under numbers N000140510048
and N000140510043.
2 TFLEX
To use the TRIPS grammar in LCFLEX we  rst ex-
tracted a CFG backbone from the TRIPS grammar,
with CFG non-terminals corresponding directly to
TRIPS constituent categories. To each CFG rule
we attach a corresponding TRIPS rule. Whenever
a CFG rule completes, a TRIPS uni cation function
is called to do all the uni cation operations associ-
ated with the TRIPS rule. If the uni cation fails, the
constituent built by the CFG is cancelled.
The TFLEX pruning algorithm uses ambiguity
packing to provide good pruning points. For exam-
ple, in the sentence  we have a heart attack victim
at marketplace mall the phrase  a heart attack vic-
tim has two interpretations depending on whether
 heart modi es  attack or  attack victim . These
interpretations will be ambiguity packed in the CFG
structure, which offers an opportunity to make prun-
ing more strategic by focusing speci cally on com-
peting interpretations for the same utterance span.
For any constituent where ambiguity-packed non-
head daughters differ only in local features, we
prune the interpretations coming from them to a
speci ed prune beam width based on their TRIPS
scores. In the example above, pruning will happen
at the point of making a VP  have a heart attack vic-
tim . The NP will be ambiguity packed, and we will
prune alternative VP interpretations resulting from
combining the same sense of the verb  have and
different interpretations of the NP.
This approach works better than the original
TRIPS best- rst algorithm, because for long sen-
tence the TRIPS chart contains a large number
194
of similar constituents, and the parser frequently
reaches the chart size limit before  nding the correct
constituent to use. Ambiguity packing in TFLEX
helps chose the best constituents to prune by prun-
ing competing interpretations which cover the same
span and have the same non-local features, thus
making it less likely that a constituent essential for
building a parse will be pruned.
3 Evaluation
Our evaluation data is an excerpt from the Monroe
corpus that has been used in previous TRIPS re-
search on parsing speed and accuracy (Swift et al.,
2004). The test contained 1042 utterances, from 1
to 45 words in length (mean 5.38 words/utt, st. dev.
5.7 words/utt). Using a hold-out set, we determined
that a beam width of 3 was an optimal setting for
TFLEX. We then compared TFLEX at beam width
3 to the TRIPS parser with chart size limits of 1500,
5000, and 10000. As our evaluation metrics we re-
port are average parse time per sentence and proba-
bility of  nding at least one parse, the latter being a
measure approximating parsing accuracy.
The results are presented in Figure 1. We grouped
sentences into equivalence classes based on length
with a 5-word increment. On sentences greater
than 10 words long, TFLEX is signi cantly more
likely to produce a parse than any of the TRIPS
parsers (evaluated using a binary logistic regression,
p < .001). Moreover, for sentences greater than
20 words long, no form of TRIPS parser returned
a complete parse. TFLEX is signi cantly faster
than TRIPS-10000, statistically indistinguishable in
terms of parse time from TRIPS-5000, and signi -
cantly slower than TRIPS-1500 (p < .001).
Thus, TFLEX presents a superior balance of cov-
erage and ef ciency especially for long sentences
(10 words or more) since for these sentences it is
signi cantly more likely to  nd a parse than any ver-
sion of TRIPS, even a version where the chart size is
expanded to an extent that it becomes signi cantly
slower (i.e., TRIPS-10000).
4 Conclusions
In this paper, we described a combination of ef -
cient parsing techniques to improve parsing speed
and coverage with the TRIPS deep parsing grammar.
Figure 1: Parse times and probability of getting a
parse depending on (aggregated) sentence lengths.
5 denotes sentences with 5 or fewer words, 25 sen-
tences with more than 20 words.
The TFLEX system uses an all-paths left-corner
parsing from the LCFLEX parser, made tractable
by a pruning algorithm based on ambiguity packing
and local features, generalizable to other uni cation
grammars. Our pruning algorithm provides a bet-
ter ef ciency-coverage balance than best- rst pars-
ing with chart limits as utilised by the TRIPS parser.

References
M. O. Dzikovska. 2004. A Practical Semantic Represen-
tation For Natural Language Parsing. Ph.D. thesis,
University of Rochester.
C. P. Ros·e and A. Lavie. 2001. Balancing robustness
and ef ciency in uni cation-augmented context-free
parsers for large practical applications. In J.C. Junqua
and G Van Noord, editors, Robustness in Language
and Speech Technology. Kluwer Academic Press.
M. Swift, J. Allen, and D. Gildea. 2004. Skeletons in
the parser: Using a shallow parser to improve deep
parsing. In Proceedings of COLING-04.
J. Tetreault, M. Swift, P. Prithviraj, M. Dzikovska, and J.
Allen. 2004. Discourse annotation in the monroe cor-
pus. In ACL-04 workshop on Discourse Annotation.
