Académique Documents
Professionnel Documents
Culture Documents
Table 7: Per-word results for preposition disambiguation with T REEBANK semantic roles. Freq gives the
frequency for the prepositions. Entropy measures non-uniformity of the role distributions. The Baseline
experiment selects the most-frequent role. The Word Only experiment just uses word collocations, whereas
Combined uses both word and hypernym collocations. Both columns show averages for percent correct over
ten trials. Total averages the values of the individual experiments (except for F req).
Table 8: Per-word results for preposition disambiguation with F RAME N ET semantic roles. See Table 7 for
the legend.
goal. This reflects the bias of F RAME N ET towards tic role assignments using all the annotations in
roles that are an integral part of the frame under con- F RAME N ET, for example, covering all types of ver-
sideration: location and time apply to all frames, so bal arguments. They use several features derived
these cases are not generally annotated. from the output of a parser, such as the constituent
Table 9 shows the results of classification when type of the phrase (e.g., NP) and the grammatical
all of the prepositions are classified together. The function (e.g., subject). They include lexical fea-
overall results are not that high due to the very large tures for the headword of the phrase and the predi-
number of roles. However, the combined colloca- cating word for the entire annotated frame. They re-
tion approach still shows slight improvement (49.4% port an accuracy of 76.9% with a baseline of 40.6%
versus 49.0%). Table 8 shows the results when us- over the F RAME N ET semantic roles. However, due
ing individual classifiers. This shows that the com- to the conditioning of the classification on the pred-
bined collocations produce better results: 70.3% icating word for the frame, the range of roles for a
versus 68.5%. Unlike the case with Treebank, the particular classification is more limited than in our
performance is below that of the individual classi- case.
fiers. This is due to the fine-grained nature of the Blaheta and Charniak (2000) classify semantic
role inventory. When all the roles are considered to- role assignments using all the annotations in T REE -
gether, prepositions are prone to being misclassified BANK. They use a few parser-derived features, such
with roles that they might not have occurred with in as the constituent labels for nearby nodes and part-
the training data, such as whenever other contextual of-speech for parent and grandparent nodes. They
clues are strong for that role. This is not a problem also include lexical features for the head and al-
with Treebank given its small role inventory. ternative head (since prepositions are considered as
the head by their parser). They report an accu-
4 Related work racy of 77.6% over the form/function tags from the
P ENN T REEBANK with a baseline of 37.8%,3 Their
Until recently, there has not been much work specif- task is somewhat different, since they address all ad-
ically on preposition classification, especially with juncts, not just prepositions, hence their lower base-
respect to general applicability in contrast to spe- line. In addition, they include the nominal and ad-
cial purpose usages. Halliday (1956) did some early verbial roles, which are syntactic and presumably
work on this in the context of machine translation. more predictable than the others in this group. Van
Later work in that area addressed the classification den Bosch and Bucholz (2002) also use the Tree-
indirectly during translation. In some cases, the is- bank data to address the more general task of assign-
sue is avoided by translating the preposition into a ing function tags to arbitrary phrases. For features,
corresponding foreign function word without regard they use parts of speech, words, and morphological
to the prepositions underlying meaning (i.e., direct clues. Chunking is done along with the tagging, but
transfer). Other times an internal representation is they only present results for the evaluation of both
helpful (Trujillo, 1992). Taylor (1993) discusses tasks taken together; their best approach achieves
general strategies for preposition disambiguation us- 78.9% accuracy.
ing a cognitive linguistics framework and illustrates
them for over. There has been quite a bit of work 5 Conclusion
in this area but mainly for spatial prepositions (Jap-
kowicz and Wiebe, 1991; Zelinsky-Wibbelt, 1993). Our approach to classifying prepositions according
to the P ENN T REEBANK annotations is fairly accu-
There is currently more interest in this type of
rate (78.5% individually and 86.1% together), while
classification. Litkowski (2002) presents manually-
retaining ability to generalize via class-based lexi-
derived rules for disambiguating prepositions, in
cal associations. These annotations are suitable for
particular for of. Srihari et al. (2001) present
3
manually-derived rules for disambiguating preposi- They target all of the T REEBANK function tags but give
performance figures broken down by the groupings defined in
tions used in named entities. the Treebank tagging guidelines. The baseline figure shown
Gildea and Jurafsky (2002) classify seman- above is their recall figure for the baseline 2 performance.
default classification of prepositions in case more M.A.K. Halliday. 1956. The linguistic basis of a
fine-grained semantic role information cannot be de- mechanical thesaurus, and its application to English
preposition classification. Mechanical Translation,
termined. For the fine-grained F RAME N ET roles,
3(2):8188.
the performance is less accurate (70.3% individu-
ally and 49.4% together). In both cases, the best Nathalie Japkowicz and Janyce Wiebe. 1991. Translat-
ing spatial prepositions using conceptual information.
accuracy is achieved using a combination of stan-
In Proc. 29th Annual Meeting of the Assoc. for Com-
dard word collocations along with class collocations putational Linguistics (ACL-91), pages 153160.
in the form of WordNet hypernyms.
K. C. Litkowski. 2002. Digraph analysis of dictionary
Future work will address cross-dataset experi- preposition definitions. In Proceedings of the Asso-
ments. In particular, we will see whether the word ciation for Computational Linguistics Special Interest
and hypernym associations learned over FrameNet Group on the Lexicon. July 11, Philadelphia, PA.
can be carried over into Treebank, given a mapping Mitchell Marcus, Grace Kim, Mary Ann Marcinkiewicz,
of the fine-grained FrameNet roles into the coarse- Robert MacIntyre, Ann Bies, Mark Ferguson, Karen
grained Treebank ones. Such a mapping would be Katz, and Britta Schasberger. 1994. The Penn Tree-
similar to the one developed by Gildea and Jurafsky bank: Annotating predicate argument structure. In
Proc. ARPA Human Language Technology Workshop.
(2002).
Tom OHara and Janyce Wiebe. 2003. Classifying func-
Acknowledgements tional relations in Factotum via WordNet hypernym as-
sociations. In Proc. Fourth International Conference
The first author is supported by a generous GAANN fellowship
on Intelligent Text Processing and Computational Lin-
from the Department of Education. Some of the work used com- guistics (CICLing-2003).
puting resources at NMSU made possible through MII Grants
Tom OHara, Janyce Wiebe, and Rebecca F. Bruce. 2000.
EIA-9810732 and EIA-0220590.
Selecting decomposable models for word-sense dis-
ambiguation: The GRLING - SDM system. Computers
and the Humanities, 34 (1-2):159164.
References
J. Ross Quinlan. 1993. C4.5: Programs for Machine
Don Blaheta and Eugene Charniak. 2000. Assigning Learning. Morgan Kaufmann, San Mateo, California.
function tags to parsed text. In Proc. NAACL-00.
Rohini Srihari, Cheng Niu, and Wei Li. 2001. A hybrid
Rebecca Bruce and Janyce Wiebe. 1999. Decomposable approach for named entity and sub-type tagging. In
modeling in natural language processing. Computa- Proc. 6th Applied Natural Language Processing Con-
tional Linguistics, 25 (2):195208. ference.
A. Van den Bosch and S. Buchholz. 2002. Shallow pars- John R. Taylor. 1993. Prepositions: patterns of polysem-
ing on the basis of words only: A case study. In Pro- ization and strategies of disambiguation. In Zelinsky-
ceedings of the 40th Meeting of the Association for Wibbelt (Zelinsky-Wibbelt, 1993).
Computational Linguistics (ACL02), pages 433440.
Philadelphia, PA, USA. Arturo Trujillo. 1992. Locations in the machine transla-
tion of prepositional phrases. In Proc. TMI-92, pages
P. Edmonds and S. Cotton, editors. 2001. Proceedings of 1320.
the S ENSEVAL 2 Workshop. Association for Compu-
tational Linguistics. Janyce Wiebe, Kenneth McKeever, and Rebecca Bruce.
1998. Mapping collocational properties into machine
Charles J. Fillmore, Charles Wooters, and Collin F. learning features. In Proc. 6th Workshop on Very
Baker. 2001. Building a large lexical databank which Large Corpora (WVLC-98), pages 225233, Montreal,
provides deep semantics. In Proceedings of the Pa- Quebec, Canada. Association for Computational Lin-
cific Asian Conference on Language, Information and guistics SIGDAT.
Computation. Hong Kong.
Ian H. Witten and Eibe Frank. 1999. Data Mining: Prac-
C. Fillmore. 1968. The case for case. In Emmon Bach tical Machine Learning Tools and Techniques with
and Rovert T. Harms, editors, Universals in Linguistic Java Implementations. Morgan Kaufmann.
Theory. Holt, Rinehart and Winston, New York.
Cornelia Zelinsky-Wibbelt, editor. 1993. The Semantics
Daniel Gildea and Daniel Jurafsky. 2002. Automatic la- of Prepositions: From Mental Processing to Natural
beling of semantic roles. Computational Linguistics, Language Processing. Mouton de Gruyter, Berlin.
28(3):245288.