Entities on land
5 Annotation Results
The distribution of the four types of [±movement] feature specified at V and NP1 is summarized in Table 3. Overall, over half of the PPs are attached to the verb [+, -] (51.5%). Noun-attached [+, -] and undetermined structures [+, +] fall into close range (17.4% and 19.8%, respectively).
The lowest proportion is co-attachment [-, -] (11.3%) which is expected due to its atypical semantic properties.
Table 3: Distribution of four types of feature specification.
Hypothesized
Next, we compare our annotation results with that extracted directly from the Penn Treebank which was generated from the traditional binary attachment approach. Table 4 demonstrates the feature distribution of the Penn Treebank binary attachment in percentage and counts in parentheses. A match of 65.6% for verb-attached PPs and 44.8% for noun-attached PPs can be
3 An example of noun-attached structure: (VP (VBZ parses) (NP#1230 (NNS names)) (PP-CLR (IN into) (NP (DT every) (JJ conceivable) (NN interest) (NN group)))).
4 An example of noun-attached structure: (VP (VB plot) (NP (DT a) (JJ peaceful) (NN course)) (PP-CLR (IN into) (NP#1080 (DT the) (NN future)))).
are found under the verb-attached category in the Penn Treebank annotation. More than half of these instances are communication and cognitive verbs (100/137). Examples include masterminded, consolidated, picked, implies, mandate, requested and protested. For those verb-attached PPs classified under the noun category, the main verb categories are change and motion verbs (39/76), exemplified by verbs like fermented, infusing, integrating, transforming, reached, walk, sweeping, threw and drops.
Table 4: Feature analysis of the Penn Treebank computer-generated binary attachment.
Hypothesized attachment site
[± movement] Penn Treebank Binary Attachment V NP1 Verb-attached PPs Noun-attached PPs
Verb-attached + - 60.9% (874) 18.5% (76)
Noun-attached - + 9.6% (137) 44.8% (184)
Co-attached - - 14.2% (204) 1.0% (4)
Undetermined + + 15.3% (219) 35.8% (147)
Total 100% (1434) 100% (411)
A further analysis of the semantic domains of all the co-attached PPs can be found in Table 5.
The social, emotion, cognition and communication domains comprise of nearly 90% of the data.
Table 5: The distribution of semantic domains of verbs in [- -] category.
Verb domains
Communication Social Emotion Cognition Stative Consumption Perception Total
Count 94 Both the social and communication domains consist of verbs that can only be interpreted with rhetorical force as shown bolded in (13a) and (13b), respectively.
(13a) …punish Iran and Iraq into an agreement on each other's production quotas…
(WSJ-V42)
(13b) …the courts have refused to uphold contracts in which people have voluntarily contracted themselves into peonage or slavery. (WSJ-V30)
Emotion verbs are also found to have their emotional sensation evoked and passed onto NP2 as shown bolded in (14). Some of them collocate with into (e.g., intimidate, scare, and galvanize).
This domain is almost uniformly co-occurs with human subjects (93%).
(14a) It happened in the 1970s when the government panicked itself into an "energy crisis,"
(WSJ-V202)
(14a) …deluded ourselves into thinking we were safe. (WSJ-V70)
In contrast, the result for cognition verbs is rather difficult for interpretation as the category could have covered too broad a range. Verbs like trick, sorted, plugged, instilling, reclassify, parsing, and categorizing, clearly denote change at the cognitive level and are supposedly assigned to the verb-attached group. In brief, the above annotation results show that the co-attached instances for the into PP are more commonly associated with interpersonal functions such as communication, social and emotion.
6 Conclusion
In this study, we adopt the construction grammar framework to provide a different means to reformulate the PP attachment problem. In addition to the conventional approach that makes a
structure that requires a co-attachment to both the verb and noun. By exploiting the linguistic properties of caused-motion construction, we propose the use the semantic feature [movement]
to parse tree structure. The co-attached structure lacks the [movement] feature at both V and NP1 because the sense of motion resides in the construction per se rather than being imposed on the attachment sites. The results indicate that 11.3% (208 instances) of the 1835 types of verbs in V NP1 into NP2 constructions extracted from the WSJ corpus are co-attached. It is therefore worthy of further consideration in NLP tasks involving PP-attachment.
However, there are some limitations to the feature specification approach of this study. First, more stringent criteria for feature annotation are necessary. For example, some words in communication, cognition and social interaction domains denote rhetorical forces (e.g., entice, allure, pressure) and their movement feature may have been overlooked. Furthermore, refinement on the undermined category [+, +] is necessary to provide more accurate figures to support our approach. Future work should also include analysis of the nouns in depth, and extend the results of this study to other prepositions and PPs in other constructions.
References
Carrier, J. and J. Randall. 1992. The argument structure and syntactic structure of resultatives.
Linguistic Inquiry, 23(2), 173–234.
Charniak, E., D. Blaheta, N. Ge, K. Hall, J. Hale and M. Johnson. 2000. BLLIP 1987-89 WSJ Corpus Release 1. Linguistic Data Consortium, Philadelphia
Fellbaum, C. 1990. English verbs as a semantic net. International Journal of Lexicography, 3(4), 278-301.
Fodor, J. 1970. Three reasons for not deriving “kill” from “cause to die”. Linguistic Inquiry, 1(4), 429–438.
Goldberg. A. 1995. Constructions: A construction grammar approach to argument structure.
Chicago: The University of Chicago Press.
Hindle, D. and M. Rooth. 1993. Structural ambiguity and lexical relations. Computational Linguistics, 19(1), 103-120.
Jackendoff, R. 2002. Foundations of language: Brain, meaning, grammar, and evolution. Oxford:
Oxford University Press.
Litkowski, K. and O. Hargraves. 2007. Sem-Eval-2007 Task 06: Word-sense disambiguation of prepositions. Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pp. 24-29.
Merlo, P. and E. E. Ferrer. 2006. The notion of argument in prepositional phrase attachment.
Computational Linguistics, 32(3), 341-377.
Radford. A. 2004. Minimalist syntax: Exploring the structure of English. Cambridge:
Cambridge University Press.
Pollard, C. and I. A. Sag. 1994. Head-driven phrase structure grammar. Chicago, IL:
University of Chicago Press.
Talmy, L. (2000). Toward a cognitive semantics. Vol. II: Typology and process in concept structuring. Cambridge, MA: The MIT Press.
Tyler, A. and V. Evans. 2003. The semantics of English prepositions. New York: Cambridge University Press.
Volk. M. 2006. How bad is the problem of PP-attachment? A comparison of English, German and Swedish. Proceedings of the Third ACL-SIGSEM Workshop on Prepositions, pp. 81-88, Trento, Italy, April 2006.
Wechsler, S. and B. Noh. 2001. On resultative predicates and clauses: Parallels between Korean and English resultatives. Language Sciences, 23, 391-423.
Perspektiven auf Sprache Corpus Linguistics and
Interdisciplinary Perspectives on Language
Bd. / Vol. 1
Herausgeber / Editorial Board:
Holger Keibel, Marc Kupietz, Christian Mair Gutachter / Advisory Board:
Heike Behrens, Mark Davies, Martin Hilpert,
Reinhard Köhler, Ramesh Krishnamurthy, Ralph Ludwig, Michaela Mahlberg, Tony McEnery, Anton Näf,
Michael Stubbs, Elke Teich, Heike Zinsmeister
Christian Mair / František Štícha Ulrich H. Waßner (Hgg.)