Unsupervised Paraphrasability Prediction for Compound Nominalizations

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3254 - 3263 July 10-15, 2022 ©2022 Association for Computational Linguistics Unsupervised Paraphrasability Prediction for Compound Nominalizations John S. Y. Lee, Ho Hung Lim Department of Linguistics and Translation City University of Hong Kong Hong Kong SAR, China jsylee@cityu.edu.hk hhlim3@cityu.edu.hk Carol Webster College of Professional and Continuing Education The Hong Kong Polytechnic University Hong Kong SAR, China carol.webster@cpce-polyu.edu.hk Abstract Commonly found in academic and formal texts, a nominalization uses a deverbal noun to describe an event associated with its corre- sponding verb. Nominalizations can be diffi- cult to interpret because of ambiguous seman- tic relations between the deverbal noun and its arguments. Automatic generation of clausal paraphrases for nominalizations can help dis- ambiguate their meaning. However, previous work has not identified cases where it is awk- ward or impossible to paraphrase a compound nominalization. This paper investigates unsu- pervised prediction of paraphrasability, which determines whether the prenominal modifier of a nominalization can be re-written as a noun or adverb in a clausal paraphrase. We adopt the approach of overgenerating candi- date paraphrases followed by candidate rank- ing with a neural language model. In exper- iments on an English dataset, we show that features from an Abstract Meaning Represen- tation graph lead to statistically significant im- provement in both paraphrasability prediction and paraphrase generation. 1 Introduction A nominalization is a noun (e.g., “response”) that is morphologically derived from a verb (“respond”), and that designates some aspects of the event re- ferred to by the verb (Quirk et al., 1985). In a com- pound nominalization, this deverbal noun may have both prenominal and postnominal modifiers. The prenominal modifier can be a noun (e.g., “police response to the rioting”) or an adjective (“bodily injury to a friend”), while postnominal modifiers are prepositional phrases (“presidential nomination of Harrison”). Academic and other formal texts utilize nomi- nalization extensively to produce a compact and abstract writing style. The meaning of compound nominalizations can however be difficult to inter- pret because of ambiguous semantic relations be- tween the deverbal noun and its modifiers. In par- ticular, the prenominal modifier can play multi- ple semantic roles in the corresponding predicate or clausal paraphrase: as a subject (e.g., “the po- lice response” → “the police responds”); as an object (“bodily injury” → “injure the body”); as an oblique (“presidential nomination” → “nomi- nate as president”; as an adverb (“symbolic admis- sion” → “admit symbolically”; or none of the above (“stellar performance” ̸→ “a star performs”). The paraphrasability of the prenominal modi- fier — whether it describes an entity, the manner of an action, or neither — therefore has direct im- pact on NLP tasks that require interpretation of compound nominalizations. This ambiguity affects accuracy in relation extraction, which is impor- tant for information retrieval and question answer- ing (Greenwood, 2004; Klein et al., 2020). A ma- chine translation system must also be able to ren- der the deverbal noun …

Content truncated to 500 words