Evaluating Composition Models for Verb Phrase Elliptical Sentence Embeddings

Proceedings of NAACL-HLT 2019, pages 261–271 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 261 Evaluating Composition Models for Verb Phrase Elliptical Sentence Embeddings Gijs Wijnholds Queen Mary University of London g.j.wijnholds@qmul.ac.uk Mehrnoosh Sadrzadeh Queen Mary University of London mehrnoosh.sadrzadeh@qmul.ac.uk Abstract Ellipsis is a natural language phenomenon where part of a sentence is missing and its information must be recovered from its sur- rounding context, as in “Cats chase dogs and so do foxes.”. Formal semantics has differ- ent methods for resolving ellipsis and recov- ering the missing information, but the prob- lem has not been considered for distributional semantics, where words have vector embed- dings and combinations thereof provide em- beddings for sentences. In elliptical sentences these combinations go beyond linear as copy- ing of elided information is necessary. In this paper, we develop different models for embedding VP-elliptical sentences. We ex- tend existing verb disambiguation and sen- tence similarity datasets to ones containing el- liptical phrases and evaluate our models on these datasets for a variety of non-linear com- binations and their linear counterparts. We compare results of these compositional mod- els to state of the art holistic sentence en- coders. Our results show that non-linear addi- tion and a non-linear tensor-based composition outperform the naive non-compositional base- lines and the linear models, and that sentence encoders perform well on sentence similarity, but not on verb disambiguation. 1 Introduction Compositional distributional semantics has so far relied on a tight connection between syntactic and semantic resources. Based on the assembly prin- ciple of compositionality, these models assign a sentence vector by applying a linear map to the in- dividual word embeddings therein. The meaning of “cats chase dogs” is as follows in (1) additive, (2) multiplicative, and (3) tensor-based models: (1) −−→ cats + −−−→ chase + −−→ dogs (2) −−→ cats ⊙ −−−→ chase ⊙ −−→ dogs (3) −−→ cats⊤ × (chase × −−→ dogs) Some linguistic phenomena, however, rely on copying resources while computing meaning; canonical examples thereof are anaphora and el- lipsis, exemplified below: (a) Cats clean themselves. (b) Cats chase dogs, children do too. More complex examples involve a structural am- biguity such as the following: (c) Cats chase their tail, dogs too. These lend themselves to a strict (dogs chase the cat’s tail) and a sloppy reading (dogs chase their own tail). In these examples, the meaning of at least one part of the sentence is used twice, e.g. the subject in a, the verb phrase “chase dogs” in b. Such cases can often be extended to a situation in which a meaning is used more than twice, e.g. in “Cats chase their tail, dogs too, and so do foxes”. In order to develop distributional semantics for such sentences while respecting the principle of compositionality, one has a choice between a lin- ear or a non-linear composition of resources. In the linear case, no information is copied, resulting in vector embeddings such as the following one (when only considering …

Content truncated to 500 words