Representing Sentence Interpretations with Overlapping Box Embeddings

ACL ARR 2025 February Submission1784 Authors

Abstract: Most of the previous studies on sentence em- beddings aim to obtain a single representation per sentence. However, this approach is in- adequate for handling the semantic relations between sentences when a sentence has multi- ple interpretations. To address this problem, we propose a novel concept, interpretation embed- dings, which are representations of the interpre- tations of a sentence. We propose GumbelCSE, which is a contrastive learning method for learn- ing box embeddings of sentences. The inter- pretation embeddings are derived by measuring the overlap between the box embeddings of the target sentence and those of other sentences. We evaluate our method on three tasks: Recog- nizing Textual Entailment (RTE), Entailment Direction Prediction, and Ambiguous RTE. On the RTE and Entailment Direction Prediction tasks, GumbelCSE outperforms baseline sen- tence embedding methods in most cases. In the Ambiguous RTE task, it is demonstrated that the interpretation embeddings are effective in capturing the ambiguity of meaning inherent in a sentence.

Paper Type: Long

Research Area: Semantics: Lexical and Sentence-Level

Research Area Keywords: Machine Learning for NLP, Semantics: Lexical and Sentence-Level

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 1784