Comparative Study of Sentence Embeddings for Contextual Paraphrasing

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 6841–6851 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 6841 Comparative Study of Sentence Embeddings for Contextual Paraphrasing Louisa Pragst1, Wolfgang Minker1, Stefan Ultes2 1Ulm University, Ulm, Germany, {louisa.pragst, wolfgang.minker}@uni-ulm.de 2Mercedes-Benz Research & Development, Sindelfingen, Germany, stefan.ultes@daimler.com Abstract Paraphrasing is an important aspect of natural-language generation that can produce more variety in the way specific content is presented. Traditionally, paraphrasing has been focused on finding different words that convey the same meaning. However, in human-human interaction, we regularly express our intention with phrases that are vastly different regarding both word content and syntactic structure. Instead of exchanging only individual words, the complete surface realisation of a sentences is altered while still preserving its meaning and function in a conversation. This kind of contextual paraphrasing did not yet receive a lot of attention from the scientific community despite its potential for the creation of more varied dialogues. In this work, we evaluate several existing approaches to sentence encoding with regard to their ability to capture such context-dependent paraphrasing. To this end, we define a paraphrase classification task that incorporates contextual paraphrases, perform dialogue act clustering, and determine the performance of the sentence embeddings in a sentence swapping task. Keywords: Sentence Similarity, Sentence Embeddings, Sentence Encoding, Paraphrasing 1. Introduction Humans show a lot of variance in the way they express themselves in a conversation. Not only do they change their phrasing by exchanging words with similar meaning for each other, often whole sentences are used interchange- ably although they do not have much in common on a sur- face level. For example, ‘When will you be home?’ can be answered by ‘At six.’, ‘At six o’clock.’ or ‘Around six.’. Those variants can be considered word-level paraphrases. However, the sentence ‘I just got on the bus.’ can fulfil the same function of providing the expected time of ar- rival, while differing greatly with regard to syntactic struc- ture and the words used. We refer to such sentences that can fulfil the same function in the context of a conversa- tion despite being dissimilar in their surface realisation as contextual paraphrases. In the area of dialogue systems, a number of contributions are concerned with the generation of variety in the utter- ances of the dialogue system (e.g (Wen et al., 2015; Ko- zlowski et al., 2003; Langkilde and Knight, 1998)). How- ever, those efforts are mainly focused on word-level para- phrases and little work has been dedicated to the generation of contextual paraphrases. Pragst and Ultes (2018) have made a first effort in this direction by proposing an ap- proach for exchanging sentences, using a dialogue vector model to assess whether two sentences can be used inter- changeably. They find that the ability to identify contex- tual paraphrases is of great importance to the overall per- formance of the approach. However, the model used in that work is highly specialised to the evaluation corpus and un- …

Content truncated to 500 words