Vector-space calculation of semantic surprisal for predicting word pronunciation duration

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 763–773, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Vector-space calculation of semantic surprisal for predicting word pronunciation duration Asad Sayeed, Stefan Fischer, and Vera Demberg Computational Linguistics and Phonetics/M2CI Cluster of Excellence Saarland University 66123 Saarbr¨ucken, Germany {asayeed,sfischer,vera}@coli.uni-saarland.de Abstract In order to build psycholinguistic mod- els of processing difficulty and evaluate these models against human data, we need highly accurate language models. Here we specifically consider surprisal, a word’s predictability in context. Existing ap- proaches have mostly used n-gram models or more sophisticated syntax-based pars- ing models; this largely does not account for effects specific to semantics. We build on the work by Mitchell et al. (2010) and show that the semantic prediction model suggested there can successfully predict spoken word durations in naturalistic con- versational data. An interesting finding is that the training data for the semantic model also plays a strong role: the model trained on in- domain data, even though a better lan- guage model for our data, is not able to predict word durations, while the out-of- domain trained language model does pre- dict word durations. We argue that this at first counter-intuitive result is due to the out-of-domain model better matching the “language models” of the speakers in our data. 1 Introduction The Uniform Information Density (UID) hypothe- sis holds that speakers tend to maintain a relatively constant rate of information transfer during speech production (e.g., Jurafsky et al., 2001; Aylett and Turk, 2006; Frank and Jaeger, 2008). The rate of information transfer is thereby quantified using as each words’ Surprisal (Hale, 2001), that is, a word’s negative log probability in context. Surprisal(wi) = − log P(wi|w1..wi−1) This work makes use of an existing measure of semantic surprisal calculated from a distributional space in order to test whether this measure ac- counts for an effect of UID on speech production. Our hypothesis is that a word in a semantically surprising context is pronounced with a slightly longer duration than the same word in a seman- tically less-expected context. In this way, a more uniform rate of information transfer is achieved, because the higher information content of the un- expected word is stretched over a slightly longer time. To our knowledge, the use of this form of surprisal as a pronunciation predictor has never been investigated. The intuition is thus: in a sentence like the sheep ate the long grass, the word grass will have relatively high surprisal if the context only con- sists of the long. However, a distributional repre- sentation that retains the other content words in the sentence, thus representing the contextual similar- ity of grass to sheep ate, would able to capture the relevant context for content word prediction more easily. In the approach taken here, both types of models are combined: a standard language model is reweighted with semantic similarities in order to capture both …

Content truncated to 500 words