Neuramanteau: A Neural Network Ensemble Model for Lexical Blends

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 576–583, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Neuramanteau: A Neural Network Ensemble Model for Lexical Blends Kollol Das kolloldas@gmail.com Independent Researcher Bangalore, India Shaona Ghosh sg811@cam.ac.uk University of Cambridge Department of Engineering, Cambridge, UK Abstract The problem of blend formation in genera- tive linguistics is interesting in the context of neologism, their quick adoption in mod- ern life and the creative generative pro- cess guiding their formation. Blend qual- ity depends on multitude of factors with high degrees of uncertainty. In this work, we investigate if the modern neural net- work models can sufficiently capture and recognize the creative blend composition process. We propose recurrent neural net- work sequence-to-sequence models, that are evaluated on multiple blend datasets available in the literature. We propose an ensemble neural and hybrid model that outperforms most of the baselines and heuristic models upon evaluation on test data. 1 Introduction Blending is the formation and intentional coinage of new words from existing two or more words (Gries, 2004). These are called neolo- gisms. Neologisms effectively trace changing cul- tures and addition of new technologies. Blend- ing is one way to create a neologism. A lexi- cal blend is formed by combining parts of two or more words. Predicting a high quality lexi- cal blend is often unpredictable due to the un- certainty in the formal structure of blends (Beli- aeva, 2014). Merging source words digital and camera, the common expectation is the blend digamera, although, in practice, digicam is more common (Beliaeva, 2014). There are mul- tiple unknown factors such as phonology, seman- tics, familiarity, recognizability and lexical cre- ativity that contribute to blend formation. Some downstream applications that can leverage such Word1 Word2 Blend Structure Category Coverage aviation electronics avionics avi-onics Prefix + Suffix 16.74% communicate fake communifake communi-fake Prefix + Word 10.78% speak typo speako speak-o Word + Letter 0.14% west indiea windies w-indies Letter + Word 0.89% point broadcast pointcast point-cast Word + Suffix 22.56% scientific fiction scientifiction scienti-fic-tion Word + Word overlap 22.56% affluence influenza affluenza af-fluen-za Prefix + Suffix overlap 13.98% brad angelina brangelina br-a-ngelina Prefix + Word overlap 11.39% subvert advertising subvertising sub-vert-ising Word + Suffix overlap 16.31% Table 1: Sample blends in our dataset along with the type and coverage. There are other types of rare blends that is beyond the scope of this work. blending systems include generating names of products, brands, businesses and advertisements to name a few especially if coupled with contextual information about the business or sector. 1.1 Related Work Blends are compositional words consisting of whole word and a splinter (part of morpheme) or two splinters (Lehrer, 2007). The creative neoglo- gism of blend formation has been studied by lin- guists in an attempt to recognize patterns in the process that model human lexical creativity or to identify source words and blend meaning, context and influence. With the popularity of deep neu- ral networks (DNN)s (LeCun …

Content truncated to 500 words