Projecting Embeddings for Domain Adaption: Joint Modeling of Sentiment Analysis in Diverse Domains

Proceedings of the 27th International Conference on Computational Linguistics, pages 818–830 Santa Fe, New Mexico, USA, August 20-26, 2018. 818 Projecting Embeddings for Domain Adaptation: Joint Modeling of Sentiment Analysis in Diverse Domains Jeremy Barnes1,2, Roman Klinger1, and Sabine Schulte im Walde1 1Institut f¨ur Maschinelle Sprachverarbeitung University of Stuttgart Pfaffenwaldring 5b, 70569 Stuttgart, Germany {barnesjy,klinger,schulte}@ims.uni-stuttgart.de 2Grup de Ling¨u´ıstica Computacional Universitat Pompeu Fabra Roc Boronat 138, 08018 Barcelona, Spain jeremy.barnes@upf.edu Abstract Domain adaptation for sentiment analysis is challenging due to the fact that supervised classifiers are very sensitive to changes in domain. The two most prominent approaches to this problem are structural correspondence learning and autoencoders. However, they either require long training times or suffer greatly on highly divergent domains. Inspired by recent advances in cross-lingual sentiment analysis, we provide a novel perspective and cast the domain adaptation problem as an embedding projection task. Our model takes as input two mono-domain embedding spaces and learns to project them to a bi-domain space, which is jointly optimized to (1) project across domains and to (2) predict sentiment. We perform domain adaptation experiments on 20 source-target domain pairs for sentiment classification and report novel state-of-the-art results on 11 domain pairs, including the Amazon domain adaptation datasets and SemEval 2013 and 2016 datasets. Our analysis shows that our model performs comparably to state-of-the-art approaches on domains that are similar, while performing significantly better on highly divergent domains. Our code is available at https://github.com/jbarnesspain/domain_blse Title and Abstract in Basque Domeinu-Egokitzapenerako Bektore Proiekzioa: Domeinu Urrunetarako Sentimenduen Analisiko Eredu Bateratua Sentimenduen analisirako domeinu-egokitzapena erronka handi bat da oraindik, domeinu arteko ezberdintasunek ondorio esanguratsuak izan baititzakete sailkatzaile gainbegiratuentzat. Arazo honi aurre egiteko bi hurbilpen arrakastatsuenak egiturazko kidetasunaren ikasketa (structural correspondence learning) eta autoencoder-ak dira. Hala ere, denbora asko behar dute sistema entrenatzeko edo, domeinu arteko distantzia handia denean, ez dituzte emaitza onak lortzen. Hizkuntza-arteko sentimenduen analisian egindako azken lanetan oinarrituta, ikuspuntu berri bat eskaintzen dugu, domeinuaren egokitzapen ataza bektore proiekzio ataza gisa planteatuta. Gure sistemaren sarrera domeinu banako bi bektore espazio dira, zeinak sistemak espazio berri batera proiektatzen ikasten duen. Sistema hau optimizatuta dago (1) domeinu batetik besterako proiekzioa egiteko eta (2) esaldi baten sentimendua aurresateko. 20 jatorri-xede domeinu pareetan esperimentuak burutu ditugu eta 11 kasutan artearen egoerako emaitzarik onenak lortzen ditugu Amazon-eko domeinu-egokitzapeneko eta SemEval 2013 eta 2016 datu-multzoetan. Gure anal- isian ikus daitekeenez, gure hurbilpena artearen egoerako sistemen pareko moldatzen da antzeko domeinuetan, baina emaitza hobeak lortzen ditu oso domeinu ezberdinetan. Gure kodea eskura- garri dago helbide honetan: https://github.com/jbarnesspain/domain_blse. This work is licenced under a Creative Commons Attribution 4.0 International Licence. Licence details: http:// creativecommons.org/licenses/by/4.0/ 819 1 Introduction One of the main limitations of current approaches to sentiment analysis is that they are sensitive to differences in domain. This leads to classifiers that, after training, perform poorly on new domains (Pang and Lee, 2008; Deriu et al., 2017). Domain adaptation techniques provide a solution to reduce the discrepancy and enable models to perform well across multiple domains (Blitzer et al., 2007). The two main approaches to …

Content truncated to 500 words