Cross-lingual intent classification in a low resource industrial setting

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 6419–6424, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 6419 Cross-lingual intent classification in a low resource industrial setting Talaat Khalil* Kornel Kiełczewski* Georgios Christos Chouliaras Amina Keldibek Maarten Versteegh Booking.com Amsterdam {talaat.khalil,kornel.kielczewski}@booking.com Abstract This paper explores different approaches to multilingual intent classification in a low re- source setting. Recent advances in multi- lingual text representations promise cross- lingual transfer for classifiers. We investigate the potential for this transfer in an applied in- dustrial setting and compare to multilingual classification using machine translated text. Our results show that while the recently de- veloped methods show promise, practical ap- plication calls for a combination of techniques for useful results. 1 Introduction Classifying the intent represented in a customer message is a core functionality of natural language understanding applications in customer service. Once the intent of the customer’s message is un- derstood and various entities are resolved, actions can be triggered, such as automatically replying to the customer, routing the message to the right customer service representative, or asking the cus- tomer for more information. The rise of automation in customer service and the growing customer expectation of immediate answers is generating increasing demand for accu- rate intent detection across many languages. Re- cent advances in natural language processing have enabled impressive performance on text classifica- tion tasks, provided that large-scale labelled data are available. In industrial settings, this is often not the case, or at least not for all languages for which intent detection is required. This paper studies what we believe to be a com- mon case in industry, in which a small amount of labeled data are available for a single language, * These authors contributed equally. say English, but neither data nor labels are avail- able for other languages. This setting poses an in- teresting constraint on the application of modern machine learning techniques, in that much fewer in-domain data are available for the task than is generally required. We believe that the findings in this paper can be generalized outside our spe- cific industrial domain and apply in general to the challenge of enabling classification in a target lan- guage based on supervision in a source language. This paper investigates the application of two techniques from the literature to this problem. The first technique is an out-of-domain machine trans- lation system, as can be assumed to be generally available, applied to the source language data set in order to generate supervised data in the target language. The second technique that we inves- tigate is transfer learning based on multi-lingual document representations. We study the results of applying these techniques, as well as the combina- tion of both, in a number of experiments. 2 Related work This study is broadly positioned in the application of cross-lingual transfer, with special focus on low resource applications. We study two approaches to this problem: machine …

Content truncated to 500 words