Building a Task-oriented Dialog System for Languages with no Training Data: the Case for Basque

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 2796–2802 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 2796 Building a Task-oriented Dialog System for languages with no training data: the Case for Basque Maddalen L´opez de Lacalle, Xabier Saralegi, I˜naki San Vicente Elhuyar Foundation Osinalde Industrialdea 3, 20170 Usurbil, Spain {m.lopezdelacalle, x.saralegi, i.sanvicente}@elhuyar.eus Abstract This paper presents an approach for developing a task-oriented dialog system for less-resourced languages in scenarios where no training data is available. Both intent classification and slot filling are tackled. We project the existing annotations in rich-resource languages by means of Neural Machine Translation (NMT) and posterior word alignments. We then compare training on the projected monolingual data with direct model transfer alternatives. Intent Classifiers and slot filling sequence taggers are implemented using a BiLSTM architecture or by fine-tuning BERT transformer models. Models learnt exclusively from Basque projected data provide better accuracies for slot filling. Combining Basque projected train data with rich-resource languages data outperforms consistently models trained solely on projected data for intent classification. At any rate, we achieve competitive performance in both tasks, with accuracies of 81% for intent classification and 77% for slot filling. Keywords: Dialog Systems, Neural language representation models, Less-resourced languages 1. Introduction Task-oriented dialog systems are one of the most trending topics in today’s Natural Language Processing. Such sys- tems are focused on helping users with specific tasks such as booking a ticket or a restaurant, scheduling a reminder for a call, etc. Task-oriented dialog systems understand what users mean when they speak or write an utterance (the request of the user) by identifying users intents and the corresponding slots. Intents are what the user wants, and slots are the key entities corresponding to the intent. Let’s take a look at example 1. The user’s intent is to set a new reminder. Intents are given a name such as ”set reminder” in this case. Intent-arguments or slots modify the intent, which in this case are ”to take out the garbage” and ”on Thurs- day”. Intent-arguments are given a name, such as ”to do” and ”datetime” in our example. Example 1. Remind me [to take out the garbage] [on Thursday] The main tasks involved in a task-oriented dialog system are intent classification and slot filling, which can be stated as sentence classification and sequence labeling problems respectively. These two tasks can be learnt in a joint fash- ion or as two separate tasks. Most successful approaches proposed for both strategies rely on deep learning algo- rithms (Xu and Sarikaya, 2013; Mesnil et al., 2013; Liu and Lane, 2016; Goo et al., 2018), and hence, they require large training datasets. Unfortunately, building labeled di- alog datasets involves a great manual effort, rendering it unfeasible for non major languages. Scenarios where task- oriented dialog systems must deal with many languages can experiment a similar problem, because the cost for building training datasets multiplies. In this work we deal with the problem of building …

Content truncated to 500 words