Low-Resource Natural Language Understanding in Task-Oriented Dialogue

Task-oriented dialogue (ToD) systems need to interpret the user's input to understand the user's needs (intent) and corresponding relevant information (slots). This process is performed by a Natural Language Understanding (NLU) component, which maps the text utterance into a semantic frame representation, involving two subtasks: intent classification (text classification) and slot filling (sequence tagging). Typically, new domains and languages are regularly added to the system to support more functionalities. Collecting domain-specific data and performing fine-grained annotation of large amounts of data every time a new domain and language is introduced can be expensive. Thus, developing an NLU model that generalizes well across domains and languages with less labeled data (low-resource) is crucial and remains challenging. This thesis focuses on investigating transfer learning and data augmentation methods for low-resource NLU in ToD. Our first contribution is a study of the potential of non-conversational text as a source for transfer. Most transfer learning approaches assume labeled conversational data as the source task and adapt the NLU model to the target task. We show that leveraging similar tasks from non-conversational text improves performance on target slot filling tasks through multi-task learning in low-resource settings. Second, we propose a set of lightweight augmentation methods that apply data transformation on token and sentence levels through slot value substitution and syntactic manipulation. Despite its simplicity, the performance is comparable to deep learning-based augmentation models, and it is effective on six languages on NLU tasks. Third, we investigate the effectiveness of domain adaptive pre-training for zero-shot cross-lingual NLU. In terms of overall performance, continued pre-training in English is effective across languages. This result indicates that the domain knowledge learned in English is transferable to other languages. In addition to that, domain similarity is essential. We show that intermediate pre-training data that is more similar – in terms of data distribution – to the target dataset yields better performance.