Enhancing NLU with Paraphrasing in Task-Oriented Dialogue Systems Toward Data-Efficient Generalization in Low-Resource Scenarios

Abstract

In task-oriented dialogue systems (TODS), Natural Language Understanding (NLU) is fundamental to interpreting user intent and extracting semantic information. However, training robust NLU models often requires large amounts of annotated data, which is costly and difficult to obtain in many real-world domains. While large language models (LLMs) have demonstrated impressive zero- and few-shot performance, their deployment in NLU pipelines remains constrained by high inference cost, cloud dependency, and lack of controllability—especially in latency-sensitive or privacy-critical settings. This thesis investigates an alternative approach: using LLMs as offline paraphrase generators to augment training data for lightweight NLU models. We evaluate this method across three core NLU tasks—domain classification, intent recognition, and slot labeling— on two benchmark datasets: Banking77 and MultiWOZ 2.2. Results show that paraphrasing improves model generalization under data-scarce conditions, particularly for intent recognition. For slot labeling, however, performance gains are limited due to span misalignment in paraphrased utterances. Our slot-level analysis reveals that categorical slots (e.g., food type, price range) are more robust to paraphrasing than span-sensitive slots (e.g., time, location). These findings suggest that LLM-based paraphrasing is a practical and resource-efficient augmentation strategy, but its success depends on task granularity, label structure, and alignment fidelity.