Efficient Task-Oriented Dialogue State Tracking using Language Models

Abstract

In the digital era, humans depend on automated systems to accomplish a wide variety of tasks that usedto require talking to a human agent, such as booking flights and hotels for their trips. This demand has driven the development of AI assistants that enable users to interact with databases and APIs through conversational interfaces. Training these AI assistants requires large amounts of data, and annotating conversational data is expensive. Therefore, developing models that can learn from small amounts of data is crucial. As the scale of large language models (LLM) driving AI assistants continues to grow, their computational requirements also increase significantly. Thus, the key challenges this thesis addresses are creating computation-efficient and data-efficient models that maintain high task performance. A core component of a task-oriented AI assistant is Dialogue State Tracking (DST), which deciphersuser intents from conversational histories, thereby mapping user goals to predefined schema for executing task-specific queries. In this thesis, we introduce our proposed solutions to the critical challenges in building computation-efficient and data-efficient dialogue state tracking systems. The initial approach concerns how to build effective DST systems with smaller LMs (SLM) that have lower computation requirements. To facilitate this line of research, we first introduce a prompt-based finetuning framework SDP-DST, that enriches SLM inputs with schema-specific textual information. This enables task-aware contextualization of the input conversations and significantly benefits SLM-based DST systems without losing generalization abilities. This schema-driven prompting framework advances the SOTA on a DST dataset despite using fewer model parameters. Second, learning efficiently from conversations is critical to facilitate building task-oriented agents onnew services that have limited annotated data. We introduce a few-shot learning framework IC-DST, that leverages the in-context learning (ICL) abilities of LLM where models can make predictions given task instructions or a few exemplars. This framework leverages a prompt format with DST ontology descriptions and reformulates DST as a text-to-SQL task to improve ICL performance, advancing the state-of-the-art (SOTA) few-shot DST without any parameter updates. To combine the benefits of the two previous approaches, we propose a novel routing framework ORCHESTRALLM, that simultaneously leverages an SLM and an LLM along with a router that dynamically assigns instances to either LM in inference time to save computing. Hypothesizing that conversational turns with similar semantic embeddings are of the same difficulty level, the router selects an appropriate LM expert based on embedding distances between the testing instance and instances in pre-stored LM expert pools. We demonstrate the effectiveness of the routing framework on two different multi-domain DST benchmarks under low-resource settings. The SLM can be trained with less data when combined with an LLM to handlemore difficult cases. The proposed framework capitalizes on the expertise of different LMs, outperforming standalone systems while also achieving a substantial reduction of over 50% in computational costs. Lastly, we propose an alternative strategy to achieve better data efficiency and computation efficiencyusing a self-correction system with only an SLM, capitalizing on recent technological advances that have made SLMs more powerful. LLMs have been developed to make corrections to their outputs through self- provided (or external) feedback when prompted, particularly in code and math reasoning tasks. While self-correction improves task performance, it requires multiple inference passes from the LLM, including feedback/verification outputs and final correction outputs. To enhance the efficiency of DST systems, we introduce a novel paradigm CORRECTIONLM, where we fine-tune SLMs to correct predictions using a small amount of annotated data, alleviating the need for feedback from LLMs. Evaluation results on two DST datasets reveal that SLMs can correct errors effectively.