Large language models (LLMs), such as those powering the artificial intelligence (AI) chatbot ChatGPT, have come to the fore as a cornerstone technology in the bustling AI landscape, driving advancements in natural language processing (NLP). These models are designed to understand and respond to human language in a meaningful way, enabling more natural interactions between humans and machines.

At the SMU School of Computing and Information Systems (SCIS), research by Assistant Professor of Computer Science Liao Lizi revolves around priming LLMs for conversation-based applications.

“Imagine trying to ask your smart assistant to make a restaurant reservation, but it only understands a limited set of commands,” says Asst Prof Liao. “My work aims to make these assistants smarter by teaching them to understand a wider range of requests, and even anticipate what you might need next.”

Her work aligns with the broader objective of LLMs in conversational AI: to create systems that not only respond accurately to user inputs, but also engage proactively in various dialogues, making the overall user experience more pleasant and natural.

Conversational AI, driven by LLMs, is turning how we interact with technology on its head. Whether in customer service chatbots or virtual personal assistants, these models are enhancing communication, making AI more intuitive and accessible than before.

With its many challenges ahead – from the handling of ambiguous inputs, maintaining context, to ensuring unbiased interactions – Asst Prof Liao’s work boldly pushes the boundaries of what conversational AI can do.

Task-oriented dialogues in conversational AI

In task-oriented dialogues, LLMs enable AI-based systems to assist users in achieving specific goals through natural language interactions. Virtual assistants – designed to handle tasks like making restaurant reservations or furnishing a well-planned itinerary for an overseas trip—are a common product of this domain.

However, their effectiveness relies heavily on pre-defined domain ontologies, which limit their flexibility in dynamic, real-world scenarios.

“To mitigate the limitations of conversational AI, our research on task-oriented dialogues targets the adaptability aspect by allowing the system to learn and incorporate new concepts dynamically,” Asst Prof Liao explains.

Traditional methods tend to focus on detecting out-of-vocabulary values or identifying new slots using unsupervised or semi-supervised learning paradigms. These often yield noisy and arbitrary results due to their sole reliance on conversational data patterns.

To overcome this, Asst Prof Liao’s work incorporates an active learning framework that involves human-in-the-loop learning.

The method leverages existing language tools to extract value candidates, which are then used as supervision signals. A bi-criteria selection scheme is integrated to allow the system to dynamically adapt to new concepts and scenarios, enhancing performance and reliability.

Proactive conversations by AI

As the name suggests, proactive conversations involve AI systems initiating conversations. LLMs have shown remarkable proficiency in understanding context and generating responses, but they still face significant limitations.

For instance, current LLMs often fail to ask clarifying questions for ambiguous queries or refuse unreasonable requests, both of which are crucial for a conversational AI agent’s proactivity,” says Asst Prof Liao.

In a study on AI-based proactive dialogue models, Asst Prof Liao and her team explored how the models can take the initiative in conversations to ask clarifying questions and guide the interaction to achieve specific goals. Her team proposed the Proactive Chain-of-Thought (ProCoT) prompting scheme, which enhances the proactivity of LLMs by improving their goal-planning capabilities through descriptive reasoning chains.

This entails instructing the system to generate intermediate steps of reasoning and planning before deciding on the next action, such as by generating more precise questions, to improve its ability in handling proactive dialogue tasks.

Besides generating more clarifying prompts, ProCoT enables a smoother topic transition in target-guided dialogues. This mitigates another limitation in LLMs: a tendency to steer the conversation aggressively towards their designated targets.

Target-driven conversations by AI

Besides understanding user preferences, strategically nudging users towards accepting a designated item is the main goal of LLMs in target-driven conversations.

It is a unique and challenging domain within conversational AI.  However, current conversational recommendation methods often fall short in strategically suggesting specific items, focusing more on acquiring user preferences than on engaging resources that lead to user acceptance of targeted items.

Asst Prof Liao and her team addressed this issue with the development of the Reinforced Target-driven Conversational Promotion (RTCP) framework. It is designed to balance short- and long-term planning through a sophisticated gating mechanism that integrates knowledge-driven strategies and reinforcement learning rewards to guide conversation.

Based on a conversational plan, the framework predicts action tuples aimed at engaging the user, while steering the dialogue towards the targeted item.

“For instance, a virtual assistant might engage a user by discussing a popular artist, before subtly recommending a related movie. Such a framework ensures that conversations remain engaging and relevant, while promoting the target item effectively,” Asst Prof Liao illustrates.

RTCP also employs an action-guided prefix tuning method to generate responses that align with the planned dialogue strategy. This allows the system to adapt quickly to new scenarios without the need for extensive re-training—ensuring robustness and flexibility in dynamic conversational environments. The team’s experimental results have demonstrated that their framework outshines state-of-the-art models in both automatic metrics and human evaluations.

Talk of the town

As algorithms become more intelligent and computers more powerful, conversational AI will grow in importance across various domains, from customer service and healthcare to education. It is research such as that by Asst Prof Liao and her team that are instrumental in advancing LLMs to be not just more intuitive and adaptive, but also deeply engaging in their interactions.

“Looking ahead, the integration of multimodal capabilities—where systems understand and respond to multiple forms of input, such as text, voice, and visual cues—will be crucial. Smarter, more responsive systems that can anticipate and meet user needs will stand the test of time,” she concludes.

Learn more about Asst Prof Liao Lizi’s research and the exciting possibilities that large language models and artificial intelligence applications bring to real-world applications in the SMU Master of IT in Business (MITB) programme.