Infinitus Systems · NLP Researcher Intern · June–September 2023
This project explored how graph structure and language representations can support next-action prediction in complex phone calls. The work was published as Graph Integrated Language Transformers for Next Action Prediction in Complex Phone Calls at NAACL 2024.
Project overview
We developed a GPT-empowered graph neural network that modeled relationships among dialogue turns, entities, and actions in call-automation workflows. The approach was designed to help an AI assistant select appropriate next actions while preserving the context of a long, multi-step conversation.
Selected contributions
- Designed a graph-integrated language modeling approach that improved the successful-call rate by 30%.
- Developed data preparation and automated annotation workflows for human–AI dialogue turns.
- Co-designed evaluation analysis for LLM-based call-automation quality and next-action prediction.
Why it matters
The project connected structured reasoning with language understanding, providing a path toward more reliable and efficient automation in conversations where the correct next action depends on several earlier turns.