Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Conversational Question Answering with Reformulations over Knowledge Graph

Lihui Liu, Blaine Hill, Boxin Du, Fei Wang, Hanghang Tong

TL;DR

Confronts the challenge of inexplicit, multi-turn questions in knowledge-graph QA by introducing CoRnNet, an RL-based framework that exploits reformulations. A teacher–student architecture aligns reformulation representations from human writing and LLM-generated reformulations, while a knowledge-grounded RL agent traverses the KG to locate answers, aided by a knowledge-based soft reward. The approach achieves state-of-the-art results on ConvQuestions and ConvRef, with ablations showing substantial gains from reformulations and the teacher–student distillation, and improved efficiency relative to baselines. This work advances practical convQA over KGs by effectively leveraging reformulations to disambiguate queries and guide evidence-grounded reasoning, with potential for scalable deployment in real-world systems.

Abstract

Conversational question answering (convQA) over knowledge graphs (KGs) involves answering multi-turn natural language questions about information contained in a KG. State-of-the-art methods of ConvQA often struggle with inexplicit question-answer pairs. These inputs are easy for human beings to understand given a conversation history, but hard for a machine to interpret, which can degrade ConvQA performance. To address this problem, we propose a reinforcement learning (RL) based model, CornNet, which utilizes question reformulations generated by large language models (LLMs) to improve ConvQA performance. CornNet adopts a teacher-student architecture where a teacher model learns question representations using human writing reformulations, and a student model to mimic the teacher model's output via reformulations generated by LLMs. The learned question representation is then used by an RL model to locate the correct answer in a KG. Extensive experimental results show that CornNet outperforms state-of-the-art convQA models.

Conversational Question Answering with Reformulations over Knowledge Graph

TL;DR

Confronts the challenge of inexplicit, multi-turn questions in knowledge-graph QA by introducing CoRnNet, an RL-based framework that exploits reformulations. A teacher–student architecture aligns reformulation representations from human writing and LLM-generated reformulations, while a knowledge-grounded RL agent traverses the KG to locate answers, aided by a knowledge-based soft reward. The approach achieves state-of-the-art results on ConvQuestions and ConvRef, with ablations showing substantial gains from reformulations and the teacher–student distillation, and improved efficiency relative to baselines. This work advances practical convQA over KGs by effectively leveraging reformulations to disambiguate queries and guide evidence-grounded reasoning, with potential for scalable deployment in real-world systems.

Abstract

Conversational question answering (convQA) over knowledge graphs (KGs) involves answering multi-turn natural language questions about information contained in a KG. State-of-the-art methods of ConvQA often struggle with inexplicit question-answer pairs. These inputs are easy for human beings to understand given a conversation history, but hard for a machine to interpret, which can degrade ConvQA performance. To address this problem, we propose a reinforcement learning (RL) based model, CornNet, which utilizes question reformulations generated by large language models (LLMs) to improve ConvQA performance. CornNet adopts a teacher-student architecture where a teacher model learns question representations using human writing reformulations, and a student model to mimic the teacher model's output via reformulations generated by LLMs. The learned question representation is then used by an RL model to locate the correct answer in a KG. Extensive experimental results show that CornNet outperforms state-of-the-art convQA models.
Paper Structure (20 sections, 15 equations, 3 figures, 5 tables)

This paper contains 20 sections, 15 equations, 3 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Problem Definition
  3. Preliminaries: Reinforcement Learning
  4. Proposed Method
  5. Student Model: LLMs Reformulation Encoder.
  6. Teacher Model: Human Writing Reformulation Encoder
  7. Inferring the Topic Entity
  8. Question Answering
  9. Policy Network
  10. Knowledge-Based Soft Reward
  11. Training
  12. Experiments
  13. Experimental Setting
  14. Main Results
  15. Ablation Studies and Efficiency Results
  16. ...and 5 more sections

Figures (3)

  • Figure 1: Training Process of CoRnNet. The light gray part denotes the architecture of the student model. The light green part shows the framework of RL-based question answering model. Both are trained end-to-end.
  • Figure 2: Reformulation imitator.
  • Figure 3: CoRnNet Training and Test Time.