Research map / Label-efficient and robust learning / Semi-supervised learning
Response Enhanced Semi-supervised Dialogue Query Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SemiDQG, a semi-supervised framework that trains a response-augmented T5 query producer to pseudo-label unlabeled dialogues, filters them by similarity to a standard query producer, and refines the producer with REINFORCE using RA-based rewards.
Problem
Dialogue query generation (producing search queries from dialogue history) suffers from scarce annotated data and poor domain adaptation, and naive self-training injects low-quality pseudo queries.
Approach
A response-augmented query producer (RA) generates better pseudo queries than the standard QP; a similarity-based selection (max unigram F1 against QP's sampled queries) filters RA's pseudo labels, and RA's length-normalized log-probabilities serve as fine-grained REINFORCE rewards for QP.
Key results
- On WoW cross-domain test, SemiDQG reaches 57.12 Uni. F1 and 63.50 R@1, versus 48.11 Uni. F1 for T5-base and 41.38 for 8-shot ChatGPT.
- On KdConv cross-domain test, SemiDQG reaches 67.21 Uni. F1 versus 57.13 for T5-base and 60.10 for the best self-training variant.
- With 300 labeled WoI instances, SemiDQG outperforms a T5-base trained on 3k instances, about 10x data efficiency.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SemiDQG rather than T5-base | T5-base query producer · WoW test | Unigram F1: 48.11 -> 57.12 | Helped |
| use SemiDQG rather than QP-ext | query producer · WoW test | R@1: 62.41 -> 63.50 | Helped |
| use SemiDQG rather than QP-ext | query producer · WoW test | R@3: 72.91 -> 74.89 | Helped |
| use SemiDQG rather than KD (RA → QP) | query producer · KdConv test | Unigram F1: 65.60 -> 67.21 | Helped |
| train on RA-labeled pseudo instances rather than QP-labeled pseudo instances (ablation) | QP · KdConv test | Unigram F1: 58.43 -> 65.60 | Helped |
| add QP-probability selection to QP-labeled pseudo instances (ablation) | QP · KdConv test | Unigram F1: 58.43 -> 59.10 | Helped |
| select RA-labeled pseudo instances by similarity rather than using them without selection (ablation) | QP · KdConv test | Unigram F1: 65.60 -> 66.06 | Helped |
| select RA-labeled pseudo instances by similarity rather than RA predictive probability (ablation) | QP · KdConv test | Unigram F1: 65.73 -> 66.06 | Helped |
| select RA-labeled pseudo instances by similarity rather than QP predictive probability (ablation) | QP · KdConv test | Unigram F1: 65.01 -> 66.06 | Helped |
| add RA-guided rank-based reinforcement learning after Stage 2 (ablation) | QP with similarity-selected pseudo instances · KdConv test | Unigram F1: 66.06 -> 67.21 | Helped |
| use rank-based rather than probability-based RA rewards (ablation) | QP reinforcement learning · KdConv test | Unigram F1: 66.39 -> 67.21 | Helped |
| add RA-guided reinforcement learning directly after Stage 1, without Stage 2 (ablation) | QP · KdConv test | Unigram F1: 57.13 -> 64.39 | Helped |
Surprising findings
- Traditional self-training fails to improve QP on WoW and can decrease performance below the T5-base baseline.
- Probability-based rewards work only with small candidate sets and underperform rank-based rewards in the cross-domain setting because RA is poorly calibrated under domain shift.
Limitations
- Performance depends on the similarity threshold alpha, which varies by dataset and can hurt when the domain gap is large.
- Query-producer training relies on RA, which may generate low-quality queries when overly influenced by the response.
Experimental setup
Datasets: Wizard-of-Internet Wizard-of-Wikipedia DuSinc KdConv
Benchmarks: Wizard-of-Wikipedia KdConv Wizard-of-Internet
Metrics: Unigram F1 Recall-1 Recall-3 BLEU-1/2 ROUGE-1/2/L
Baselines: T5-base Self-training(scratch) Self-training(QP) Self-training(joint) QP-ext QP-gen KD (RA→QP) ChatGPT
Builds on: T5 REINFORCE self-training knowledge distillation
Closest papers by meaning
- From Retrieval to Generation: A Simple and Unified Generative Model for End-to-End Task-Oriented Dialogue (AAAI 2024 · similarity 0.85)
- Learning to Clarify: Multi-turn Conversations with Action-Based Contrastive Self-Training (ICLR 2025 · similarity 0.85)
- ChatQA: Surpassing GPT-4 on Conversational QA and RAG (NeurIPS 2024 · similarity 0.84)
- MaFeRw: Query Rewriting with Multi-Aspect Feedbacks for Retrieval-Augmented Large Language Models (AAAI 2025 · similarity 0.84)
- Improving Open-Domain Dialogue Response Generation with Multi-Source Multilingual Commonsense Knowledge (AAAI 2024 · similarity 0.84)
- RankRAG: Unifying Context Ranking with Retrieval-Augmented Generation in LLMs (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.