Research map / Large language models / Retrieval and knowledge
Improving the Robustness of Knowledge-Grounded Dialogue via Contrastive Learning
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 EnCo, an entity-guided contrastive learning framework that builds paraphrased positives and entity-perturbed negatives to make knowledge-grounded dialogue models robust to real-world noise.
Problem
Knowledge-grounded dialogue systems are hurt by real-world noise such as misspellings, abbreviations, incomplete, erroneous or outdated knowledge graph facts, and prior KGD work neglects robustness.
Approach
Entity-guided paraphrasing (entity boundary tokens keep entities unchanged) creates semantic-irrelevant positives, while entity-guided negative augmentation (delete/replace entities in both context and KG) creates semantic-relevant negatives; a contrastive loss on encoder representations plus cross-entropy on positives trains the model.
Key results
- On DuConv, EnCo reaches BLEU-1 40.10 and BLEU-2 29.69 vs. 39.06/28.64 for the second-best KRP-DS.
- On KdConv (travel), EnCo BLEU-1 46.61 vs. 45.00 for KRP-DS.
- In the robustness study on DuConv with noisy contexts, EnCo keeps the best BLEU-1/2 across all noise types (e.g., 34.07 BLEU-1 under ChatGPT-generated noise vs. 31.58 for BART+Know).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add entity-based contrastive learning and positive-sample augmentation | BART-large KGD model · KdConv (music) | BLEU-1: 36.88 -> 39.39 | Helped |
| add entity-based contrastive learning and positive-sample augmentation | BART-large KGD model · KdConv (travel) | BLEU-1: 45.00 -> 46.61 | Helped |
| add entity-based contrastive learning and positive-sample augmentation | BART-large KGD model · KdConv (film) | BLEU-1: 30.14 -> 31.43 | Helped |
| add entity-based contrastive learning and positive-sample augmentation | BART-large KGD model · DuConv | BLEU-1: 39.06 -> 40.10 | Helped |
| add entity-based contrastive learning and positive-sample augmentation | BART-large KGD model · DuRecDial | BLEU-1: 39.02 -> 40.05 | Helped |
| add positive-sample cross-entropy loss (ablation) | EnCo · KdConv (music) | BLEU-1: 36.59 -> 39.39 | Helped |
| add positive-sample cross-entropy loss (ablation) | EnCo · KdConv (travel) | BLEU-1: 43.29 -> 46.61 | Helped |
| add contrastive loss (ablation) | EnCo · KdConv (music) | BLEU-1: 36.02 -> 39.39 | Helped |
| add contrastive loss (ablation) | EnCo · KdConv (travel) | BLEU-1: 42.80 -> 46.61 | Helped |
| train with entity-guided contrastive learning and augmentation | EnCo · DuConv with noisy inputs | BLEU-1 | Helped |
Surprising findings
- PPL generally increases when knowledge is incorporated into KGD models (e.g., HRED 16.82 to HRED+Know 17.69 on KdConv music).
- BART alone achieves the lowest PPL on several datasets, even though EnCo wins on BLEU and DISTINCT.
Limitations
- Robustness is evaluated only with synthetic noise on DuConv test data, not on real noisy user inputs.
- Negative and positive samples depend on an off-the-shelf NER toolkit (TexSmart), so noise in NER propagates into the augmentations.
- Hyperparameter tuning was minimal, with a fixed 30% perturbation rate and 5 positives/negatives per sample.
Experimental setup
Datasets: KdConv DuConv DuRecDial ParaZh-22M
Benchmarks: KdConv DuConv DuRecDial
Metrics: PPL BLEU-1/2/3/4 DISTINCT-1/2/3/4 human fluency/coherence/informativeness
Baselines: Seq2Seq HRED Seq2Seq+Know. HRED+Know. KIC SDAN MGCG_G KSPN BART BART+Know KRP-DS
Builds on: BART TransR contrastive learning ParaZh-22M
Closest papers by meaning
- Zero-Resource Knowledge-Grounded Dialogue Generation (NeurIPS 2020 · similarity 0.89)
- Sequential Latent Knowledge Selection for Knowledge-Grounded Dialogue (ICLR 2020 · similarity 0.88)
- KPT: Keyword-Guided Pre-training for Grounded Dialog Generation (AAAI 2023 · similarity 0.88)
- Diverse and Faithful Knowledge-Grounded Dialogue Generation via Sequential Posterior Inference (ICML 2023 · similarity 0.88)
- Improving Open-Domain Dialogue Response Generation with Multi-Source Multilingual Commonsense Knowledge (AAAI 2024 · similarity 0.85)
- Learning to Clarify: Multi-turn Conversations with Action-Based Contrastive Self-Training (ICLR 2025 · 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.