Source-linked AI summary
A Controllable Model of Grounded Response Generation
Zeqiu Wu, Michel Galley, Chris Brockett, Yizhe Zhang, Xiang Gao, Chris Quirk, Rik Koncel-Kedziorski, Jianfeng Gao, Hannaneh Hajishirzi, Mari Ostendorf, Bill Dolan
TL;DR
Neural conversation models can produce fluent, human-like responses but may hallucinate facts, while grounding alone does not guarantee relevant, specific responses. CGRG combines grounding and lexical control, with control phrases supplied by users or predicted automatically. The authors report that it outperforms strong generation baselines in both control settings.
Problem
Neural response generation can hallucinate facts, while grounding alone may leave responses vague or irrelevant without semantic targeting.
Source: §1 · p. 1Method
Results
CGRG outperforms strong baselines when control phrases are user-provided or automatically extracted.
Source: §1 · p. 2Takeaways & Limitations
The framework combines grounding for information reliability with lexical control for specificity in response generation.
Source: §7 · p. 7Combining control during training and decoding may be complementary, but the paper leaves investigation of such combinations to future research.
Source: §5.1 · p. 5
Abstract
from arXiv · showhide
Current end-to-end neural conversation models inherently lack the flexibility to impose semantic control in the response generation process, often resulting in uninteresting responses. Attempts to boost informativeness alone come at the expense of factual accuracy, as attested by pretrained language models' propensity to "hallucinate" facts. While this may be mitigated by access to background knowledge, there is scant guarantee of relevance and informativeness in generated responses. We propose a framework that we call controllable grounded response generation (CGRG), in which lexical control phrases are either provided by a user or automatically extracted by a control phrase predictor from dialogue context and grounding knowledge. Quantitative and qualitative results show that, using this framework, a transformer based model with a novel inductive attention mechanism, trained on a conversation-like Reddit dataset, outperforms strong generation baselines.
1 Introduction
Modern neural dialogue models can produce human-like responses but may hallucinate facts; grounding can improve reliability while leaving responses vague. The paper proposes combining grounding with lexical control to generate more specific, reliable responses and reports that CGRG outperforms strong baselines.
1 Introduction
- Recent diversity-enhancing strategies and large GPT-2-style models reduce blandness, but generated responses can still hallucinate or contain fake output.The paper illustrates this problem with an ungrounded response in Figure 1.Source: §1 · p. 1
- Figure 1 contrasts ungrounded, uncontrolled responses with responses using grounding, control, or both: grounding improves reliability, control boosts specificity, and their combination is contentful and reliable.The combined condition is contrasted with grounding alone, which may lead to vague responses.Source: Figure 1 · p. 1
- Grounding can inhibit factual hallucinations, yet without semantic targeting it may produce accurate but vague or irrelevant responses.The paper argues that lexical controls enforce specificity and help filter lengthy, irrelevant, or incoherent grounding text.Source: §1 · p. 1
- Lexical control may come from a user indicating semantic intent or be predicted automatically in a fully automated system.Figure 2 depicts the editorial-assistant scenario, where the machine suggests responses using conversation history, partial user input, and grounding knowledge.Source: §1 · p. 1, Figure 2 · p. 2
- CGRG outperforms strong baselines with both user-provided control phrases and automatically extracted phrases.The framework generates responses from dialogue context, lexical control phrases, and grounding knowledge.Source: §1 · p. 2
2 Approach
CGRG generates responses from dialogue context, lexical control phrases, and relevant grounding; inductive attention limits input attention to links relevant to those controls. Control phrases can be user-provided or predicted automatically.
2 Approach
- CGRG generates a response guided by dialogue context, lexical control phrases, and grounding sentences relevant to those phrases.The control phrases may be provided by a user or automatically derived by a predictor.Source: §2 · p. 2
2.1 Inductive Attention
- Inductive attention masks potentially uninformative links using the structure between control phrases and relevant grounding, while response tokens retain access to all input tokens and earlier response tokens.For example, a control phrase found only in one grounding sentence attends only to that sentence; cross-sentence grounding links and links between different control phrases are removed.Source: §2.1 · p. 2, Figure 3 · p. 3
2.2 Control Phrase Selection
- User control phrases are informative n-grams of up to five words shared by the grounding document and reference response, excluding phrases already in the dialogue context.When matched phrases differ only by a function word or punctuation, the shorter form is retained.Source: §2.2 · p. 3
- The retrieval-based predictor selects two frequent n-grams from the 50 grounding sentences ranked highest by IDF-weighted overlap with the dialogue context, restricting candidates to noun phrases.It drops the second phrase if it fully overlaps with the first.Source: §2.2 · p. 3
- The second automatic predictor fine-tunes BERT question answering with dialogue context as query, grounding as document, and control phrases as answers, then selects its top two predicted answers.It also drops the second phrase when its string fully overlaps with the first.Source: §2.2 · p. 3
3 Controllable Conversation Dataset
The dataset uses Reddit conversations associated with web pages across a broad range of topics, filtered to emphasize responses that use grounding. Crowd workers also assessed whether simulated user-control phrases were central to their reference responses.
3 Controllable Conversation Dataset
- The grounded Reddit dataset contains conversations about web pages, including news stories and Wikipedia articles, across 178 subreddit topics and multiple writing styles.Topics range from news and technology to literature and music.Source: §3 · p. 3
- After strict lexical matching kept only responses with at least one phrase found in the grounding document, the dataset contained 390K train, 6.7K development, and 21K test utterances.Average reference-response length was 26.5; mean phrase counts in C were 1.32, 1.27, and 1.38 for train, development, and test, respectively.Source: §3 · p. 3
- In 2,000 annotated examples, simulated control phrases had a median centrality score of 4.33, and 67.4% scored above 4 on the six-point scale.Three judges rated each example; inter-rater agreement was fair (Krippendorff’s alpha 0.32).Source: §3 · p. 3
4 Experimental Setup
Experiments fine-tuned GPT-2-based systems on a controllable, grounded Reddit dataset and compared settings that vary context, grounding, and control. Automatic metrics supported tuning and selection, while human judgments assessed relevance and consistency with background text.
4.1 Training and Inference Setup
- GPT-2 small with 117M parameters served as the base model, and GPT-2-based baselines were fine-tuned on the same Reddit dataset using greedy decoding except for a grid beam search comparison.Input or target sequences were capped at 512 tokens; learning rate and warmup were tuned on development perplexity.Source: §4.1 · pp. 3–4
4.2 Evaluated Systems
- Evaluated settings varied whether systems received dialogue context, grounding, control phrases, or grounding relevant to control, including control applied only at decoding with grid beam search.The compared systems included GPT-2, GPT2IA, and CMR; GPT2IA adds inductive attention.Source: §4.2 · p. 4
4.3 Automatic Evaluation
- Automatic evaluation measured relevance with BLEU-4 and NIST-4 and response diversity with Div-2, using the highest score across up to five references for each metric.Because automatic generation metrics can be unreliable, human evaluation was used to verify improvements among systems selected by automatic metrics.Source: §4.3 · p. 4
4.4 Human Evaluation
- Crowd workers compared randomized response pairs for relevance and appropriateness to dialogue and consistency with background text, allowing ties on five-point Likert scales.Three to four judges evaluated each pair; inter-rater agreement was fair (Krippendorff’s coefficient 0.32).Source: §4.4 · pp. 4–5, Table 2 · p. 5
5 Results and Analysis
Human evaluations favor the grounded, controlled GPT2IA model, and predicted controls also improve fully automatic generation. Qualitative analyses indicate that grounding and control together help prioritize correct, intended entities.
5.1 User-controlled Response Generation
- Comparisons show gains from adding control, constraint-sensitive grounding, explicit control phrases, hidden-state control, and inductive attention.GPT-2 performs better than the LSTM-based CMR model; hidden-state control outperforms decoding-only control. The authors note that combining training- and decoding-stage controls remains for future work.Source: §5.1 · p. 5
- X+C+GC+GPT2IA outperforms other systems in human evaluation, except that its Consistency score is not statistically different from X+C+GC+GPT2.For single-reference evaluation, the authors note low BLEU/NIST scores consistent with differences between human references.Source: §5.1 · p. 5
5.2 Predictor-controlled Response Generation
- 32% of paired judgments favored fully automatic X+ ˜C+G ˜ C +GPT2IA as most informative and relevant, versus 20-21% for X+GPT-2 (DialoGPT).Retrieval- and BERT-QA-based control predictors also outperform X+GPT-2 (DialoGPT) and achieve good Div-2 results, though BLEU is low.Source: §5.2 · p. 5
5.3 Qualitative Analysis
- Compared with X+C, the grounded model assigns higher probabilities to contextual words such as “graduated” and “thesis” and correct entity token “2017,” and lower probability to incorrect “economics.”Source: §5.3 · p. 6
- With the same context and grounding, adding grounding ranks locations using background knowledge, while adding the control phrase “Canada” helps locate the intended answer.Source: §5.3 · p. 6
- 0.773 (X+C/X+C+GC) and 0.886 (X+G/X+C+GC): both ratios are below 1.0, indicating higher probability for correct entities with both grounding and control than with either alone.Source: §5.3 · p. 6
6 Related Work
Prior work explores grounding and controllable generation, but the cited approaches often treat them separately or make narrower assumptions about domains, grounding, or dialogue context.
6.1 Grounded Response Generation
- Earlier grounded-response work includes domain-specific datasets and single-sentence grounding, while this work targets scalable responses blending multiple grounding sources.Other approaches copy from context and grounding or co-encode them with a reading-comprehension model.Source: §6.1 · p. 6
- This work uses text-based grounding without requiring preconstructed knowledge graphs, unlike cited relational-knowledge and graph-path approaches.Source: §6.1 · p. 7
6.2 Controlled and Content-Planned Generation
- Prior lexical-constraint methods apply constraints at inference only, whereas this work applies them during training and optionally at inference.The paper says training-time application incorporates constraints into the latent space for better predictions.Source: §6.2 · p. 7
- Related non-lexical control work does not examine grounding and often assumes gold constraints; this paper also evaluates predicted constraints.Style-control methods use labels such as positive/negative or formal/informal, while content-planned generation does not account for dialogue context.Source: §6.2 · p. 7
7 Conclusion
CGRG combines soft semantic control with grounding, and introduces inductive attention and a control phrase predictor for response generation. The paper also identifies broader control types and document-writing assistance as future directions.
7 Conclusion
- CGRG uses grounding to contextualize users’ semantic intents and boost information reliability.It allows users to inject soft semantic control into generation.Source: §7 · p. 7
- An inductive attention mechanism is introduced for self-attention-based generation models like GPT-2 to boost performance.Source: §7 · p. 7
- Integrating a control phrase predictor lets CGRG benefit standard automatic response generation.Source: §7 · p. 7
- Future directions include exploring different user-desired controls and extending controllable grounded generation to broader tasks such as document-writing assistance.Source: §7 · p. 7