Source-linked AI summary
Affect-Driven Dialog Generation
Pierre Colombo, Wojciech Witon, Ashutosh Modi, James Kennedy, Mubbasir Kapadia
TL;DR
Current end-to-end dialogue systems generally lack explicit control over responses’ affective content. EMOTICONS uses emotion representations and affective training and inference methods to generate controlled emotional responses; the authors report improved BLEU and response diversity, while noting that different emotions are not generated equally well.
Problem
Most end-to-end dialogue systems focus on response quality without explicit control over affective content.
Source: §1 · p. 1Method
EMOTICONS uses a desired-emotion vector, an affective regularizer, affective sampling, and inference-time reranking to generate emotion-specific responses.
Source: §1 · p. 1Results
EMOTICONS improves over Zhou et al. (2018) and vanilla seq2seq by up to 7.7% in BLEU and 52% in response diversity.
Source: §1 · p. 1Takeaways & Limitations
A 22-person user study found the system could generate grammatically correct, coherent, emotionally rich text in a controlled manner.
Source: §1 · p. 1EMOTICONS does not generate different emotions equally well, and future work could add contextual information to better capture emotional content.
Source: §8 · p. 8
Abstract
from arXiv · showhide
The majority of current systems for end-to-end dialog generation focus on response quality without an explicit control over the affective content of the responses. In this paper, we present an affect-driven dialog system, which generates emotional responses in a controlled manner using a continuous representation of emotions. The system achieves this by modeling emotions at a word and sequence level using: (1) a vector representation of the desired emotion, (2) an affect regularizer, which penalizes neutral words, and (3) an affect sampling method, which forces the neural network to generate diverse words that are emotionally relevant. During inference, we use a reranking procedure that aims to extract the most emotionally relevant responses using a human-in-the-loop optimization process. We study the performance of our system in terms of both quantitative (BLEU score and response diversity), and qualitative (emotional appropriateness) measures.
1 Introduction
Most conversational systems emphasize functional response quality but lack social intelligence. EMOTICONS adds controlled emotional expression while aiming to preserve grammatical correctness, coherence, and relevance.
1 Introduction
- Most existing conversational systems focus on functional qualities such as understanding and pertinence, while lacking social intelligence.The introduction connects social intelligence to avoiding inappropriate responses and building rapport with users.Source: §1 · p. 1
- EMOTICONS generates emotion-specific responses using explicit controls and an affective regularizer without direct control over the desired emotion.The paper reports improvements over Zhou et al. (2018) and vanilla seq2seq of up to 7.7% in BLEU and 52% in response diversity.Source: §1 · p. 1
- A 22-person user study found EMOTICONS could generate grammatically correct, coherent, emotionally rich text in a controlled manner.Source: §1 · p. 1
2 Related Work
Prior dialogue-generation work emphasizes response quality and diversity, while emotion-focused systems use different ways to represent or control affect. EMOTICONS builds on these approaches and uses established inference techniques.
2 Related Work
- Sequence-to-sequence dialogue work has emphasized BLEU and perplexity, and this paper uses GRUs for their fewer parameters and faster training.Source: §2 · pp. 1–2
- Table 1 is captioned as showing example responses from the seq2seq baseline and four EMOTICONS models with different emotions.Source: Table 1 · p. 2
- Beam search has been used to improve answer quality, while MMI has improved diversity; EMOTICONS builds on these inference techniques for affective generation.Source: §2 · p. 2
- Prior affective systems use emotion tokens, external and internal memory, or Valence-Arousal representations, whereas EMOTICONS extends VA modeling to Valence-Arousal-Dominance.Asghar et al. (2017) use VAD but do not generate different emotional outputs for the same input or a specified emotion.Source: §2 · p. 2
3 System Architecture
The system comprises emotion labeling, affective training, and affective inference, which generates candidate responses and reranks them by emotional content.
3 System Architecture
- Emotion labeling assigns sentence-level emotion categories and word-level VAD values; affective training uses forward prompt-response and reversed response-prompt seq2seq networks.Source: §3 · p. 2
- Affective inference generates multiple plausible responses and reranks them based on emotional content.Source: §3 · p. 2
3.1 Preliminaries
The system represents emotion as a distribution over six categories and uses a target-emotion vector to guide response generation. During training that vector represents the target response; at test time it specifies the desired final-response emotion.
3.1 Preliminaries
- The sequence emotion vector E_X is a probability distribution over six emotions associated with sequence X.Source: §3.1 · p. 2
- E_0 represents the target response's emotion during training and the desired final-response emotion during testing, when it can be set manually.Source: §3.1 · p. 2
- The system feeds the input sequence and target emotion into seq2seq, then reranks candidates using reversed-seq2seq output, length, and emotional content.Source: Figure 1 · p. 3
3.2 Affect Modeling
Affect modeling adds emotion-specific information during training and inference, using six categorical emotions and continuous VAD word representations. The section develops sequence-level and word-level mechanisms to align generated responses with a target emotion while retaining grammaticality.
3.2 Affect Modeling
- The model uses six categorical emotions and VAD vectors for word-level emotion, with valence, arousal, and dominance each ranging from 0 to 1.The VAD lexicon maps 20,000 words to three-dimensional vectors.Source: §3.2 · p. 3
3.2.1 Emotion Classifier
- The Emotion Classifier supplies target-emotion labels for corpus sentences, and its Cornell predictions are reported as highly correlated with human predictions.The classifier predicts a distribution over six emotion classes.Source: §3.2.1 · p. 3
3.2.2 Sequence-Level Explicit Encoder Model (SEE)
- SEE prepends a learned embedding of the target emotion to the encoder input sequence.The embedding maps the target-emotion representation into a three-dimensional space.Source: §3.2.2 · p. 3
3.2.3 Sequence-Level Explicit Decoder Model (SED)
- SED supplies the target-emotion embedding at each decoding step, allowing the desired emotion to be specified in a continuous space.Its emotion embedding mapping differs from SEE's.Source: §3.2.3 · p. 3
3.2.4 Word-Level Implicit Model (WI)
- WI adds an affective regularizer that penalizes deviation between the generated response's VAD content and the target response's affect.Generated-response affect is represented as a confidence-weighted sum of vocabulary-word VAD vectors.Source: §3.2.4 · pp. 3–4
3.2.5 Word-Level Explicit Model (WE)
- WE uses adaptive affective sampling to choose words in light of the sequence's remaining emotional content, so later words can offset neutral openings.After the first half of the maximum response length, λ is set to 1 to carry target affect in early words while preserving grammatical correctness of the whole response.Source: §3.2.5 · p. 4
3.3 Affective Inference
Affective inference generates a diverse candidate list and reranks responses to select an emotion-specific output. Its scoring objective combines response likelihood, a reverse prompt likelihood, and a penalty for deviation from the desired affect.
3.3 Affective Inference
- The proposed inference strategy generates B diverse sequences with length-normalized beam search, then reranks them to seek an emotion-specific response.The approach builds on N-best reranking and MMI during inference.Source: §3.3 · pp. 4–5
- The reverse probability p(S|RC) is computed by a model trained on reversed prompt-response pairs.The reverse model uses the same architecture as the response-given-prompt model.Source: §3.3 · p. 5
- The modified MMI objective penalizes candidate responses whose emotional content deviates from the desired emotion.The authors tune α, β, and γ for grammatical correctness, diversity, and emotional content.Source: §3.3 · p. 5
4 Model Training
The experiments use Cornell and filtered OpenSubtitles2018 dialogue data, an affective dictionary, and vanilla seq2seq and ECM baselines. Models are tuned and evaluated using specified preprocessing, optimization, and data-split procedures.
4.1 Corpora
- OpenSubtitles2018 is filtered to 2.5 million utterances from consecutive 5–30-word sequences within five seconds and conversations of at least four turns.Cornell contains around 10,000 movie characters and 220,000 dialogues.Source: §4.1 · p. 5
4.2 Affective Dictionary
- Vocabulary words without a sufficiently similar VAD-lexicon entry receive the neutral vector [0.5, 0.5, 0.5].The similarity threshold is greater than 90%.Source: §4.2 · p. 5
4.3 Baselines
- The comparison baselines are vanilla seq2seq and Zhou et al.'s ECM, with the affective dictionary used for ECM's external memory.ECM is trained using the authors' default parameters.Source: §4.3 · p. 5
4.4 Training Details
- Hyperparameters are optimized on validation BLEU; models use 94%/1%/5% train/validation/test splits, with maximum sequence lengths of 20 for Cornell and 30 for OpenSubtitles.The encoder uses two-layer bidirectional GRUs and the decoder a one-layer unidirectional GRU.Source: §4.4 · p. 5
4 Model Training
- Table 2 reports proposed-model results on Cornell and OpenSubtitles, including distinct-1 and distinct-2 normalized over generated tokens and performance boost over vanilla seq2seq.The measures count distinct unigrams and bigrams across 200 candidate responses.Source: Table 2 · p. 6
5 Quantitative Evaluation for Model Selection
The evaluation compares emotional models on response diversity and BLEU, finding that emotional features can increase diversity without generally reducing BLEU. Word-level emotion models perform particularly well on diversity, while the results suggest they use emotional patterns in generation.
5.1 Response Diversity
- WI and WE achieve the best distinct-1 and distinct-2 scores among models without reranking, with improvements of up to 13.1%.The authors attribute their diversity to training that promotes less frequent words close to target words in VAD values.Source: §5.1 · p. 6
- Across Cornell and OpenSubtitles, SED, WI, and WE each outperform vanilla seq2seq and ECM on at least one distinct measure.The authors state that observed improvements after adding emotions suggest the systems extract and use emotional patterns to improve the general language model.Source: §5.1 · p. 6, §5 · p. 6
- EMOTICONS, combining WI + WE for response scoring and WI for reverse scoring, beats MMIbas. by up to 52.5% in distinct-1.The authors note that the OpenSubtitles MMI diversity gain is smaller than Li et al. (2016) reported, possibly due to data filtering, beam search, or Cornell-only hyperparameter tuning.Source: §5.1 · p. 7
5.2 Response Quality
- Introducing emotional features generally does not reduce BLEU, and WI and SED sometimes improve it.The authors compute BLEU on beam-search results to reduce the potential impact of inappropriate first words.Source: §5.2 · p. 7
6 Human-in-the-Loop Hyper-Parameter Estimation
The authors use human judgments of generated responses to tune the affective reranking parameter γ, since standard metrics do not directly measure affect. The average distance between judged and desired emotion is lowest at γ = 4.2, though results differ by emotion.
6 Human-in-the-Loop Hyper-Parameter Estimation
- The authors tune γ using human judgments because BLEU and diversity metrics do not explicitly measure affective content.They optimize γ after an earlier evaluation with γ = 0.Source: §6 · p. 7
6.1 Experiment Description
- Twelve annotators rated 400 Cornell sequences using AffectButton; prompts came from the 200 test sequences with the most diverse MMIbas. responses.Generated responses targeted fear, anger, joy, or surprise, and γ was sampled uniformly from [0, 10].Source: §6.1 · p. 7
6.2 Experiment Results
- Average judged-to-target VAD distance reaches its minimum at γopt = 4.2, with lower distance for joy than anger.The authors connect the emotion difference to the classifier's detection accuracy: joy 77%, anger 57%, surprise 62%, and fear 69%.Source: §6.2 · pp. 7–8, Figure 2 · p. 7
7 Qualitative Evaluation
The user study compares EMOTICONS with MMIbas. on grammatical correctness, preference, and emotional appropriateness. Users strongly prefer EMOTICONS, and most user-assigned faces portray the desired emotion accurately, although surprise is a weaker case.
7 Qualitative Evaluation
- The study tests EMOTICONS responses for four target emotions against MMIbas. responses using Cornell prompts selected for MMIbas. diversity.The experiments cover grammatical correctness, user preference, and emotional appropriateness.Source: §7 · p. 8
7.1 Grammatical Correctness
- EMOTICONS does not substantially sacrifice grammatical correctness in the user rating study.Three native English speakers rated responses for 40 prompts; Fleiss’ κ was 0.4128, corresponding to moderate agreement.Source: §7.1 · p. 8
7.2 User Preference
- Users strongly prefer EMOTICONS over MMIbas. in the response-choice study.Eighteen annotators chose among eight answers, with each of 45 sentences rated by three annotators.Source: §7.2 · p. 8
7.3 Emotional Appropriateness
- Most user-assigned faces portray the desired emotion accurately, but surprise is expressed poorly, partly because users mismatch it with joy.The authors also note that surprise is one of the most difficult emotions to judge.Source: §7.3 · p. 8
8 Conclusion
EMOTICONS generates responses with controlled emotions and can be used with neural architectures that fit the encoder-decoder framework. It does not generate different emotions equally well, and contextual information is suggested as future work to better capture emotional content.
8 Conclusion
- EMOTICONS generates responses with controlled emotions.Source: §8 · p. 8
- The solution can be used with any neural architecture that fits the encoder-decoder framework.Source: §8 · p. 8
- EMOTICONS does not generate different emotions equally well; future work could incorporate contextual information to better capture emotional content.Source: §8 · p. 8