Source-linked AI summary
Affective Neural Response Generation
Nabiha Asghar, Pascal Poupart, Jesse Hoey, Xin Jiang, Lili Mou
TL;DR
Existing neural conversation models largely capture lexico-syntactic information while omitting affective content. The paper adds cognitively engineered affective embeddings, affect-based training objectives, and affectively diverse decoding to LSTM encoder-decoder systems, and reports improved response quality, especially emotional appropriateness. Direct comparison with Emotional Chatting Machine is infeasible because its required emotion-category input and experimental setting differ.
Problem
Existing neural conversation models capture syntactic alignment and low-level word semantics but do not explicitly model the emotional aspects of dialogue.
Method
The paper introduces affective word embeddings, affect-based objective functions augmenting cross-entropy, and affectively diverse beam-search decoding for neural dialogue generation.
Results
The three affective strategies can be combined, with performance generally increasing as components are added and significantly exceeding original Seq2Seq performance, especially for emotional appropriateness.
Takeaways & Limitations
The combined strategies improve the quality of generated open-domain responses and advance affectively cognizant neural encoder-decoder dialogue systems.
Takeaways & Limitations
Direct comparison with Emotional Chatting Machine is infeasible because ECM requires a desired affect category as input and uses a different experimental setting.
Abstract
from arXiv · showhide
Existing neural conversational models process natural language primarily on a lexico-syntactic level, thereby ignoring one of the most crucial components of human-to-human dialogue: its affective content. We take a step in this direction by proposing three novel ways to incorporate affective/emotional aspects into long short term memory (LSTM) encoder-decoder neural conversation models: (1) affective word embeddings, which are cognitively engineered, (2) affect-based objective functions that augment the standard cross-entropy loss, and (3) affectively diverse beam search for decoding. Experiments show that these techniques improve the open-domain conversational prowess of encoder-decoder networks by enabling them to produce emotionally rich responses that are more interesting and natural.
Introduction
Neural dialogue models produce well-formed responses but often remain off-context, dull, vague, and emotionally insensitive. This paper augments open-domain Seq2Seq models with three affective strategies and reports improved response quality, especially emotional appropriateness.
- Existing neural conversation models can generate syntactically well-formed responses but remain prone to off-context, short, dull, or vague outputs.
- These models capture syntactic alignment and low-level word semantics but do not explicitly model emotional aspects of language.
- The proposed approach augments neural dialogue models with affective intelligence through three strategies.
- It uses affective word embeddings, affect-based objectives added to cross-entropy loss, and affectively diverse beam search for decoding.
- The emotional strategies can be combined to further improve the quality of generated responses in open-domain dialogue.
Related Work
Affective neural dialogue research builds on earlier efforts to incorporate emotion into dialogue systems. The paper situates its approach alongside Affect-LM and Emotional Chatting Machine while distinguishing its broader dialogue setting and affective mechanisms.
- Earlier emotional dialogue systems commonly used hand-crafted speech or text features in retrieval-based or slot-based spoken dialogue systems.
- Affect-LM is an LSTM-RNN language model that extracts binary affective features with LIWC and conditions generation on affect features and strength.
- Emotional Chatting Machine is a Seq2Seq model that takes a prompt and desired emotion as input and generates a response across eight emotion categories.
Background
Word embeddings represent tokens as fixed-dimensional real-valued vectors, while Seq2Seq models encode variable-length inputs and decode outputs. The paper motivates affective extensions because ordinary co-occurrence statistics do not explicitly represent emotional features.
- Word Embeddings: Word embeddings map words or tokens to real-valued vectors of fixed dimensionality and support accuracy across NLP tasks.
- Word Embeddings: Traditional embeddings learn from word co-occurrence, placing words with similar syntactic and semantic contexts near one another.
- Affective Extensions: The approach uses a cognitively engineered dictionary with Valence, Arousal, and Dominance scores to guide affective strategies for Seq2Seq input, training, and inference.
- Affective Extensions: Co-occurrence statistics can miss sentiment because words with opposing sentiments may share context, motivating explicit affective features.
- Seq2Seq Model: Seq2Seq is an encoder-decoder framework that maps variable-length input sequences to variable-length output sequences using recurrent networks, typically LSTMs.
- Seq2Seq Model: The encoder sequentially processes input-word embeddings into a vector, and the decoder uses that vector to generate an output sequence.
- Seq2Seq Model: Seq2Seq models are typically trained with cross-entropy loss and generate responses by maximizing log p(Y|X), using greedy or beam-search variants.
The Proposed Affective Approaches
The paper augments Seq2Seq dialogue generation with affective word representations, affect-aware training objectives, and affectively diverse decoding. These strategies use a 3D VAD space, affective losses, and beam-group dissimilarity to incorporate emotional content into responses.
- Overview: The approach combines affective word embeddings, affective training objectives, and affectively diverse beam search for Seq2Seq dialogue generation.The strategies are designed to add emotional cognizance to traditional neural conversation models.
- Affective Word Embeddings: The affective dictionary rates 13,915 lemmatized English words on Valence, Arousal, and Dominance dimensions.Valence measures pleasantness, Arousal emotional intensity, and Dominance power or control; the ratings form the word-level affective embeddings.
- Affective Word Embeddings: Traditional and affective embeddings are concatenated and supplied as input to both the encoder and decoder.Words with similar emotional connotations are close in the affective space, while affectively dissimilar words are farther apart.
- Affective Loss Functions: Affective training objectives penalize affective distance between source and generated responses or between generated words and a neutral vector, encouraging affective consistency or emotionally rich words.The objectives include minimizing affective dissonance, maximizing affective dissonance, and maximizing affective content.
- Affectively Diverse Decoding: Beam search retains the top-B likely subsequences at each step, while the proposed decoding variants use affective dissimilarity to diversify the retained samples.Standard beam search can produce nearly identical responses, motivating affect-sensitive diversity across beam groups.
- Affectively Diverse Decoding: Affectively diverse beam search injects affective dissimilarity across beam groups at either the word or sentence level during decoding.The sentence-level metric addresses the limitation that word-level diversity does not capture each group’s overall sentence affect.
Experiments
Experiments evaluate affective word embeddings, affective loss functions, and affectively diverse decoding individually and in combination. Across human judgments and diversity evaluation, the strategies improve emotional appropriateness and other response qualities, with combined components performing significantly better than the original Seq2Seq model.
- Evaluation: Human judges rated responses on syntactic coherence, naturalness, and emotional appropriateness using scores from 0 to 2.Five workers evaluated 100 test samples per model variant; inter-annotator agreement was κ = 0.447, interpreted as moderate agreement.
- Affective word embeddings: Affective word embeddings consistently improve syntactic coherence, naturalness, and emotional appropriateness, with the largest effect on emotional appropriateness.The results indicate that traditional end-to-end word embeddings do not sufficiently capture emotional aspects, whereas additional affective knowledge improves affect awareness.
- Affective loss functions: Affective loss functions DMIN and DMAX both outperform XENT, while AC generally performs better than them on naturalness and emotional appropriateness.AC encourages affective content without specifying affect direction; DMIN and DMAX impose minimum or maximum affective dissonance relationships.
- Affectively diverse decoding: WL-ADBS and SL-ADBS outperform original beam search and Hamming-based DBS on affective diversity and the number of emotionally appropriate responses.SL-ADBS is slightly better than WL-ADBS because it considers cumulative sentence-level affect rather than individual-word affect.
- Combined strategies: Combining affective embeddings, LAC, and SL-ADBS generally increases performance as components are added and yields significantly better results than the original Seq2Seq model, especially for emotional appropriateness.The combined strategies are described as directly combinable and complementary across experiments.
- Case study: The case study shows generic, non-committal baseline responses, while affective models produce responses with stronger or more subtle emotional connotations.LAC generates highly affective words such as “suicide” and “crazy,” whereas affective embeddings produce subtler affective responses.
Conclusion and Future Work
The work advances affectively cognizant neural encoder-decoder dialogue systems through three affective strategies. Future work will investigate affect-based attention and personalization using reinforcement learning.
- The paper addresses affective neural dialogue generation for applications such as emotional conversation partners.
- It advances affectively cognizant encoder-decoder systems through affective-space embeddings, affect-based objectives, and affectively diverse decoding.
- Future work will investigate affect-based attention mechanisms for neural conversational models.
- Future work will explore affect-based personalization using reinforcement learning.