Affective Neural Response Generation

Affective Neural Response Generation

Nabiha Asghar, Pascal Poupart, Jesse Hoey, Xin Jiang, Lili Mou
Cheriton School of Computer Science, University of Waterloo, Canada
Noah’s Ark Lab, Huawei Technologies, Hong Kong

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.

Affective Neural Response Generation

Nabiha Asghar, Pascal Poupart, Jesse Hoey, Xin Jiang, Lili Mou Cheriton School of Computer Science, University of Waterloo, Canada {nasghar,ppoupart,jhoey}, Noah’s Ark Lab, Huawei Technologies, Hong Kong


Human-computer dialogue systems have wide applications ranging from restaurant booking (?) to emotional virtual agents (?). Inspired by the recent success of deep neural networks in natural language processing (NLP) tasks such as language modeling (?), machine translation (?), and text summarization (?), the artificial intelligence (AI) community is aggressively exploring the paradigm of neural dialogue generation.

In a neural network-based dialogue system, discrete words are mapped to real-valued vectors, known as embeddings, capturing abstract meanings of words (?); then an encoder-decoder framework—with long short term memory (LSTM)-based recurrent neural networks (RNNs)—generates a response conditioned on one or several previous utterances. Latest advances in this direction have demonstrated its efficacy for both task-oriented dialogue systems (?) and open-domain response generation (???).

While most of the existing neural conversation models generate syntactically well-formed responses, they are prone to being off-context, short, dull, or vague. Latest efforts to address these issues include diverse decoding (??), diversity-promoting objective functions (?), adversarial learning (?), latent variable modeling for diversity (?), human-in-the-loop reinforcement learning (?), online active learning (?), latent intention modeling (?), and content-introducing approaches (??). These advancements are promising, but we are still far from our goal of building autonomous neural agents that can consistently carry out interesting human-like conversations.

One shortcoming of these existing open-domain neural conversation models is the lack of affect modeling of natural language. These models, when trained over large dialogue datasets, do not capture the emotional states of the two humans interacting in the textual conversation, which are typically manifested through the choice of words, phrases, or emotions. For instance, the attention mechanism in a sequence-to-sequence (Seq2Seq) model can learn syntactic alignment of words within the generated sequences  (?). Similarly, neural word embedding models like Word2Vec learn word vectors by context, and can preserve low-level word semantics (e.g., “king”“male”“queen”“woman”). However, emotional aspects are not explicitly captured by existing methods.

Our goal is to alleviate this issue in open-domain neural dialogue models by augmenting them with affective intelligence. We do this in three ways.

  1. We embed words in a 3D affective space by using a cognitively engineered word-level affective dictionary (?), where affectively similar constructs are close to one other. In this way, the ensuing neural model is aware of words’ emotional features.

  2. We propose to augment the standard cross-entropy loss with affective objectives, so that our neural models are explicitly taught to generate more emotional utterances.

  3. We inject affective diversity into the responses generated by the decoder through affectively diverse beam search algorithms, and thus our model actively searches for affective responses during decoding.

We also show that these emotional aspects can be combined to further improve the quality of generated responses in an open-domain dialogue system.

Related Work

Affectively cognizant virtual agents are generating interest both in the academia (?) and the industry,111 due to their ability to provide emotional companionship to humans. Endowing text-based dialogue generation systems with emotions is also an active area of research. Past research has mostly focused on developing hand-crafted speech and text-based features to incorporate emotions in retrieval-based or slot-based spoken dialogue systems (??).

Despite these, our work is mostly related to two very recent studies:

  • Affect Language Model (?, Affect-LM) is an LSTM-RNN language model which leverages the Linguistic Inquiry and Word Count (?, LIWC) text analysis program for affective feature extraction through keyword spotting. Affect-LM considers binary affective features, namely positive emotion, angry, sad, anxious, and negative emotion; at prediction time, Affect-LM generates sentences conditioned on the input affect features and a learned parameter of affect strength.

    Our work differs from Affect-LM in that we consider affective dialogue systems instead of merely language models, and we have explored more affective aspects including training and decoding.

  • Emotional Chatting Machine (?, ECM) is a Seq2Seq model. It takes as input a prompt and the desired emotion of the response, and then produces a response. It has eight emotion categories, namely anger, disgust, fear, happiness, like, sadness, surprise, and other. Additionally, ECM contains an internal memory and an external memory. The internal memory models the change of the internal emotion state of the decoder, and therefore encodes how much an emotion has already been expressed. The external memory decides whether to choose an emotional or generic (non-emotional) word at a given step during decoding.

    Our approach does not require the input of desired emotion as in ? (?), which is unrealistic in applications. Instead, we intrinsically model emotion by affective word embeddings as input, as well as objective functions and inference criterion based on these affective embeddings.


Word Embeddings

In NLP, word embeddings map words (or tokens) to real-valued vectors of fixed dimensionality. In recent years, neural network-based embedding learning has gained tremendous popularity, e.g., Word2Vec (?). They have been shown to boost the accuracy of computational models on various NLP tasks.

Typically, word embeddings are learned from the co-occurrence statistics of words in large natural language corpora, and the learned embedding vector space has such a property that words sharing similar syntactic and semantic context are close to each other. However, it is known that co-occurrence statistics are insufficient to capture sentiment/emotional features, because words different in sentiment often share context (e.g., “a good book” vs. “a bad book”). Therefore, we enhance traditional embeddings with a cognitively engineered dictionary, explicitly representing word affective features from several perspectives.

Figure 1: Overview of our approach, which is built upon a traditional Seq2Seq model. We propose three affective strategies for the input, training, and inference of Seq2Seq based on a cognitively engineered dictionary with Valence, Arousal, and Dominance (VAD) scores.

Seq2Seq Model

A sequence-to-sequence (Seq2Seq) model is an encoder-decoder neural framework that maps a variable length input sequence to a variable length output sequence (?). It consists of an encoder and a decoder, both of which are RNNs (typically with LSTM units). In the context of NLP, the encoder network sequentially accepts the embedding of each word in the input sequence, and encodes the input sentence as a vector. The decoder network takes the vector as input and sequentially generates an output sequence.

Given a message-response pair , where and are sequences of words/tokens, Seq2Seq models are typically trained with cross entropy loss (XENT):


where denotes model parameters. At prediction time, the model generates a response to a prompt by computing


either greedily or by variants of beam search.














Valence (V)

Arousal (A)

(a) V-A plot.














Valence (V)

Dominance (D)

(b) V-D plot.














Dominance (D)

Arousal (A)

(c) D-A plot.
Figure 2: The relationship between several examples of adjectives, nouns, and verbs on the 3-dimensional VAD scale.

The Proposed Affective Approaches

In this section, we propose affective neural dialogue generation, which augments traditional neural conversation models with emotional cognizance.

Figure 1 delineates an overall picture of our approach. We leverage a cognitively engineered dictionary, based on which we propose three strategies for affective dialogue generation, namely affective word embeddings as input, affective training objectives, and affectively diverse beam search. As will be shown in the experiments, these affective strategies can be combined to further improve Seq2Seq dialogue systems.

Affective Word Embeddings

As said, traditional word embeddings trained with co-occurrence statistics are insufficient to capture affect aspects. We propose to augment traditional word embeddings with a 3D affective space by using an external cognitively-engineered affective dictionary (?).222Available at

The dictionary we use consists of 13,915 lemmatized English words, each of which is rated on three traditionally accepted continuous and real-valued dimensions of emotion: Valence (V, the pleasantness of a stimulus), Arousal (A, the intensity of emotion produced, or the degree of arousal evoked, by a stimulus), and Dominance (D, the degree of power/control exerted by a stimulus). Sociologists hypothesize that the VAD space (also known as the EPA333EPA refers to the affective dimensions Evaluation, Potency, and Activity, which are congruent to Valence, Arousal, and Dominance, respectively. space) structures the semantic relations of linguistic concepts across languages and cultures; it captures almost 70% of the variance in affective meanings of concepts (?). VAD ratings have been previously used in sentiment analysis and empathetic tutors, among other affective computing applications (??). To the best of our knowledge, we are the first to introduce VAD to dialogue generation.

The scale of each dimension in the VAD space is from 1 to 9, where a higher value corresponds to higher valence, arousal, or dominance. Thus, V and corresponds to a word being very negative (e.g., pedophile), neutral (e.g., tablecloth) and very positive (e.g., happiness), respectively. This axis is traditionally used on its own in most sentiment analysis techniques. Similarly, A and corresponds to a word having very low (e.g., dull), moderate (e.g., watchdog), and very high (e.g., insanity) emotional intensity, respectively. Finally, D and corresponds to a word/stimulus that is very powerless (e.g., dementia), neutral (e.g., waterfall) and very powerful (e.g., paradise), respectively. The VAD ratings of each word were collected through a survey in ? (?) over 1800 participants. We directly take them as the 3-dimensional word-level affective embeddings.

Some examples of words (including nouns, adjectives, and verbs) and their corresponding VAD values are depicted in Figure 2. For instance, the VAD vectors of the words ecstatic and bored are and , respectively. This means that an average human rates the feeling of being bored as more unpleasant (V), less intense (A), and slightly weaker (D), compared with the feeling of being ecstatic. Similarly, the VAD vectors for the nouns mother and granny are and , respectively. Thus, mothers are perceived to be more pleasant (V) and more powerful (D) than grannies, and evoke more intense emotions (A). From Figure 2a, we also see some clusters {angry, hate, enraged} and {depressed, sad, bored} that are slightly apart on the A axis. Also, the cluster {sword, table, granny} is fairly neutral on the V axis, compared with the cluster {happy, mother, love} in Figure 2b.

For words missing in this dictionary, such as stop words and proper nouns, we set the VAD vector to be the neutral vector , because these words are neutral in pleasantness (V) and power (D), and evoke no arousal (A). Formally, we define “word to affective vector” (W2AV) as:


where is the lemmatization of the word .

In this way, words with similar emotional connotations are close together in the affective space, and affectively dissimilar words are far apart from each other. Thus W2AV is suitable for neural processing.

The simplest approach to utilize W2AV, perhaps, is to feed it to a Seq2Seq model as input. Concretely, we concatenate the W2AV embeddings of each word with its traditional word embeddings, the resulting vector being the input to both the encoder and the decoder.

Affective Loss Functions

Equipped with affective vectors, we further design affective training loss functions to explicitly train an affect-aware Seq2Seq conversation model. The philosophy of manipulating loss function is similar to ? (?), but we focus on affective aspects (instead of diversity in general). We have several heuristics as follows.

Minimizing Affective Dissonance.

We start with the simplest approach: maintaining affective consistency between prompts and responses. This heuristic arises from the observation that typical open-domain textual conversations between two humans consist of messages and responses that, in addition to being affectively loaded, are affectively similar to each other. For instance, a friendly message typically elicits a friendly response and provocation usually results in anger or contempt. Assuming that the general affective tone of a conversation does not fluctuate too suddenly and too frequently, we emulate human-human interactions in our model by minimizing the dissonance between the prompts and the responses, i.e., the Euclidean distance between their affective embeddings. This objective allows the model to generate responses that are emotionally aligned with the prompts.

Thus, at the time step , the loss is computed by


where denotes -norm. The first term is the standard XENT loss as in Equation 1. The sum is the average affect vector of the source sentence, whereas is the average affect vector of the target sub-sentence generated up to the current time step .

In other words, we penalize the distance between the average affective embeddings of the source and the target sentences. Notice that this affect distance is not learnable and that selecting a single predicted word makes the model indifferentiable. Therefore, we relax hard prediction of a word by its predicted probability . is a hyperparameter balancing the two factors.

Maximizing Affective Dissonance.

Admittedly, minimizing the affective dissonance does not always make sense while we model a conversation. An over-friendly message from a stranger may elicit anger or disgust from the recipient. Furthermore, from the perspective of training an open-domain conversational agent whose main purpose is to entertain, it is interesting to generate responses that are not too affectively aligned with the prompts. Therefore, we design an objective function that maximizes the dissonance by flipping the sign in the second term in Equation 4. (Details are not repeated here.)

Maximizing Affective Content.

Our third heuristic encourages Seq2Seq to generate affective content, but does not specify the polarity of sentiment. This explores the hypothesis that most of the casual human responses are not vague, dull, or emotionally neutral, while the model could choose the appropriate sentiment. Concretely, we maximize the affective content of the model’s responses, so that it avoids generating generic responses like “yes,” “no,” “I don’t know,” and “I’m not sure.” That is, at the time step , the loss function is


The second term is an affect objective that discourages non-affective words. We penalize the distance between ’s affective embedding and the affectively neutral vector , so that the model pro-actively chooses emotionally rich words.

Affectively Diverse Decoding

In this subsection, we propose affectively diverse decoding that incorporates affect into the decoding process of neural response generation.

Traditionally, beam search (BS) has been widely used for decoding in Seq2Seq models because it provides a tractable approximation of searching an exponentially large solution space. However, in the context of open-domain dialogue generation, BS is known to produce nearly identical samples like “This is great.” “This is great!” and “This is so great!” that lack semantic and/or syntactic diversity (?).

Diverse beam search (DBS) (?) is a recently proposed variant of BS that explicitly considers diversity during decoding; it has been shown to outperform BS and other diverse decoding techniques in many NLP tasks.

In the following, we give an overview of BS and DBS, followed by a description of our proposed affective variants of DBS.

Beam Search (BS).

BS maintains top- most likely (sub)sequences, where is known as the beam size. At each time step , the top- subsequences at time step are augmented with all possible actions available; then the top- most likely branches are retained at time , and the rest are pruned.

Let be the set of vocabulary tokens and let be the input sequence. Ideally, decoding of an entire sequence is given by


where is the length. BS approximates Equation 6 by computing and storing only the top- high scoring (sub)sequences (called beams) at each time step. Let denote the th beam stored at time , and let denote the set of beams stored by BS at time . Then at time , the BS objective is given by:


where is the set of all possible extensions (i.e., ) based on the beams stored at time (i.e., ).

Diverse Beam Search (DBS).

DBS aims to overcome the diversity problem in BS by incorporating diversity among candidate outputs. DBS divides the top- beams into groups (each group containing beams) and incorporates diversity between these groups by maximizing the standard likelihood term as well as a dissimilarity metric among the groups.

Concretely, DBS adds to traditional BS (Equation 7) a dissimilarity term which measures the dissimilarity between group and previous groups if token is selected to extend any beam in group . This is given by


where is a hyperparameter controlling the diversity strength. Intuitively, DBS modifies the probability in BS as a general scoring function by adding a dissimilar term between a particular sample (i.e., ) and samples in other groups (i.e., ). We refer readers to ? (?) for the details of DBS. Here, we focus on the dissimilarity metric that can incorporate affective aspects into the decoding phase.

Affectively Diverse Beam Search (ADBS).

The dissimilarity metric for DBS can take many forms as used in ? (?): Hamming diversity that penalizes tokens based on the number of times they are selected in the previous groups, n-gram diversity that discourages repetition of n-grams between groups, and neural-embedding diversity that penalizes words with similar embeddings across groups. Among these, the neural-embedding diversity metric is the most relevant to us. When used with Word2Vec embeddings, this metric discourages semantically similar words (e.g., synonyms) to be selected across different groups.

To decode affectively diverse samples, we propose to inject affective dissimilarity across the beam groups based on affective word embeddings. This can be done either at the word level or sentence level. We formalize these notions below.

Word-Level Diversity for ADBS (WL-ADBS). We define the word level affect dissimilarity metric to be


where denotes a similarity measure between two vectors. In our experiments, we use the cosine similarity function. denotes the token under consideration at the current time step for beam in group , and denotes the token chosen for beam in a previous group at time .

Intuitively, this metric computes the cosine similarity of group ’s current beam with all the beams generated in groups . The metric operates at the word level, ensuring that the word affect at time is diversified across the groups.

Sentence-Level Diversity for ADBS (SL-ADBS). The word-level metric in Equation Affectively Diverse Beam Search (ADBS). does not take into account the overall sentence affect for each group up to time . Thus, as an alternative, we propose a sentence-level affect diversity metric, given by

where (11)

Here, for is the th beam in the th group stored at time ; is the concatenation of and .

The intuition is that, to capture sentence level affect, this metric computes the cumulative dissimilarity (given by the function ) between the current beam and all the previously generated beams in other groups. This bag-of-affective-words approach is simple but works well in practice, as will be shown in the experimental results.

It should be also noticed that several other beam search-based diverse decoding techniques have been proposed in recent years, including DivMBest (?), MMI objective (?), and segment-by-segment re-ranking (?). All of them use the notion of a diversity term within BS; therefore our affect-injecting technique can be potentially used with these algorithms.


Dataset and Implementation Details

We evaluated our approach on the Cornell Movie Dialogs Corpus444The dataset is available at (?). It contains  300k utterance-response pairs, and we kept a vocabulary size of 12,000. All our model variants used a single-layer LSTM encoder and a single-layer LSTM decoder, each layer containing 1024 cells. We used Adam (?) for optimization with a batch size of 64 and other default hyperparameters. Listed as follows are detailed settings for each model.

  • For the baseline loss, we used 1024-dimensional Word2Vec embeddings as input and trained the Seq2Seq model for 50 epochs by using Equation 1.

  • For the affective embeddings as input, we used 1027-dimensional vectors, each a concatenation of 1024-D Word2Vec and 3-D W2AV embeddings. Training was also done for 50 epochs.

  • For affective loss functions (, , and ), we trained the models in two phases. In the first phase, each model was trained using loss for 40 epochs. In the second phase, each model was fine-tuned using the affective loss functions for 10 epochs. This two-phase approach was adopted because we observed inferior results (syntactical errors in particular) for single-phase training with the full loss functions for 50 epochs.

  • The ADBS decoding was deployed at test time (both word-level and sentence-level metrics, and in Equations Affectively Diverse Beam Search (ADBS). and Affectively Diverse Beam Search (ADBS)., respectively). We set for simplicity, that is, each group contains a single beam. Therefore, diversification among groups in our case was equivalent to diversification among all the beams.

  • The hyperparameters for , , and were manually tuned through validation and set to 0.5, 0.4, and 0.5, respectively. For affectively diverse BS, was set to 0.7 (Equation 8).

    Model Syntactic Natural Emotional
Coherence Approp.
Word embed. 1.48 0.69 0.41
Word + Affective 1.71 1.05 1.01
Table 1: The effect of affective word embeddings as input.
        Model Syntactic Natural Emotional
Coherence Approp.
1.48 0.69 0.41
1.75 0.83 0.56
1.74 0.85 0.58
1.71 0.95 0.71
Table 2: Effect of affective loss functions.


Recent work employs both automated metrics (e.g., BLEU, ROUGE, and METEOR) and human judgments to evaluate dialogue systems. While automated metrics enable high-throughput evaluation, ? (?) show that these metrics have weak or no correlation with human judgments. It is also unclear how to evaluate affective aspects by automated metrics. Therefore, in this work, we recruited human judges to evaluate our models, following several previous studies (??).

To evaluate the quality of the generated responses, we had 5 workers to evaluate 100 test samples for each model variant in terms of syntactic coherence (Does the response make grammatical sense?), naturalness (Could the response have been plausibly produced by a human?) and emotional appropriateness (Is the response emotionally suitable for the prompt?). For each axis, the judges were asked to assign each response an integer score of 0 (bad), 1 (satisfactory), or 2 (good). The scores were then averaged for each axis. We also evaluated inter-annotator consistency by Fleiss’ score (?), and obtained a score of 0.447, interpreted as “moderate agreement” among the judges.555 kappa

Model Syntactic Affective No. of Emotionally
Diversity Diversity Approp. Responses
BS 1.23 0.87 0.89
H-DBS 1.47 0.79 0.78
WL-ADBS 1.51 1.25 1.30
SL-ADBS 1.45 1.31 1.33
Table 3: Effect of affectively diverse decoding. H-DBS refers to Hamming-based DBS used in ? (?). WL-ADBS and SL-ADBS are the proposed word-level and sentence-level affectively diverse beam search, respectively.
Model Syntactic Coherence Naturalness Emotional Appropriateness
Traditional Seq2Seq 1.48 0.69 0.41
Seq2Seq + Affective Embeddings 1.71 1.05 1.01
Seq2Seq + Affective Embeddings & Loss 1.76 1.03 1.07
Seq2Seq + Affective Embeddings & Loss & Decoding 1.69 1.09 1.10
Table 4: Combining different affective strategies.

The evaluation of diversity was conducted separately (i.e., the results in Table 3). In this experiment, an annotator was presented with top-three decoded responses and was asked to judge syntactic diversity (How syntactically diverse are the five responses?) and emotional diversity (How affectively diverse are the five responses?). The rating scale was 0, 1, 2, and 3 with labels bad, satisfactory, good, and very good, respectively. The annotator was also asked to state the number of beams that were emotionally appropriate to the prompt. The scores obtained for each question were averaged. Moreover, we had three annotators in this experiment (fewer than the previous one), as it required more annotations (3 responses for every test sample). The Fleiss’ score for this protocol was 0.471, also signifying “moderate agreement” between the judges.

In the following, we first evaluate the performance of three affective strategies individually, namely affective word embeddings as input, affective loss functions, and affectively diverse decoding. Then we show how these strategies can be integrated.

Experiment #1: Affective word embeddings as input.

Table 1 compares Seq2Seq open-domain dialogue systems with and without the affective word embeddings. We see that the cognitively engineered affective embeddings, even with three additional features, largely improve the Seq2Seq model. The improvement is consistent in terms of all three evaluation aspects, and its effect is the most significant in emotional appropriateness.

The results show that traditional word embeddings learned end-to-end during training are not sufficient to capture emotional aspects, and that using additional knowledge makes the model more aware of affect.

Experiment #2: Affective loss functions.

We compare in Table 2 the proposed loss functions—namely minimizing affective dissonance (DMIN), maximizing affective dissonance (DMAX), and maximizing affective content (AC)—with traditional cross-entropy loss (XENT).

As shown in Table 2, DMIN and DMAX yield similar results, both outperforming XENT. Moreover, AC generally outperforms DMIN and DMAX in terms of naturalness and appropriateness. The results imply that forcing the affect vector in either direction (towards or against the previous utterance) helps the model, but its performance is worse than AC. The mediocre performance of and could be explained by the fact that the relationship between a prompt and a response is not always as simple as minimum or maximum affective dissonance. It is usually much more subtle; therefore it makes more sense to model this relationship through established sociological interaction theories like the Affect Control Theory (?). By contrast, the AC loss function encourages affective content without specifying the affect direction; it works well in practice.

Considering both Tables 1 and 2, we further notice that the affective loss function alone is not as effective as affective embeddings. This makes sense because the loss function itself does not provide additional knowledge to the neural network. However, as will be seen in Experiment #4, these affective aspects are “orthogonal” to each other, and can be directly combined.

Message Model Response
Traditional Seq2Seq Um, you know…
But you work for the casino?
I wish You are personally responsible.
I was a It’s suicide!
billion- Aff. Emb. Don’t trouble yourself about it.
aire. Aff. Emb. Would it matter in the end.
Aff. Emb. Would it change you?
I hate Traditional Seq2Seq It was supposed to be.
you. Aww c all guys do…
Keeps fairies away!
I haven’t done anything crazy.
Aff. Emb. Don’t say that!
Aff. Emb. I still love you!
Aff. Emb. I don’t want to fight you.
Table 5: Examples of the responses generated by the baseline (traditional Seq2Seq) and affective models.

Experiment #3: Affectively Diverse Decoding.

We now evaluate our affectively diverse decoding methods. Since evaluating diversity requires multiple decoded utterances for a test sample, we adopted a different evaluation setting as described before.

Table 3 compares both word-level and sentence-level affectively diverse BS (WL-ADBS and SL-ADBS, respectively) with the original BS and Hamming-based DBS used in ? (?). We see that WL-ADBS and SL-ADBS beat the baselines BS and Hamming-based DBS by a fair margin on affective diversity as well as number of emotionally appropriate responses. SL-ADBS is slightly better than WL-ADBS as expected, since it takes into account the cumulative affect of sentences as opposed to individual words.

Experiment #4: Putting them all together.

We show in Table 4 how the affective word embeddings, loss functions, and decoding methods perform when they are combined. Here, we chose the best variants in the previous individual tests: the loss function maximizing affective content () and the sentence level diversity measure (SL-ADBS).

As shown, the performance of our model generally increases when we gradually add new components to it. This confirms that the three affective strategies can be directly combined for further improvement, and we achieve significantly better performance compared with the original Seq2Seq model, especially in terms of emotional appropriateness.

Note that our setting is different from the Emotional Chat Machine (ECM) (?), the only other known emotion-based neural dialogue system to the best of our knowledge. ECM requires a desired affect category as input, which is unrealistic in applications. It also differs from our experimental setting (and our research goal), making direct comparison infeasible. However, our proposed affective approaches can be potentially integrated to ECM in addition to their manually specified emotion category.

Case study.

We present several sample outputs of all the models in Table 5 to give readers a taste of how the responses differ. responses are generic and non-committal, as expected. tries to match the affect of the word billionaire with casino, responds to hate with fairies, maximizes affective content of the responses with the words suicide and crazy. with affective embeddings produces responses with more subtle affective connotations.

Conclusion and Future Work

In this work, we address the problem of affective neural dialogue generation, which is useful in applications like emotional conversation partners to humans. We advance the development of affectively cognizant neural encoder-decoder dialogue systems by three affective strategies. We embed linguistic concepts in an affective space with a cognitively engineered dictionary, propose several affect-based heuristic objective functions, and introduce affectively diverse decoding methods.

In the future, we would like to investigate affect-based attention mechanisms for neural conversational models. We would also like to explore affect-based personalization of neural dialogue systems using reinforcement learning.


We thank Marc-André Cournoyer for his helpful Github repositories on neural conversation models, and the annotators who helped us evaluate our models.


  • [Alhothali and Hoey 2015] Alhothali, A., and Hoey, J. 2015. Good news or bad news: Using affect control theory to analyze readers’ reaction towards news articles. In NAACL-HLT, 1548–1558.
  • [Asghar et al. 2017] Asghar, N.; Poupart, P.; Jiang, X.; and Li, H. 2017. Deep active learning for dialogue generation. In Proc. Conf. Lexical and Computational Semantics, 78–83.
  • [Bahdanau, Cho, and Bengio 2015] Bahdanau, D.; Cho, K.; and Bengio, Y. 2015. Neural machine translation by jointly learning to align and translate. In ICLR.
  • [Callejas, Griol, and López-Cózar 2011] Callejas, Z.; Griol, D.; and López-Cózar, R. 2011. Predicting user mental states in spoken dialogue systems. EURASIP J. Advances in Signal Processing 2011(1):6.
  • [Cao and Clark 2017] Cao, K., and Clark, S. 2017. Latent variable dialogue models and their diversity. In EACL, volume 2, 182–187.
  • [Danescu-Niculescu-Mizil and Lee 2011] Danescu-Niculescu-Mizil, C., and Lee, L. 2011. Chameleons in imagined conversations: A new approach to understanding coordination of linguistic style in dialogs. In Proc. Workshop on Cognitive Modeling and Computational Linguistics, 76–87.
  • [Fleiss 1971] Fleiss, J. L. 1971. Measuring nominal scale agreement among many raters. Psychological Bulletin 76(5):378–382.
  • [Ghosh et al. 2017] Ghosh, S.; Chollet, M.; Laksana, E.; Morency, L.-P.; and Scherer, S. 2017. Affect-LM: A neural language model for customizable affective text generation. In ACL, 634–642.
  • [Gimpel et al. 2013] Gimpel, K.; Batra, D.; Dyer, C.; Shakhnarovich, G.; and Tech, V. 2013. A systematic exploration of diversity in machine translation. In EMNLP, 1100–1111.
  • [Hoey, Schröder, and Alhothali 2016] Hoey, J.; Schröder, T.; and Alhothali, A. 2016. Affect control processes: Intelligent affective interaction using a partially observable markov decision process. Artificial Intelligence 230:134–172.
  • [Kingma and Ba 2015] Kingma, D., and Ba, J. 2015. Adam: A method for stochastic optimization. In ICLR.
  • [Li et al. 2016a] Li, J.; Galley, M.; Brockett, C.; Gao, J.; and Dolan, B. 2016a. A diversity-promoting objective function for neural conversation models. In NAACL-HLT, 110–119.
  • [Li et al. 2016b] Li, J.; Monroe, W.; Ritter, A.; and Jurafsky, D. 2016b. Deep reinforcement learning for dialogue generation. In EMNLP, 1192–1202.
  • [Li et al. 2017] Li, J.; Monroe, W.; Shi, T.; Ritter, A.; and Jurafsky, D. 2017. Adversarial learning for neural dialogue generation. In EMNLP, 2147–2159.
  • [Li, Monroe, and Jurafsky 2016] Li, J.; Monroe, W.; and Jurafsky, D. 2016. A simple, fast diverse decoding algorithm for neural generation. arXiv preprint arXiv:1611.08562.
  • [Liu et al. 2016] Liu, C.-W.; Lowe, R.; Serban, I.; Noseworthy, M.; Charlin, L.; and Pineau, J. 2016. How not to evaluate your dialogue system: An empirical study of unsupervised evaluation metrics for dialogue response generation. In EMNLP, 2122–2132.
  • [Malhotra et al. 2015] Malhotra, A.; Yu, L.; Schröder, T.; and Hoey, J. 2015. An exploratory study into the use of an emotionally aware cognitive assistant. In AAAI Workshop: Artificial Intelligence Applied to Assistive Technologies and Smart Environments.
  • [Mikolov et al. 2010] Mikolov, T.; Karafiát, M.; Burget, L.; Cernockỳ, J.; and Khudanpur, S. 2010. Recurrent neural network based language model. In INTERSPEECH, 1045–1048.
  • [Mikolov et al. 2013] Mikolov, T.; Sutskever, I.; Chen, K.; Corrado, G. S.; and Dean, J. 2013. Distributed representations of words and phrases and their compositionality. In NIPS, 3111–3119.
  • [Mou et al. 2016] Mou, L.; Song, Y.; Yan, R.; Li, G.; Zhang, L.; and Jin, Z. 2016. Sequence to backward and forward sequences: A content-introducing approach to generative short-text conversation. In COLING, 3349–3358.
  • [Osgood 1952] Osgood, C. E. 1952. The nature and measurement of meaning. Psychological Bulletin 49(3):197–237.
  • [Pennebaker, Francis, and Booth 2001] Pennebaker, J. W.; Francis, M. E.; and Booth, R. J. 2001. Linguistic Inquiry and Word Count. Erlbaum Publishers.
  • [Pittermann, Pittermann, and Minker 2010] Pittermann, J.; Pittermann, A.; and Minker, W. 2010. Emotion recognition and adaptation in spoken dialogue systems. Int. J. Speech Technology 13(1):49–60.
  • [Robison, McQuiggan, and Lester 2009] Robison, J.; McQuiggan, S.; and Lester, J. 2009. Evaluating the consequences of affective feedback in intelligent tutoring systems. In Proc. Int. Conf. Affective Comput. and Intell. Interaction and Workshops, 1–6.
  • [Rush, Chopra, and Weston 2015] Rush, A. M.; Chopra, S.; and Weston, J. 2015. A neural attention model for abstractive sentence summarization. In EMNLP, 379–389.
  • [Serban et al. 2017] Serban, I. V.; Sordoni, A.; Lowe, R.; Charlin, L.; Pineau, J.; Courville, A.; and Bengio, Y. 2017. A hierarchical latent variable encoder-decoder model for generating dialogues. In AAAI, 3295–3301.
  • [Shang, Lu, and Li 2015] Shang, L.; Lu, Z.; and Li, H. 2015. Neural responding machine for short-text conversation. In ACL, 1577–1586.
  • [Shao et al. 2017] Shao, L.; Gouws, S.; Britz, D.; Goldie, A.; Strope, B.; and Kurzwell, R. 2017. Generating high-quality and informative conversation responses with sequence-to-sequence models. arXiv preprint arXiv:1701.03185.
  • [Sutskever, Vinyals, and Le 2014] Sutskever, I.; Vinyals, O.; and Le, Q. V. 2014. Sequence to sequence learning with neural networks. In NIPS, 3104–3112.
  • [Vijayakumar et al. 2016] Vijayakumar, A. K.; Cogswell, M.; Selvaraju, R. R.; Sun, Q.; Lee, S.; Crandall, D.; and Batra, D. 2016. Diverse beam search: Decoding diverse solutions from neural sequence models. arXiv preprint arXiv:1610.02424.
  • [Warriner, Kuperman, and Brysbaert 2013] Warriner, A. B.; Kuperman, V.; and Brysbaert, M. 2013. Norms of valence, arousal, and dominance for 13,915 English lemmas. Behavior Research Methods 45(4):1191–1207.
  • [Wen et al. 2015] Wen, T.-H.; Gasic, M.; Mrkšić, N.; Su, P.-H.; Vandyke, D.; and Young, S. 2015. Semantically conditioned lstm-based natural language generation for spoken dialogue systems. In EMNLP, 1711–1721.
  • [Wen et al. 2017] Wen, T.-H.; Miao, Y.; Blunsom, P.; and Young, S. 2017. Latent intention dialogue models. In ICML, 3732–3741.
  • [Xing et al. 2017] Xing, C.; Wu, W.; Wu, Y.; Liu, J.; Huang, Y.; Zhou, M.; and Ma, W.-Y. 2017. Topic aware neural response generation. In AAAI, 3351–3357.
  • [Zhou et al. 2017] Zhou, H.; Huang, M.; Zhang, T.; Zhu, X.; and Liu, B. 2017. Emotional chatting machine: Emotional conversation generation with internal and external memory. arXiv preprint arXiv:1704.01074.
Comments 0
Request Comment
You are adding the first comment!
How to quickly get a good reply:
  • Give credit where it’s due by listing out the positive aspects of a paper before getting into which changes should be made.
  • Be specific in your critique, and provide supporting evidence with appropriate references to substantiate general statements.
  • Your comment should inspire ideas to flow and help the author improves the paper.

The better we are at sharing our knowledge with each other, the faster we move forward.
The feedback must be of minimum 40 characters and the title a minimum of 5 characters
Add comment
Loading ...
This is a comment super asjknd jkasnjk adsnkj
The feedback must be of minumum 40 characters
The feedback must be of minumum 40 characters

You are asking your first question!
How to quickly get a good answer:
  • Keep your question short and to the point
  • Check for grammar or spelling errors.
  • Phrase it like a question
Test description