In the emotion detection stage, we try to classify and map an utterance to an emotional category. It is important to note, that emotion detection is strongly tied with response generation as similar approaches are used for both stages.

One of the first challenges is setting the number of emotion categories to be used for classification. There is no universally accepted model of emotions and the number of emotions differs drastically depending on the underlying model. One of the most popular models in emotion detection is Ekman’s six basic emotions — happiness, sadness, fear, anger, disgust, and surprise (Ekman, 1992) . Other popular models include Plutchik’s wheel of emotions (Plutchik, 1991) or Parrot’s Emotional Layers (Parrott, 2001) consisting of thirty-one different emotions. However, the latter two models are seldomly used for emotion detection, since more emotion categories mean additional complexity for the classification task.

Seyeditabari et al. (2018) review existing works and approaches in the field of emotion detection and provide a good overview of the state-of-theart of emotion detection in text. They list different resources used for detecting emotions in text such as labelled text, emotion lexicons, or word embeddings and elaborate on common approaches used for emotion detection. Seyeditabari et al. (2018) conclude that there is still potential for improving emotion detection in text. Thereby, the complex nature of emotion expression, the shortage of quality data and inefficient models induce most challenges for future work.

In this survey, we distinguish three major approaches used for emotion detection in text — rule-based, non-neural machine learning and deep learning.

Rule-based approaches mainly use emotion lexicons or word embeddings. Both approaches are based on keyword lookup from text to detect the underlying emotion. The rule-based approach is only as good, as is its parsing algorithm, and the quality of the lexical resource used for the lookup. Emotion lexicons list emotion-bearing words and classify them to single or multiple emotional categories. Word embeddings, on the other hand, also take into account frequently co-occurring words that are semantically similar.

Emotion lexicons can be built from scratch. However, there exist good off-the-shelf solutions. One of the most popular being WordNetAffect (Strapparava et al., 2004) . These off-theshelf solutions differ tremendously in terms of their number of entries. WordNet-Affect contains close to five thousand words, whereas DepecheMood (Liu and Zhang, 2012) , another popular lexicon, contains more than thirty-five thousand words. Nonetheless, the quality of the lexicon is not solely dependent on its size. The vocabulary used for the lexicon also impacts its quality. Bandhakavi et al. (2017) argue that generalpurpose lexicons such as WordNet-Affect perform not as good as domain-specific emotion lexicons. Therefore, a smaller domain-specific lexicon might yield in better results than a larger general-purpose lexicon. LIWC-based lexicons (Linguistic Inquiry and Word Count) are also widely used, since these dictionaries list grammatical, psychological, and content word categories, and thus also emotion categories with thousands of entries (Chung and Pennebaker, 2012) .

The idea behind word embeddings is similar to emotion lexicons. Each word is represented as a vector in the vector space. Thereby, frequently cooccurring words are considered semantically similar and therefore, close in the vector space (Seyeditabari et al., 2018) . Among the most popular word embedding methods is word2Vec (Mikolov et al., 2013) . Word embeddings are also often used to train machine learning models, like LSTM, which usually take word vectors as inputs.

Both rule-based approaches are straightforward. However, there are some drawbacks to them. The emotional meaning of keywords can be ambiguous and is context-sensitive. The sentences She hates me, and I hate her, could both be classified as anger based on the keyword hate. However, when looking at the sentence level information, the first utterance could also be perceived as sad. Ignoring the syntactic structure and semantics of the whole sentence, can therefore lead to misinterpretations. Furthermore, sentences without any emotional keywords cannot be classified. Even if they might contain an implicit expression of emotions, for example in the form of a metaphor (Kao et al., 2009) . As a consequence, especially emotion lexicons often lack accuracy compared to more complex approaches.

Unlike rule-based approaches, non-neural learning-based approaches are trying to detect emotions using trained classifiers, such as the Support Vector Machine (SVM) (Teng et al., 2006) , Naive Bayes, or Decision Trees.

We distinguish between supervised and unsupervised learning. Unsupervised approaches are an evolution of the rule-based approaches and are learning from test data that is not annotated with emotional labels. Most commonly, these approaches use movie dialogues (Banchs, 2017; Honghao et al., 2017) or children’s fairy tales (Kim et al., 2010) to build emotional lexicons and train their models.

Supervised approaches, on the other hand, learn from labelled data such as Twitter messages. Common labels are annotations, hashtags or emojis. There exist a few good sources for emotionally labelled text, one of the most prominent being the Swiss Center for Affective Sciences7 providing datasets like the International Survey On Emotion Antecedents And Reactions (ISEAR) and other useful tools for emotion detection. Other well-known datasets are EmotiNet (Balahur et al., 2011) and SemEval-2007 (Strapparava and Mihalcea, 2007) .

As stated by Seyeditabari et al. (2018), one of the major challenges for supervised approaches is the lack of quality training data. Oftentimes, these datasets are unbalanced in terms of emotion categories. Banchs (2017) analyse the large movie dialogue dataset MovieDiC and conclude that emotions such as love, or joy occur much more frequently than fear or surprise. The classifiers trained on such datasets will therefore underperform for these emotional categories. 2.2.3

Most recent advances that showed to be effective in the field of emotion detection, have been made using deep learning (Xu et al., 2017) .

Oftentimes, deep learning approaches are covering both, the emotion detection and the response generation, for example when using an EncoderDecoder architecture (Serban et al., 2015) . This architecture consists of two stages — the encoding and decoding stage. In the encoding stage, the raw text input is turned into a feature representation, usually in the form of a vector. The vector is then used as an input for the decoding stage to generate a response by applying the same strategies as in the encoding stage, but in the opposite direction.

A well-known approach applying the encodingdecoding architecture is Long Short-Term Memory (LSTM) (Jithesh et al., 2017) . LSTM is a Recurrent Neural Network (RNN) that allows to capture long-term dependencies and store sequential information over a longer time. It can retain and forget the previous state and memorise extracted information from the input data depending on its importance (Xu et al., 2017; Sun et al., 2019) .

The commonly used Sequence to Sequence (Seq2Seq) model also uses LSTMs and the Encoding-Decoding architecture (Sutskever et al., 2014) . There is one LSTM for the encoding stage, transforming the raw text input into a fixed-length vector representation, whereas another LSTM is used for the decoding to a variable-length text out7https://www.unige.ch/cisa/ put (Cho et al., 2014; Xu et al., 2017; Chan and Lui, 2018) .

To improve the model’s efficiency, Chan and Lui (2018) investigate different approaches on embedding emotional information for Seq2Seq models. Different styles, positioning, and embeddings of emotional information are tested. The authors conclude that the positioning in general matters and impacts the emotion detection. 2.3

One of the most difficult tasks for empathic chatbots is generating an empathic response. Firstly, because it faces similar challenges as the emotion detection stage. Secondly, because it not only has to ensure that the response is appropriate in terms of content level information, but also in terms of emotion level information. This balancing act is tremendously difficult, as one usually has to sacrifice accuracy for one of the information levels, when trying to optimise the other (Xu et al., 2017; Zhou et al., 2017) .

In terms of empathic response generation, we distinguish between two strategies — retrievalbased approaches and dynamic generation.

These approaches look up common responses to the user’s utterance in conversation datasets. However, this method is very limited in its applicability. Similar inputs yield in the same responses, making the conversation repetitive and less natural. Furthermore, huge datasets of emotional conversations are required for such systems in order to achieve acceptable results. As such datasets are scare, these types of approaches tend to yield in responses such as I don’t know in cases where no candidate response can be found.

A more advanced version of retrieval-based systems uses word embeddings on the input text to find the closest candidate responses, thus yielding in slightly more diverse responses (Bartl and Spanakis, 2017) . However, it still requires lots of emotionally charged sample conversations.

Empathic response generation in general requires similar datasets as emotion detection, but with a bigger focus on conversation and dialogue turns. Movie dialogues are a good source for emotionally charged conversations. However, they oftentimes do not resemble daily conversation and seem more artificial and theatrical (Chan and Lui, 2018) . Furthermore, emotions in movie dialogues are after all still acted and not naturally occurring, which might also have an impact on the quality of the training data.

Other common datasets include chat conversations, for example from Twitter service requests (Xu et al., 2017) that might yield in more natural conversations. 2.3.2

These approaches are strongly tied with the deep-learning approaches used for emotion detection from Section 2.2.3 and usually based on the encoder-decoder architecture, such as the Sequence-to-Sequence model.

The input sentence is encoded on a word-byword basis by embedding each word separately, whilst taking into account already encoded words using hidden states. Thereby, the last word embedding will produce a vector representation of the whole input sequence, encapsulating all relevant sentence level information. Semantically similar sentences are therefore close to each other in a vector space (Sutskever et al., 2014) .

The decoder will then use the sentence vector or rather sentence embedding to produce an output sentence using inverted encoding mechanisms on a word-by-word basis. This allows encodingdecoding architectures to generate variable length responses. Thereby, it will consider already decoded words to ensure that the generated response is also grammatically correct. To find appropriate responses to a given word from the encoding stage, vocabularies or word-embeddings are being used.

However, the longer the input sequence, the more challenging to capture the full meaning in a single sentence embedding. For an input sentence of 30 words, the decoder would have to consider, what was encoded 30 steps ago, just to decode the first word. This long-range dependency problem oftentimes results in poor responses, like I don’t know (Chan and Lui, 2018) .

Attention mechanisms are commonly used to tackle the issue of long-range dependencies (Sutskever et al., 2014) . Using attention, the decoder has direct access to the hidden state of each encoded word and can weight each word correspondingly. This allows the decoder to attend and weight on relevant parts of an input sentence, when generating the output. This mechanism is also applied in Neural Machine Translation (Bahdanau et al., 2014) .

However, these Recurrent Neural Networks (RNN) still suffer from the vanishing gradient problem, that causes issues with long-range dependencies (Hochreiter, 1998) . LSTMs also apply attention and in addition allow to retain and forget information, therefore handling the long-range dependency problem better than other approaches.

All these mechanisms are essentially required to ensure an appropriate response in terms of content level information. When we also want to consider emotion level information, we add additional complexity to the model. Emotions either have to be additionally encoded during the encoding stage or fed directly to the decoding stage to generate emotional responses (Zhou et al., 2017) .

The Emotional Chatting Machine as proposed by Zhou et al. (2017) is a recent and noteworthy approach for assessing the emotional state of conversations and to generate appropriate emotional responses. Therefore, it belongs to the dynamic generation approaches. The ECM deep learning algorithm is trained with 22.300 Chinese blog posts that are manually annotated with Ekman’s six basic emotions.

In terms of architecture, the ECM is based on the Seq2Seq model with an encoding and decoding phase. In addition, Zhou et al. (2017) introduce an internal and external memory to the model to capture changes in the emotion state throughout the sentence and to map explicit emotion expressions to emotion categories. Figure 2 provides a good overview of the ECM architecture. As input, the ECM requires the user’s text message and one of the Ekman’s six emotion categories to condition the response. Based on the input message, the ECM will generate an appropriate response and condition it using the input emotion category. Zhou et al. (2017) benchmark the ECM with other approaches, such as the traditional Seq2Seq model or lexicon-based approaches and show that ECM performs best across all emotion categories. However, it lacks behind slightly on the content level that the authors put down to an imbalance in the training set.

One drawback of the ECM is that the input emotion has to be set manually. This hardwiring of the output emotion can be useful, if the chatbot should always respond in the same emotional state, or express certain personality traits such as being angry all the time. However, if we want the chatbot to dynamically react to the user’s emotions and make the interaction natural, then we have to change the emotion category based on the emotional state of the user’s message automatically.

As discussed in Section 1, empathy requires understanding of the user’s emotion and replying to them on an appropriate emotional level. Using emotion detection, we can achieve good results in understanding the user’s emotions. The difficult part is actually selecting the appropriate emotion to condition the response with.

One approach could be to simply mirror the user’s emotion. However, in humanconversations, empathy finds expression, when one tries to feel with the opponent and not necessarily similar to the opponent. One’s own emotion must not be confused with the opponent’s emotion. When resonating to the opponent’s emotion, one is still aware that it might be different from the personal emotion (Singer and Klimecki, 2014) . If someone is sad, you might understand this sadness and try to cheer them up, instead of responding in a sad way as well. However, it does not mean that you are necessarily feeling sad as well. Thus, simply mirroring the user’s emotion does not necessarily yield in empathic responses.

This is also an important aspect with regards to a chatbot’s personality, since in these cases, one should think about the chatbot’s own emotion, as well as how it might resonate on someone else’s emotions.

Another crucial aspect is taking into account the user’s emotional evolution throughout the whole conversation. When just considering the latest user utterance for emotion detection, misinterpretations or frequent switches in the emotions expressed by the chatbot’s response might occur. For example, the user could genuinely be in a bad mood, but laugh at a joke one just made. If we would consider just the latest user utterance to detect the user’s emotion and condition the response accordingly, the chatbot’s expressed emotion would switch from negative to positive within a single sentence. Modelling the user’s emotion over a longer period might also be important when applying empathic chatbots in treating mental illnesses, or when building personality profiles to monitor the emotional state of the user.

We observe that only little research actually focusses on the generation of empathic responses in Computer Science, compared to the efforts done for emotion detection. There exist some approaches, such as the Encoder-Decoder architecture, that cover emotion detection as well as emotional response generation. Nonetheless, an emotional response is not necessarily an empathic response as elaborated before.

How humans are generating empathic responses is still an ongoing field of research in neuroscience (Shamay-Tsoory and Lamm, 2018) . Similar to emotions, there is no universally accepted model for empathic responses, except that empathy is heavily context-dependent (Singer and Klimecki, 2014) . 2.4

In a normal conversation, non-verbal cues such as facial expressions or gestures can help indicate a person’s emotional state. However, with chatbots, we are missing such information and have to focus solely on the user’s text, to detect the emotional state. Similar constraints also apply to the response expression by the chatbot. It is difficult to transport the intended emotion from the generated response back to the user in a text format. Some approaches try to simulate non-verbal cues by displaying the chatbot as a 3D simulation of a person (Tatai et al., 2003) . We note that in general, chatbots do not express responses in any other way than text. Because such non-verbal cues are missing in traditional electronic messaging systems, people are using emojis to supply such cues. The usage of emojis has increased heavily over the previous years. In 2017, Facebook revealed that on an average day, over 5 Billion emojis are being sent over Messenger only8. Hu et al. (2017) state that 8https://www.adweek.com/digital/ facebook-world-emoji-day-stats-theemoji-movie-stickers/ the main reason for the usage of emojis in messages is to express emotions or strengthen expressions.

Emojis could therefore also be considered when detecting the user’s emotion. However, two challenges arise when using emojis as possible emotional labels. First, the emoji label could be contradictory to the perceived emotional state from the text, for example implying a sarcastic utterance. Figure 3 shows, how emojis can lead to such contradicting interpretations. Second, emojis are prone to cultural differences as stated by Ljubesˇic´ and Fisˇer (2016). Chatbots with a global scope should therefore take into account, that emojis might be used and perceived differently depending on the country.

DeepMoji (Felbo et al., 2017) is an impressive tool translating text into a set of emojis expressing a similar emotional state returning the five most likely emojis together with their probabilities. It is trained on 1.2 billion tweets containing emojis and uses LSTM to predict the most appropriate emojis. Generated emojis could be mapped to different emotion categories and used to express emotions in the response to the user. We leave the validation of this method for future work.

We state that the usage of emojis in conversational agents might be a clue to make them more human-like and to help expressing non-verbal cues that otherwise might go missing. Future work should therefore focus on validating this hypothesis. 3