Optimizing emotional response devices using common sense knowledge graphs integrated with emoticons

Abstract

The present invention provides a device for optimizing emotional responses using a commonsense knowledge graph incorporating emoticons. The device comprises a dictionary graph storage module for storing an emotional lexicon and an emoticon-enhanced commonsense knowledge graph; an emotional context knowledge subgraph generation module for generating an emotional context knowledge subgraph based on the current conversation, the emotional lexicon, and the emoticon-enhanced commonsense knowledge graph; and a response generation module comprising a conditional variational autoencoder for inputting the emotional context knowledge subgraph, the current utterance, and the emotion label into the conditional variational autoencoder to generate an emotional response. In summary, this method can generate emotional responses that are rich in emotional expression and accurate.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence, and specifically relates to an emotional response device that optimizes an emotional response by utilizing a common sense knowledge graph integrated with emoticons. Background Art

[0002] Rationality and emotion are two fundamental elements of being human. Imbuing conversational agents with both rationality and emotion has been a key milestone in artificial intelligence. The difference between humans and machines lies in their ability to connect conversations to their own background knowledge, while machines can only capture limited information from the surface text of conversational messages. Consequently, machines struggle to fully understand the context of conversations and utterances, often generating generic, meaningless responses. Furthermore, the lack of external knowledge makes it difficult for emotional dialogue systems to perceive underlying emotions and learn emotional interactions from limited conversational history.

[0003] Emotion is crucial for conversational models. By building conversational models and incorporating emotional information, we can improve the ability to simulate human conversations, enhance conversation quality, and promote the development of conversation generation. Previous research aimed to enable AI to perceive and understand emotions and classify explicit or implicit emotions contained in text. In recent years, researchers have begun incorporating emotions into conversation generation models, enabling machines to learn to recognize, understand, and express emotions. Emotional conversations attempt to organically combine sentiment analysis and conversation generation, but go beyond simple summation. Emotion-integrated conversational models require comprehensive control over knowledge, context, emotion, and even personality, ensuring consistent context, coherent and diverse responses, and emotional perception and expression. These challenges confront emotion-enhanced open-domain conversation generation.

[0004] While existing knowledge-based conversational models have achieved impressive results by incorporating large-scale knowledge graphs to enhance conversational quality and bridge the knowledge gap between humans and machines, they lack emotion. In real social scenarios, people combine their background knowledge with their perception and understanding of the speaker's emotions to organize their speech and fully express their feelings and attitudes. However, existing conversational models are unable to generate responses with rich emotional expression based on the conversation and existing knowledge graphs. Summary of the Invention

[0005] The present invention is made to solve the above-mentioned problems, and its purpose is to provide an emotional response device that optimizes the common sense knowledge graph by incorporating emoticons.

[0006] The present invention provides an emotional reply device optimized by using a common sense knowledge graph integrated with emoticons, which is used to generate corresponding emotional replies according to current discourse and emotional labels, and has the following characteristics: a dictionary graph storage module, which is used to store the common sense knowledge graph enhanced with emotional dictionaries and emoticons; an emotional context knowledge subgraph generation module, which is used to generate an emotional context knowledge subgraph according to the current conversation, the emotional dictionary and the common sense knowledge graph enhanced with emoticons; a reply generation module, which includes a conditional variational autoencoder, which is used to input the emotional context knowledge subgraph, the current discourse and the emotional labels into the conditional variational autoencoder to obtain an emotional reply, wherein the common sense knowledge graph enhanced with emoticons is constructed according to the existing common sense knowledge graph and multiple emoticon triples, and the common sense knowledge graph enhanced with emoticons includes multiple triples, and the triples and the emoticon triples are both It includes head entities, relations and tail entities. The emotional context knowledge subgraph generation module includes a candidate tuple generation submodule, an emotional tuple generation submodule and a knowledge subgraph generation submodule. The candidate tuple generation submodule is used to retrieve the triple corresponding to the head entity from the emoticon-enhanced common sense knowledge graph for each word in the current discourse as the initial candidate tuple, and then use the initial triple with a confidence score greater than the confidence threshold as the candidate tuple of the word. The emotional tuple generation submodule is used to calculate the emotional scores of the tail entity word and each emotional word in the emotional dictionary for the tail entity of the candidate tuple, and then sort the emotional scores from large to small, select the emotional words corresponding to the top k emotional scores, and form emotional tuples with the corresponding words in the current discourse. The knowledge subgraph generation submodule is used to construct an emotional context knowledge subgraph based on all candidate tuples and emotional tuples.

[0007] In the device for optimizing emotional responses using the common sense knowledge graph integrated with emoticons provided by the present invention, it may also have the following features: wherein, the specific steps of generating emoticon triples based on existing discourse, specified symbolic emotional tags and corresponding replies are: step S1, extracting multiple concept words from the discourse as discourse concept words; step S2, extracting multiple concept words from the replies as reply concept words; step S3, constructing a first PPMI matrix based on all discourse concept words and reply concept words; step S4, selecting discourse concept words and reply concept words corresponding to PPMI values ​​greater than or equal to a preset first PPMI threshold from the first PPMI matrix as first strong association pairs; step S5, constructing a second PPMI matrix based on all first strong association pairs and symbolic emotional tags; step S6, selecting first strong association pairs and symbolic emotional tags corresponding to PPMI values ​​greater than or equal to a preset second PPMI threshold from the second PPMI matrix as second strong association pairs; step S7, taking the discourse concept words in the second strong association pairs as head entities, the symbolic emotional tags as relationships, and the reply concept words as tail entities, thereby obtaining emoticon triples.

[0008] The device for optimizing emotional response by utilizing common sense knowledge graphs incorporating emoticons provided by the present invention may also have the following feature: wherein the first PPMI threshold is 1.

[0009] The device for optimizing emotional response by utilizing the common sense knowledge graph incorporating emoticons provided by the present invention may also have the following features: wherein, emoticon triples are combined with the common sense knowledge graph through the knowledge graph embedding model TransE to obtain an emoticon-enhanced common sense knowledge graph.

[0010] The device for optimizing emotional response by using a common sense knowledge graph incorporating emoticons provided by the present invention may also have the following features: wherein the confidence threshold is 0.2.

[0011] In the device for optimizing emotional responses using a common sense knowledge graph incorporating emoticons provided by the present invention, the device may also have the following features: wherein, during the training process of the conditional variational autoencoder, the specific process of introducing an emotion category classifier to optimize the parameters of the conditional variational autoencoder is as follows: inputting the utterance x and the given emotion label c into the conditional variational autoencoder to generate a response x', and then inputting the response x' into the emotion category classifier to obtain a ranking of the probabilities of all emotion labels from high to low, and adjusting the parameters according to the ranking until the given emotion label c is at the top of the ranking, thereby completing the parameter optimization.

[0012] The device for optimizing emotional responses using a common sense knowledge graph incorporating emoticons provided by the present invention may also have the following features: it also includes an emoticon selection module for inserting emoticons that conform to the context and emotional labels into emotional responses.

[0013] Functions and effects of the invention

[0014] According to the present invention, the emotion reply device optimized by using the common sense knowledge graph integrated with emoticons is provided, because the dictionary graph storage module of the common sense knowledge graph enhanced with the emotion dictionary and emoticons enables the common sense knowledge graph to optimize the emotion reply device with richer external knowledge with emotion relationships; the emotion context knowledge subgraph generation module extracts relevant knowledge from external knowledge according to the current discourse to construct an emotion context knowledge subgraph for assisting in generating responses containing emotions; the response generation module including the trained conditional variational autoencoder generates an emotion response that is more in line with the emotion label according to the emotion context knowledge subgraph; the emoticon selection module inserts emotion symbols into the emotion response to make the entire response more natural and vivid. Therefore, the present invention's emotion reply device optimized by using the common sense knowledge graph integrated with emoticons can generate emotion responses that are rich in emotion expression and accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Schematic diagram of the principle of generating emotional responses in an embodiment of the present invention;

[0016] Figure 2 is a schematic diagram of a process for generating emoticon triples in an embodiment of the present invention;

[0017] Figure 3 is a schematic diagram of an emoticon-enhanced common sense knowledge graph in an embodiment of the present invention;

[0018] Figure 4 2 is a schematic diagram of a framework of a common sense knowledge graph optimization emotion response device according to an embodiment of the present invention;

[0019] Figure 5 3 is a schematic diagram of the framework of the emotional context knowledge subgraph generation module in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the following embodiments, combined with the accompanying drawings, specifically illustrate the present invention's use of common sense knowledge graphs that incorporate emoticons to optimize the emotional response device.

[0021] Figure 1 2 is a schematic diagram of the principle of generating emotional responses in an embodiment of the present invention.

[0022] like Figure 1 As shown, emoji triples are generated based on the existing discourse, specified symbolic emotional tags and corresponding replies, and then an emoji-enhanced common sense knowledge graph is constructed based on the emoji triples. Then, an emotional context knowledge subgraph is constructed based on the current discourse, the emoji-enhanced common sense knowledge graph and the emotional dictionary. Finally, the current discourse, the emotional context knowledge subgraph and the emotional tags are used to generate the corresponding emotional replies, where the emoji triples and the triples in the emoji-enhanced common sense knowledge graph include head entities, relations and tail entities, that is, the triples are in the form of (head entity, relation, tail entity).

[0023] In this embodiment, discourse concepts, emoticons, and reply concepts are extracted from existing discourses, specified symbolic emotion tags, and corresponding replies to construct emoticon triples, so that an emoticon triple can represent the emoticon relationship implied from a discourse concept to a reply concept in a complete conversation. Specifically, a method based on point mutual information (PMI) is used to extract emoticon triples, because the PMI method can find mutual dependencies between two given target objects through calculation.

[0024] Figure 2 1 is a flow chart of generating emoticon triples in an embodiment of the present invention.

[0025] like Figure 2 As shown in Figure 2, the specific steps for generating emoji triples using the PMI method based on the existing utterances, the specified symbolic emotion labels, and the corresponding replies are as follows:

[0026] Step S1: extract multiple concept words from the discourse as discourse concept words, namely discourse concepts.

[0027] Step S2: extract multiple concept words from the reply as reply concept words, namely reply concepts.

[0028] Step S3: construct a first PPMI matrix PPMI (discourse concept, reply concept) based on all discourse concept words and reply concept words.

[0029] Step S4: selecting, from the first PPMI matrix, discourse concept words and response concept words corresponding to PPMI values ​​greater than or equal to a preset first PPMI threshold as first strongly associated pairs.

[0030] Among them, the first PPMI threshold is 1.

[0031] Step S5: construct a second PPMI matrix PPMI ({discourse concept, response concept}, symbolic emotion label) according to all first strongly associated pairs and symbolic emotion labels.

[0032] Step S6: Selecting a first strongly associated pair and a symbolic emotion tag corresponding to a PPMI value greater than or equal to a preset second PPMI threshold from the second PPMI matrix as a second strongly associated pair.

[0033] In step S7, the discourse concept word in the second strongest association pair is used as the head entity, the symbolized emotion label is used as the relationship, and the reply concept word is used as the tail entity to obtain an emoticon triple.

[0034] For the multiple emoji triplets obtained in the above method, they are combined with the existing common sense knowledge graph through the knowledge graph embedding model TransE to obtain an emoji-enhanced common sense knowledge graph.

[0035] Figure 3 Schematic diagram of the common sense knowledge graph enhanced by emoticons in an embodiment of the present invention.

[0036] like Figure 3 As shown, the obtained emoji triples (cake, delicious) and (cake, greasy) is connected with the cake in the existing common sense knowledge graph, thereby obtaining a knowledge graph containing the concepts of complete conversations and corresponding emoticons as relationships, namely the emoticon-enhanced common sense knowledge graph.

[0037] In this embodiment, a common sense knowledge graph-optimized emotional response device is constructed that includes an emoticon-enhanced common sense knowledge graph, thereby utilizing the emoticon-enhanced common sense knowledge graph to generate emotional responses with richer emotional expressions.

[0038] Figure 4 It is a schematic diagram of the framework of the common sense knowledge graph optimization emotion response device in an embodiment of the present invention.

[0039] like Figure 4 As shown, the common sense knowledge graph optimization emotion response device 100 in this embodiment includes a user input module 10, a dictionary graph storage module 20, an emotion context knowledge subgraph generation module 30, a response generation module 40, an emoticon selection module 50 and a display module 60.

[0040] The user input module 10 is used for the user to input the current conversation and emotion tags.

[0041] The dictionary graph storage module 20 is used to store the emotional dictionary and the above-mentioned emoticon-enhanced common sense knowledge graph. In this embodiment, the emotional dictionary is NRC_VAD.

[0042] The emotion context knowledge subgraph generation module 30 is used to generate an emotion context knowledge subgraph based on the current conversation, the emotion dictionary and the common sense knowledge graph enhanced by emoticons.

[0043] Figure 5 3 is a schematic diagram of the framework of the emotional context knowledge subgraph generation module in an embodiment of the present invention.

[0044] like Figure 5 As shown, the emotion context knowledge subgraph generation module 30 includes a candidate tuple generation submodule 301 , an emotion tuple generation submodule 302 and a knowledge subgraph generation submodule 303 .

[0045] The candidate tuple generation submodule 301 is used to retrieve, for each word in the current utterance, the triple corresponding to the head entity from the emoticon-enhanced common sense knowledge graph as the initial candidate tuple, and then use the initial triple with a confidence score greater than a confidence threshold as the candidate tuple of the word, where the confidence threshold is 0.2.

[0046] In this embodiment, in order to ensure that more knowledge expanded with emoticons can be noticed from the knowledge graph, the reply generation module 40 can generate more words related to this knowledge, and therefore a higher confidence is given to the emoticon triples in the knowledge graph.

[0047] The sentiment tuple generation submodule 302 is used to calculate the sentiment scores of the tail entity word and each sentiment word in the sentiment dictionary for the tail entity of the candidate tuple, and then sort the sentiment scores from large to small, select the sentiment words corresponding to the first k sentiment scores, and form sentiment tuples with the corresponding words in the current discourse, that is, each word constitutes k sentiment tuples.

[0048] The knowledge subgraph generation submodule 303 is used to construct an emotional context knowledge subgraph based on all candidate tuples and emotional tuples.

[0049] The reply generation module 40 includes a conditional variational autoencoder, which is used to input the emotional context knowledge subgraph, the current discourse and the emotion label into the conditional variational autoencoder to obtain an emotional reply.

[0050] In this embodiment, the encoder of the conditional variational autoencoder is Bi-GRU, and the specific training process of the conditional variational autoencoder includes:

[0051] First, a Seq2Seq model with a global attention mechanism is pre-trained, and a recognition network and a priori network are added on this basis. The embedding vector of the specified emotion category is defined as e'. Then, the response Y is maximized under the given condition C. x,e =[o x ;e'] is used to train the model using the variational lower bound of the conditional likelihood, where KL represents the KL divergence, i.e., Kullback-Leibler divergence, z is the latent variable, and p D (x|z,c) is the prior model used to sample z from the prior distribution, q R (z|x,c) is the recognition network used to approximate the posterior distribution of the latent variable z. During training, z is passed from the recognition network to the decoder to decode and generate the response p(Y|C x,e ) and is trained by the prior network until it is close to z'. Finally, the prior network and the Seq2Seq model constitute a conditional variational autoencoder, that is, the output of the prior network is used as the processing result of the current discourse and is input into the decoder together with the encoding of the emotional context knowledge subgraph and the emotion label. The decoder obtains the probability distribution of the word and then maps it into natural language according to the vocabulary, thereby obtaining a sense response.

[0052] In addition, during the training process of the conditional variational autoencoder, an emotion category classifier is introduced to optimize the parameters of the conditional variational autoencoder. The specific process is as follows:

[0053] The utterance x and the given emotion label c are input into the conditional variational autoencoder to generate the response x', which is then input into the emotion category classifier to obtain the probability of all emotion labels ranked from high to low. The parameters are adjusted according to the ranking until the given emotion label c is ranked at the top of the ranking, completing the parameter optimization.

[0054] Through the above optimization process, the conditional variational autoencoder can be more consistent with the given emotion label c in terms of emotional expression, so that the emotional expression of the emotional response generated by the trained conditional variational autoencoder is more accurate.

[0055] The emoticon selection module 50 is used to insert emoticons that match the context and emotion tags into the emotional reply based on the current speech and emotion tags.

[0056] The display module 60 is used to display emotional responses to the user.

[0057] The above is the overall workflow of the common sense knowledge graph optimized emotional response device 100. In this embodiment, the common sense knowledge graph optimized emotional response device 100, i.e., the device of the present invention, is compared with the dialogue response device built based on the existing model for response generation. The existing models include Attn-S2S, CVAE, Mojitalk, CARE and EREI.

[0058] We now select three cases from multiple comparisons for illustration. The information of the first case is shown in the following table:

[0059]

[0060]

[0061] As shown in the above table, the first line is the emotion label specified in the verification comparison, the second line is the conversation history of the verification comparison, that is, the current conversation, the third to seventh lines are the emotion replies generated by the devices constructed according to Attn-S2S, CVAE, Mojitalk, CARE and EREI respectively, the eighth line is the words in the emotion context knowledge subgraph constructed by the device of the present invention according to the emotion label and the current conversation, that is, the knowledge, and the ninth line is the emotion reply generated by the device of the present invention. It can be seen from the above table that in this case, the emotion reply generated by the common sense knowledge graph optimized emotion reply device 100 contains more relevant emotion information and rich reactions, that is, the knowledge obtained from the knowledge graph: threat and the word used to express emotions and one's own attitude: pain.

[0062] The information for the second case is shown in the following table:

[0063]

[0064]

[0065] As shown in the above table, the first line is the emotion label specified in the verification comparison, the second line is the conversation history of the verification comparison, that is, the current conversation, the third to seventh lines are the emotion replies generated by the devices constructed according to Attn-S2S, CVAE, Mojitalk, CARE and EREI respectively, the eighth line is the words in the emotion context knowledge subgraph constructed by the device of the present invention according to the emotion label and the current conversation, that is, the knowledge, and the ninth line is the emotion reply generated by the device of the present invention. It can be seen from the above table that in this case, the emotion reply generated by the common sense knowledge graph optimized emotion reply device 100 also contains more relevant emotion information and rich reactions, that is, the knowledge obtained from the knowledge graph: thief.

[0066] The information for the third case is shown in the following table:

[0067]

[0068] As shown in the table above, the first line is the emotion tag specified in the verification comparison, the second line is the conversation history for verification comparison, i.e., the current conversation, the third to seventh lines are the emotion replies generated by the devices constructed based on Attn-S2S, CVAE, Mojitalk, CARE, and EREI, respectively, the eighth line is the words in the emotion context knowledge subgraph constructed by the device of the present invention based on the emotion tag and the current conversation, i.e., the knowledge, and the ninth line is the emotion reply generated by the device of the present invention. It can be seen from the table above that in this case, the common sense knowledge graph optimized emotion reply device 100 generates a response with appropriate positive emotions by answering "fun". However, other devices do not recognize the positive emotions hidden in the conversation context, and therefore the emotion expression is not strong, and they tend to give dull and indifferent replies.

[0069] In summary, compared with other existing methods, the common sense knowledge graph optimized emotional response device 100 can better perceive contextual emotions through the dictionary graph storage module 20 and the emotional context knowledge subgraph generation module 30, and can effectively extract meaningful knowledge as background information for the conversation, and then generate knowledge responses with more emotional relevance and content diversity through the reply generation module 40. In addition, the emoticon selection module 50 automatically judges and inserts expression symbols that are consistent with the context and characteristics of the times in the reply, and uses it as an extension of the text content to simply and directly convey complex and changeable emotional information, making the reply more humorous and harmonious.

[0070] Functions and Effects of the Embodiments

[0071] According to the present embodiment, the emotion reply device is optimized by using the common sense knowledge graph incorporating emoticons. The dictionary graph storage module 20 including the emotion dictionary and the common sense knowledge graph enhanced with emoticons enables the common sense knowledge graph optimized emotion reply device 100 to have richer external knowledge with emotion relationships; the emotion context knowledge subgraph generation module 30 extracts relevant knowledge from external knowledge based on the current discourse to construct an emotion context knowledge subgraph to assist in generating responses containing emotions; the response generation module 40 including the trained conditional variational autoencoder generates an emotion response that is more in line with the emotion label based on the emotion context knowledge subgraph; the emoticon selection module 50 inserts emotion symbols into the emotion response to make the entire response more natural and vivid. In short, this method can generate emotion responses that are rich in emotion expression and accurate.

[0072] The above embodiments are preferred examples of the present invention and are not intended to limit the scope of protection of the present invention.

Claims

1. A device for optimizing emotional responses using a common sense knowledge graph incorporating emoticons, which is used to generate corresponding emotional responses based on current speech and emotional tags, characterized in that: include: Dictionary graph storage module, used to store sentiment dictionaries and emoticon-enhanced common sense knowledge graphs; An emotional context knowledge subgraph generation module, configured to generate an emotional context knowledge subgraph based on the current utterance, the emotional dictionary, and the common sense knowledge graph enhanced by the emoticons; A response generation module includes a conditional variational autoencoder, which is used to input the emotional context knowledge subgraph, the current utterance and the emotion label into the conditional variational autoencoder to obtain the emotional response. The emoticon-enhanced common sense knowledge graph is constructed based on the existing common sense knowledge graph and multiple emoticon triples. The emoji-enhanced common sense knowledge graph includes multiple triples, The triple and the emoticon triple both include a head entity, a relation and a tail entity, The emotional context knowledge subgraph generation module includes a candidate tuple generation submodule, an emotional tuple generation submodule and a knowledge subgraph generation submodule. The candidate tuple generation submodule is used to retrieve, for each word in the current utterance, the triple corresponding to the head entity from the emoticon-enhanced common sense knowledge graph as an initial candidate tuple, and then use the initial candidate tuple with a confidence score greater than a confidence threshold as the candidate tuple of the word. The sentiment tuple generation submodule is used to calculate the sentiment scores of the tail entity word and each sentiment word in the sentiment dictionary for the tail entity of the candidate tuple, and then sort the sentiment scores from large to small, select the sentiment words corresponding to the first k sentiment scores, and form a sentiment tuple with the corresponding words in the current discourse. The knowledge subgraph generation submodule is used to construct the emotion context knowledge subgraph according to all the candidate tuples and the emotion tuples.

2. The device for optimizing emotional response by utilizing a common sense knowledge graph incorporating emoticons according to claim 1, Its characteristics are: The specific steps of generating the emoticon triples according to the existing speech, the specified symbolic emotion label and the corresponding reply are as follows: Step S1, extracting a plurality of concept words from the discourse as discourse concept words; Step S2, extracting a plurality of the concept words from the reply as reply concept words; Step S3, constructing a first PPMI matrix based on all the discourse concept words and the response concept words; Step S4, selecting the discourse concept word and the reply concept word corresponding to a PPMI value greater than or equal to a preset first PPMI threshold from the first PPMI matrix as a first strongly associated pair; Step S5, constructing a second PPMI matrix based on all the first strongly associated pairs and the symbolic emotion labels; Step S6, selecting the first strongly associated pair and the symbolic emotion label corresponding to the PPMI value greater than or equal to a preset second PPMI threshold from the second PPMI matrix as the second strongly associated pair; Step S7: Use the discourse concept word in the second strong association pair as the head entity, the symbolic emotion label as the relationship, and the reply concept word as the tail entity to obtain the emoticon triple.

3. The device for optimizing emotional responses using a common sense knowledge graph incorporating emoticons according to claim 2, characterized in that: in, The first PPMI threshold is 1.

4. The device for optimizing emotional responses using a common sense knowledge graph incorporating emoticons according to claim 1, characterized in that: in, The emoticon triples are combined with the common sense knowledge graph through the knowledge graph embedding model TransE to obtain the emoticon-enhanced common sense knowledge graph.

5. The device for optimizing emotional responses using a common sense knowledge graph incorporating emoticons according to claim 1, characterized in that: in, The confidence threshold is 0.

2.

6. The device for optimizing emotional responses using a common sense knowledge graph incorporating emoticons according to claim 1, characterized in that: in, During the training process of the conditional variational autoencoder, the specific process of introducing the emotion category classifier to optimize the parameters of the conditional variational autoencoder is as follows: The words and a given emotion label Input the conditional variational autoencoder to generate the response , and then reply Input the emotion category classifier to obtain the probability of all the emotion labels from high to low, and adjust the parameters according to the ranking until the given emotion label is obtained. The one that is ranked higher in the ranking completes the optimization of the parameters.

7. The device for optimizing emotional responses using a common sense knowledge graph incorporating emoticons according to claim 1, characterized in that: It also includes an emoticon selection module for inserting emoticons that are consistent with the context and the emotional label into the emotional reply.

Images