Abstract
Affective Computing (AC) is an emerging area of research that aims to develop intelligent computer systems that can recognize, synthesize, and respond to the various concepts of human affect. With the vast increase of textual user-generated content on social media networks, the detection of human affect from text became an imperative need. Many tasks in Natural Language Processing (NLP) are directly related to affect recognition such as sentiment analysis, opinion mining, abusive language, at-risk user detection, and also those concerning human-computer interactions such as conversational frameworks and chatbots. Subjective and affective concepts in NLP research including feelings, intentions, emotions, moods, and sentiments are used interchangeably. However, bearing in mind the differences of these affect-related terms helps for more reliable and efficient detection systems. Many traditional systems and their modern extensions employ extensive feature engineering steps for text representation including hand-crafted, lexical features, or classical static word embedding. However these models may focus on the important parts of the input text, they disregard other parts and aspects which may harm model generalization for different affective states of the different affective concepts.In order to mitigate these limitations, we introduce different models that use/extend advanced NLP deep learning models for more reliable text representation. These models use the transfer learning capabilities and are empowered with the attention mechanisms to consider all the contextual information with varied emphasis on different parts with a higher influence on the decisions. Moreover, the proposed models accord special attention to the characteristic differences of the different affective concepts. We consider the affective characteristics of the most important conscious affect-driven subjectivity concepts, precisely, the sentiment, emotion, and mood:Sentiment: We addressed the problem of sentiment analysis and proposed a deep learning model that applies transfer learning and multi-levels of self-attention layers to focus on the most important parts of the text that have a high influence on sentiments. The model is evaluated on several datasets and shows very competitive results. Furthermore, we evaluate the impact of attention mechanisms on the model's interpretability and user perceptions.Emotion: We tackle the problem of detection and classification of basic emotions in textual dialogues. We extend the basic model used for sentiment classification to model textual conversations and track the emotion over turns. We participate in the SemEval-2019 shared task on contextual emotion detection in text. The model shows very competitive results and ranked 9th out of more than 150 participants.Mood-I: However user mood can be classified into two main types - positive and negative mood, mood disturbances inflict various mental illnesses/disorders. We consider the problem of early detection of depression, anorexia, and self-harm using users' writings on Reddit. We proposed a new multi-stage architecture that models users' temporal mood variations. We participated in eRisk-2018 and eRisk-2019 tasks. The proposed models perform comparably to other contributions and ranked the 2nd out of 13 teams in eRisk-2019.Mood-II: We foster the study of the mood consequences to include the problem of suicide thoughts detection. Therefore, we propose a novel backbone-independent model that uses state-of-the-art Transformer-based models through Negative Correlation Learning (NCL) configuration. We evaluate the model on different tasks for at-risk users detection. The models achieve significant improvements over the existing state-of-the-art results reported for five out of six tasks for the different risk sources.