COLING 2022 Tutorial

A large body of NLP research has pointed to the need of developing systems that are socially aware and culturally sensitive. Two such broad fields include

• Development of socially-aware dialog systems or assistants, including both a generic system like SARA (Matsuyama et al., 2016)or a goal-oriented system (such as (Wanget al., 2020))

• Development of systems for automatic recognition of hateful and abusive speech online.

While socially aware dialogue systems need to recognise and generate language aimed at collaboration, cooperation and generating an environment of trust and positivity, the hateful language detection system needs to work at the other extreme. In spite of these seemingly related goals, the two fields have progressed relatively independent of each other. On the other hand, the fields of pragmatics and im/politeness research have provided extensive theoretical and empirical information to understand the nature and interrelationship between distinct but related phenomena including politeness, facework, impoliteness, offensiveness, aggression, etc., much of which has escaped the attention of NLP researchers. It has resulted in a divide between NLP and pragmatics. Researchers working with automatic generation and analysis of politeness have remained mostly confined to the theory proposed by Brown and Levinson (1987) (see for example, (Danescu-Niculescu-Mizilet al., 2013; Wang et al., 2020) and others). At the same time, an extensive amount of empirical and theoretical research in the field of im/politeness have argued and stressed upon several issues with Brown and Levinson’s theory and its inability to adequately explain and describe how (im)politeness is generated and evaluated in human communication (see for example (Eelen, 2001), (Watts, 2003),(Culpeper, 2011), (Kadar and Haugh, 2013) and a humongous amount of research published across various manuscripts and journals dedicated to the field). These studies have shown dynamic, contextual and interactional aspects of im/politeness production and evaluation in communication and stressed on both the culture-specific, contextual properties as well as the normative, ritualistic, conventionalised features of these aspects of human communication. These newer theoretical models of im/politeness have not only exposed the inadequacies and insufficiencies of the earlier models but have also lead to a better understanding of the phenomena, in general, and ultimately sorting out some of the most puzzling facets of im/politeness.However, lack of awareness about these models within the NLP community have led to a situation where NLP researchers are either struggling with understanding and explaining what they encounter in their research or duplication of research efforts where they arrive at conclusions similar to what was demonstrated decades ago in the sociopragmatic im/politeness literature (for example, see(Gupta et al., 2007) for conclusions and results that have been rather commonly known among im/politeness researchers).

One witnesses a similar divide in the recently "trending" research field of identification of hateful, offensive and abusive language. Verbal aggression and abuse have become a prevalent feature of digital communication on almost all kinds of social media platforms. Considering that aggression can manifest itself in practically infinite forms, it is difficult even for humans to recognise every form of online aggression, not mentioning automated systems. This problem is further exacerbated by a high level of sparsity in the field of Natural Language Processing in terms of taxonomies being used, consistency and explainability of these taxonomies being used, consistency and explainability of these taxonomies themselves and lack of an understanding of their interrelatedness. Even a cursory glance at multiple manifestations of online language aggression being annotated across datasets results in at least 5 or 6 different types of phenomena being addressed, such as ‘abuse’, ‘aggression’, ‘offense’, ‘toxicity’, ‘hate speech’. Furthermore, one may annotate online data by relying on non-pragmatic categories, such as ‘misogyny’, ‘Islamophobia’, ‘homophobia’, etc. And here we have not even mentioned the micro-level of annotation, i.e. how various aspects of aggression in a particular utterance can be and have been annotated.

The problemacy of annotation implies that it is challenging to develop interoperable and replicable datasets through which aggression and hate can be detected in online data. As mentioned earlier, a major obstacle in such work has been that most NLP researchers have remained unaware of the significant body of pragmatic research on (im)politeness and verbal aggression. Instead of relying on (im)politeness theory, many NLP experts have used folk formulations to study online aggression. A major issue with using such folk concepts is that they are often incompatible with each other even when one attempts to annotate different datasets in parallel. That is, without making an interdisciplinary effort - by bringing together NLP and (im)politeness research - it is very difficult to study pragmatic phenomena such as politeness, impoliteness, aggression and hateful speech in a rigorous and replicable way.

Given this and the huge interest of the computational linguistics community in the field, in order to give a definite direction to the research in the field, it is of utmost significance that the community becomes aware of the interdisciplinary studies, especially advances made within the field of sociopragmatics and pragmalinguistics of im/politeness and aggression. In this introductory tutorial we aim to fill this knowledge gap, by introducing the audience to the different models theoretical models of im/politeness, starting from the earlier models by Brown & Levinson (1987) and Leech (1983) to more recent ones like the interactional models proposed by Terkourafi (2001). In the second half of the tutorial, we will present the pragmalinguistic concept of RFID as an interdisciplinary approach to the study of socially and culturally sensitive, pragmatic phenomena such as im/politeness and aggression. This framework is mainly aimed at conducting large-scale corpus-based analysis of these phenomena. This framework may also be used for building datasets that could prove to be more fruitful in designing more transparent and better-performing systems through which one can detect aggression and hate in online settings.