Warning: This paper contains examples of language and images which may be offensive. Misogyny is a form of hate against women and has been spreading exponentially through the Web, especially on social media platforms. Hateful content towards women can be conveyed not only by text but also using visual and/or audio sources or their combination, highlighting the necessity to address it from a multimodal perspective. One of the predominant forms of multimodal content against women is represented by memes, which are images characterized by pictorial content with an overlaying text introduced a posteriori. Its main aim is originally to be funny and/or ironic, making misogyny recognition in memes even more challenging. In this paper, we investigated 4 unimodal and 3 multimodal approaches to determine which source of information contributes more to the detection of misogynous memes. Moreover, a bias estimation technique is proposed to identify specific elements that compose a meme that could lead to unfair models, together with a bias mitigation strategy based on Bayesian Optimization. The proposed method is able to push the prediction probabilities towards the correct class for up to 61.43% of the cases. Finally, we identified the most challenging archetypes of memes that are still far to be properly recognized, highlighting the most relevant open research directions.

In this paper, we investigate the impact of optimal hyper-parameter configuration in relational topic models. The main goal is to validate the hypothesis that single-objective Bayesian Optimization (BO) can discover a hyper-parameter setting that leads a set of relational topic models to simultaneously ensure good prediction capabilities and significant topics from a qualitative perspective. Our research, as a result of a comparative analysis performed on 7 state-of-the-art models, 5 performance measures and 3 datasets, has highlighted three main findings: (1) the majority of relational topic models are not able to offer a good trade-off between classification capabilities and topic interpretability; (2) single-objective optimization of hyper-parameters, targeted on maximizing the F1-Measure, is able to create topics that are also optimal with respect to the Kullback Leibler divergence measure; (3) the Pareto frontiers across several performance metrics reveals that the most promising trade-off between the performance metrics can be obtained by Constrained Relational Topic Models.

During the last years, the phenomenon of hate against women increased exponentially especially in online environments such as microblogs. Although this alarming phenomenon has triggered many studies both from computational linguistic and machine learning points of view, less effort has been spent to analyze if those misogyny detection models are affected by an unintended bias. This can lead the models to associate unreasonably high misogynous scores to a non-misogynous text only because it contains certain terms, called identity terms. This work is the first attempt to address the problem of measuring and mitigating unintended bias in machine learning models trained for the misogyny detection task. We propose a novel synthetic test set that can be used as evaluation framework for measuring the unintended bias and different mitigation strategies specific for this task. Moreover, we provide a misogyny detection model that demonstrate to obtain the best classification performance in the state-of-the-art. Experimental results on recently introduced bias metrics confirm the ability of the bias mitigation treatment to reduce the unintended bias of the proposed misogyny detection model.

The Social Web represents nowadays the principal means to support and foster social interactions among people through Web 2.0 technologies. Individuals interact in virtual communities to pursue mutual interests or goals, by exchanging multiple kinds of contents (i.e., textual, acoustic, visual), the so-called User-Generated Content (UGC). In this context, the SeCredISData Special Session is especially devoted at discussing the implications that the analysis of big social data has in tackling open issues related to society from different perspectives. On one side, there is the need to push forward the research on emotion and sentiment, and the investigation of affective cognitive models and their possible integration into intelligent systems. On the other side, it is urgent to address the issue of on-line information credibility assessment, in an era where trusted intermediaries have disappeared and people must rely only on their cognitive capacities to judge information. The Special Session is therefore aimed at promoting the development of models and applications able to tackle these issues.

The automatic detection of sarcasm and irony in user generated contents is one of the most challenging task of Natural Language Processing. In this paper we address this problem by introducing Bayesian Model Averaging (BMA), an ensemble approach to take into account several classifiers according to their reliabilities and their marginal probability predictions. The impact of the most used expressive signals (pragmatic particles and POS tags) have been evaluated in baseline models (traditional classifiers and majority voting) as well as in the proposed BMA approach. Experimental results highlight two main findings: (1) not all the features are equally able to characterize sarcasm and irony and (2) BMA not only outperforms traditional state of the art models, but is also able to ensure notable generalization capabilities both on ironic and sarcastic text.