Contextual Representation of Self-Disclosure and Supportiveness in Short Text Chandan Reddy Akiti 1 , Sarah Rajtmajer 1 , and Anna Squicciarini 1 Pennsylvania State University, University Park, PA 16804, USA Abstract. As user engagement in online public discourse continues to richen, voluntary disclosure of personal information and its associated risks to privacy and security are of increasing concern. Users are often unaware of the sheer amount of personal information they share across online forums, commentaries, and social networks, as well as the power of modern AI to synthesize and gain insights from this data. We develop a novel multi-modal approach for the joint classification of self-disclosure and supportiveness in short text. We take an ensemble approach for representation learning, leveraging BERT, LSTM, and CNN neural net- works. 1 Introduction As public discourse facilitated through social media and online forums grows increasingly commonplace, voluntary disclosure of personal information has been normalized. But users are often unaware or under-aware about the threat of self-disclosure to their privacy and security. We argue that public self-disclosure is often encouraged and even primed by a false sense of intimacy, as well as topics and tone of conversations. Accordingly, we aim to leverage contextual representations afforded by deep neural language models for the detection of self-disclosure and supportive text. The application of deep learning to NLP is made possible by representing words as vectors in a low-dimensional continuous space. Traditionally, these word representations were static. Each word was represented by a single vec- tor, regardless of the context. However, this approach had fundamental deficits for tasks like sentiment analysis where the representation of a word in context is critically important. Instead, recent work, e.g., [1], has shown that contextual word representations increase performance on NLP tasks. Deep neural language models such as BERT [2] and GPT-2 [3] represent suc- cessful attempts to create contextualized word representations. Replacing static with contextualized representations has led to significant improvement in a num- ber of NLP tasks [2]. In this work, we use the BERT pre-trained model from the huggingface [4] library. Following, we present our approach and results for the CL-Aff Shared Task on the OffMyChest conversation dataset put forth in the AAAI-2020 workshop on Affective Content Analysis. The task involves multi-class classification of Copyright 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). In: N. Chhaya, K. Jaidka, J. Healey, L. H. Ungar, A. Sinha (eds.): Proceedings of the 3rd Workshop of Affective Content Analysis, New York, USA, 07- FEB-2020, published at http://ceur-ws.org