A Semantics-based Approach to Disclosure Classification in User-Generated Online Content

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 3490–3499 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 3490 A Semantics-based Approach to Disclosure Classification in User-Generated Online Content Chandan Akiti, Anna Squicciarini, Sarah Rajtmajer Pennylvania State University {cra5302, acs20, smr48}@psu.edu Abstract As users engage in public discourse, the rate of voluntarily disclosed personal information has seen a steep increase. So-called self-disclosure can result in a number of privacy concerns. Users are often unaware of the sheer amount of personal information they share across on- line forums, commentaries, and social net- works, as well as the power of modern AI to synthesize and gain insights from this data. This paper presents an approach to detect emo- tional and informational self-disclosure in nat- ural language.We hypothesize that identifying frame semantics can meaningfully support this task. Specifically, we use Semantic Role La- beling to identify the lexical units and their se- mantic roles that signal self-disclosure. Exper- imental results on Reddit data show the perfor- mance gain of our method when compared to standard text classification methods based on BiLSTM, and BERT. In addition to improved performance, our approach provides insights into the drivers of disclosure behaviors. 1 Introduction With the growth of social networking sites and increased user engagement with public discourse online, heightened rates of user disclosure of personal information (henceforth, self-disclosure) have raised privacy and security concerns. Prior research (Keep et al., 2012) suggests that self- disclosure may be more common online thanks to the relative anonymity afforded in this environment and the lack of non-verbal cues to signal thoughts or feelings. Users are often unaware of the aggre- gate amount of personal information they share, as well as the power of modern AI to synthesize and gain insights from this data. Automating the process of identification and classification of private information in text is chal- lenging (Abril et al., 2011). A large volume of textual data needs to be processed, and a number Figure 1: SRL of a sentence containing emotional dis- closure. of real-time requirements need to be met (Agerri et al., 2015), (Singh and Nene, 2013), and signifi- cant ambiguities arise from nuanced use of natural language. In this work, we adopt the existing framework of Semantic Role Labeling to support self-disclosure identification and classification. Semantic role la- beling (SRL) is a process which aims to recog- nize all predicate-argument pairs along with their roles in a given sentence and its predicates (usually verbs). SRL is a task with numerous applications to Natural Language Processing (NLP) like Question- Answering (Abujabal et al., 2017), Information Ex- traction (Christensen et al., 2011), Machine Trans- lation (Xiong et al., 2012), text-to-scene genera- tion (Coyne et al., 2012), dialog systems (Chen et al., 2013) and social-network extraction (Agar- wal et al., 2014). We hypothesize that the inclusion of semantic frames can provide valuable context for the detection of self-disclosure. Our code is available here1. Self-disclosure in social media can take two non- exclusive forms: emotional disclosure, in …

Content truncated to 500 words