You Shall Know a User by the Company It Keeps: Dynamic Representations for Social Media Users in NLP

Abstract

Information about individuals can help to better understand what they say, particularly in social media where texts are short. Current approaches to modelling social media users pay attention to their social connections, but exploit this information in a static way, treating all connections uniformly. This ignores the fact, well known in sociolinguistics, that an individual may be part of several communities which are not equally relevant in all communicative situations. We present a model based on Graph Attention Networks that captures this observation. It dynamically explores the social graph of a user, computes a user representation given the most relevant connections for a target task, and combines it with linguistic information to make a prediction. We apply our model to three different tasks, evaluate it against alternative models, and analyse the results extensively, showing that it significantly outperforms other current methods.

Words are the Window to the Soul: Language-based User Representations for Fake News Detection

“…While initial work focused uniquely on the textual content of the news (Mihalcea and Strapparava, 2009), subsequent research has considered also the social context in which news are consumed, characterizing, in particular, the users who spread news in social media. In line with the results reported in other classification tasks of user-generated texts (Del Tredici et al, 2019;Pan and Ding, 2019), several studies show that leveraging user representations, together with news' ones, leads to improvements in fake news detection. In these studies, user representations are usually computed using informative but costly features, such as manually assigned credibility scores (Kirilin and Strube, 2018).…”

“…To define the social graph we follow a common approach in the literature (Yang and Eisenstein, 2017;Del Tredici et al, 2019) and create, for each dataset, a graph G = (V, E) in which V is the set of users in the dataset, and E is the set of edges between them. An unweighted and undirected edge is instantiated between two users if one retweets the other.…”

Multilingual stance detection in social media political debates

“…It is also worth mentioning that several systems which participated to the SemEval-2016 Task 6 competition involved the use of deep learning techniques, such as Wei et al (2016) who used a Convolutional Neural Network (CNN) combined with a voting scheme based on the concept of "divide and conquer", and Zarrella and Marsh (2016) who exploited a Recurrent Neural Network (RNN) with four layers containing 128 Long Short Term Memory (LSTM) units. In the last years, after the end of the contest, the dataset released for the SemEval-2016 Task 6 has been considered as a benchmark and therefore exploited to carry on research regarding SD in English tweets by several research groups (Augenstein et al, 2016;Dey et al, 2018;Wei et al, 2018;Zhou et al, 2019;Del Tredici et al, 2019). Among them, let us focus on the ones where a score on the specific stance targets addressed in this paper (Hillary Clinton or Donald Trump) is reported: Augenstein et al (2016) who proposed a neural approach based on bidirectional conditional encoding, Dey et al (2018) who implemented a two-phase LSTM using attention, Wei et al (2018) who explored the performances of a biderectional Long Short-Term Memory neural network (biLSTM), and Zhou et al (2019) who used a condensed CNN with attention over self-attention.…”

Knowledge Enhanced Masked Language Model for Stance Detection

“…More specifically, BERT uses Masked Language Modeling (MLM) to pre-train a transformer encoder by predicting masked tokens in order to learn the semantic representation of a corpus. Ghosh and colleagues (Ghosh et al, 2019) show that the original pre-trained BERT without any further finetuning outperforms other former state-of-the-art models on the SemEval set including the model that utilizes both text and user information (Del Tredici et al, 2019). Because we are interested in the 2020 US Presidential election and many temporal factors relevant to stance exist (e.g.…”