Modeling Multi-level Context for Informational Bias Detection by Contrastive Learning and Sentential Graph Network

Abstract

Informational bias is widely present in news articles. It refers to providing one-sided, selective or suggestive information of specific aspects of certain entity to guide a specific interpretation, thereby biasing the reader's opinion. Sentence-level informational bias detection is a very challenging task in a way that such bias can only be revealed together with the context, examples include collecting information from various sources or analyzing the entire article in combination with the background. In this paper, we integrate three levels of context to detect the sentence-level informational bias in English news articles: adjacent sentences, whole article, and articles from other news outlets describing the same event. Our model, MultiCTX (Multi-level ConTeXt), uses contrastive learning and sentence graphs together with Graph Attention Network (GAT) to encode these three degrees of context at different stages by tactically composing contrastive triplets and constructing sentence graphs within events. Our experiments proved that contrastive learning together with sentence graphs effectively incorporates context in varying degrees and significantly outperforms the current SOTA model sentencewise in informational bias detection.

Proceedings of the Sixth Fact Extraction and VERification Workshop (FEVER)

“…Another study by Lee et al (2021) introduce a general purpose misinformation UnifiedM2 model for bias detection in BASIL and handle tasks like fake news, clickbait and rumors. Contrastive learning and Graph Attention Network in the Mul-tiCTX (Multi-level ConTeXt) model uses triplets of BASIL to detect informational bias as proposed by Guo and Zhu (2022b). Similarly, bias sentence identification is also studied by Lei et al (2022) through local and global discourse structures using RoBERTa for addressing bias.…”

“…In order to shed light into this issue, we begin by organizing prior work and defining the notation we will utilize in the rest of the manuscript, for the sake of simplicity. We refer to sentences annotated with no bias as 'NEU', sentences annotated with 2019) is denoted as 'Fan', van den Berg and Markert (2020) as 'Berg', (Lee et al, 2021) as 'Lee', Guo and Zhu (2022b) as 'Guo', and Lei et al (2022) as 'Lei'. informational bias as 'INF', sentences annotated with lexical bias as 'LEX'. Additionally, we refer to the combination of neutral sentences with and sentences with lexical bias as 'OTH', while 'BIAS' refers to the combination of samples that have both informational and lexical bias.…”

“…Using the above definitions, we proceeded to analyze the settings proposed by previous work by Fan et al (2019), van den Berg and Markert (2020), Lee et al (2021), Guo and Zhu (2022b), and Lei et al (2022) which we summarize in Table 2, above. We can see that there is a clear disparity in the way in which the bias detection tasks are approached.…”

“…In this sense, their proposal was to differentiate informational bias from lexical bias. On one hand, the former is usually presented using speculative, imperative, and tangential clauses that convey information in a factual and neutral way (van den Berg and Markert, 2020) to sway readers' opinions towards news entities (Guo and Zhu, 2022a), therefore depending mostly on the context. On the other hand, lexical bias instead depends on linguistic attributes like word choice and syntax and originates from content realization or how things or events are expressed, and is generally not depending on the context (Greene and Resnik, 2009;Hube and Fetahu, 2019;Iyyer et al, 2014;Yano et al, 2010;Recasens et al, 2013).…”

“…Though the development of datasets such as BASIL has brought an interesting new paradigm to look at bias in news, thereby drawing the attention of several members of the research community and leading to abundant prior work, we note that these studies mostly focus on informational bias only (van den Berg and Markert, 2020; Guo and Zhu, 2022b;Fan et al, 2019;Lei et al, 2022), with lexical bias studied solely by Fan et al (2019). While we surmise that this could be partially due to the fact that lexical bias appears comparatively less frequently in news articles, making the automatic detection of both kinds of bias simultaneously difficult, we think this should be no reason to focus on either and argue that detecting lexical bias from informational bias should be equally significant.…”