EPISTEMOLOGICAL BIAS AS A MEANS FOR THE AU-TOMATED DETECTION OF INJUSTICES IN TEXT

Abstract

Injustice occurs when someone experiences unfair treatment or their rights are violated. The automated identification of injustice in text has received little attention, due in part to the fact that underlying stereotypes are rarely explicitly stated and that instances often occur unconsciously due to the pervasive nature of prejudice in society. Here, we leverage the combined use of a fine-tuned BERT-based bias detection model, two stereotype detection models, and a lexicon-based approach to show that epistemological biases (i.e., words, which presupposes, entails, asserts, hedges, or boosts text to erode or assert a person's capacity as a knower) can assist with the automatic detection of injustice in text. The news media has many instances of injustice (i.e. discriminatory narratives), thus it is our use case here.

1. INTRODUCTION

The most basic duty of the media is knowledge sharing. Yet, the tool necessary for wide-spread knowledge sharing is influence. With this influence, the media is able to shape how one will understand the intricacy of a story-line in a news story with little effort. This often results in the use of epistemological biases which involves propositions that are presupposed, entailed, asserted, hedged, or boosted in text (Recasens et al., 2013) to erode or assert a person's capacity as a knower, leading to framing issues and injustice within the text. Particularly, some of these word choices lead to testimonial injustice and what we define as character injustice. Character assassination is "the deliberate destruction of an individual's reputation or credibility" (Icks et al., 2019) . This often leads to character injustice, which is an unjustified attack on a person's character that results in an unfair criticism or inaccurate representation of them. Testimonial injustice occurs when modified believability is assigned to the statement of a subject based on widely known stereotypes (Fricker, 2007) . Both of these types of injustices can lead to affirming or perpetuating stereotypes concerning the subject. Though it has always been a harsh reality with various ringing consequences, in recent years we have publicly witnessed how the affirmation of stereotypes can lead to physical violence, prejudice, and negative self-image [(Harrison & Esqueda, 1999 ), (Gover et al., 2020 ), (Kuykendall, 1989) ]. These experiences are harmful and dangerous, explicitly for the victims but for all members of our society. We also consider framing injustice. According to (Entman, 2007) framing bias happens when the use of subjective, one-sided words that reveals the stance of an author occurs. Which means that an individual's choice from a set of options is influenced more by how the information is worded, rather than by the information itself. We recognise that news content can be positively or negatively framed to influence the narratives. In this paper, we aim to show how negative cases, when it affects particular subjects or individuals, can lead to framing injustice occurring toward those subjects. In this work, we seek to make room for subjects of text, even in creative writing, to not have their credibility shot or character assumed due to well-known stereotypes which are harmful and unfounded. Our proposed framework will be used to detect character, testimonial, and framing injustices. The framework includes a fine-tuned BERT model based on work from (Pryzant et al., 2020b) to automatically tag words associated with epistemological bias from an input text, use the (Kwon & Gopalan, 2021) CO-STAR model and (Sap et al., 2020) Social-Bias Frames to find stereotypes and the concepts of those stereotypes associated with the input text, and show when the tagged words (associated with some epistemological bias) or less credibility of a person are correlated with a stereotype which causes injustice. Though we could use examples from various fields (e.g. politics, marketing, medicine, etc...), we will use news media as a use-case throughout this work. Thus, we present the following contributions: 1. We develop a novel framework that uses the results of 3 models to detect character, testimonial, and framing injustices in News Media. 2. We produce a fine-tuned tagger model to automatically detect epistemological bias. 3. We develop a User-Interface for journalists and editors to submit text to and receive output and explanations surrounding the tagged word, referred to in this work as tagger-UI. 4. We produce empirical evidence showing how epistemological bias can translate to injustices. A goal of this work is to give journalists and editors a tool which will help them easily and quickly learn to avoid character, testimonial, and framing injustices in their work. This will be accomplished by showing users which words they use that produce epistemological bias, showing them the potential stereotype associated with the tagged words and text, offering the user explainability with the help of the stereotype concepts as defined by (Kwon & Gopalan, 2021), and offering the user resources to reference literature on the particular epistemological bias type(s) identified in their input text.

2. BACKGROUND

Many works have established it is difficult for the common person to identify a biased word in a sentence and establish the need for computational agents to take on this charge. Section 2.1, 4.1, and 6.1 and Table 8 of (Pryzant et al., 2020b) shows humans have low ability to detect bias and show humans perform worse than their detection model. Recasens et al. (2013) shows in Table 4 that the accuracy of Humans annotators on AMT (amazon mechanical turk) was not more than 37.39% for a single detected biased word. The difficulty arises due to us holding our own biases as facts and lack of education on sentence construction. However, there are specific word choices which are epistemologically biased and can lead to injustices occurring. We will focus on these word choices throughout this work. Great efforts have been put towards identifying potential words in text materials that could encourage epistemological bias [(Recasens et al., 2013) , (Hube & Fetahu, 2019) , (Pryzant et al., 2020b) ]. Following on from the work of (Pryzant et al., 2020b) we fine-tuned their tagger model to automatically tag words associated with epistemological bias from an input text. Identifying words which cause epistemological bias is a step towards awareness of social harms. This begs the question, what do we do with our new found knowledge and awareness? What kinds of implications does our use of these words impact society? What communities are affected by these word choices? These are the questions we explore in this work. Authors Kwon & Gopalan (2021) have trained a model to detect widely-known stereotypes and the concept of those stereotypes in text materials. We leverage the results of their CO-STAR model and the Social Bias Frames model (Sap et al., 2020) to offer some explainability of the word choices by the model. Associating a particular text with a stereotype and the concept of that stereotype is a critical step towards awareness of social harms that might cause character injustice to a particular individual or group. Lack of identifying the words in a sentence which imply and promote these stereotypes leaves us with the undirected burden and question of: how can we address these harms? This will be further discussed in the methods section of the paper. Beach et al. ( 2021) identify words in text that cause testimonial injustice in medical records of Black patients. Detecting such testimonial injustice is helpful in seeing the unjust realities of our society. They conclude the testimonial injustice that persists in these medical records has a high potential of causing disparity in the quality of health care for Black patients, which correlates with findings that Black patients receive a worse quality of health care (Odonkor et al., 2021) . Identifying testimonial injustice in text materials is vital to creating an environment of accountability. For accountability to be applied we must include education, which is often unexplored or left up to the user. Many users do not know where to find resources for such things, thus we provide them in our framework. Raza et al. ( 2022) developed a pipeline which takes in news articles, detects and masks words that are biased, and suggests words with more neutral text. Whilst the pipeline and library designed by the authors are very good and useful, they however do not consider the linguistic and epistemological bias features as discussed in (Recasens et al., 2013) and used by (Pryzant et al., 2020b) . Unlike

