top of page

Lesson 4: Bias in Computer Linguistics

Definition: 

Research on bias is an important sub-field of computational linguistic research which analyses the bias in machine learning, its own techniques, as well as in digital language online.

Key Concepts

  • Bias in Machine Learning

    • Historical Bias

    • Representation Bias

    • Measurement Bias

    • Aggregation Bias

    • Evaluation Bias

  • Biased Language

    • Framing Bias

    • Epistemiological Bias

UNIT 1: BIASES IN COMPUTATIONAL LINGUISTIC PROCESSES

Bias could be a new word for you, it refers to an inclination/prejudice/distortion towards a certain group/idea. Biases in machine learning, the development of computer systems that learn from data, can be divided into five different categories: historical, representation, measurement, aggregation and evaluation bias. These categories are not mutually exclusive and language models are often affected by a combination of them. They are illustrated in Figure 1 below, which gives you an overview of the sources of the biases and the steps in which they enter into the system.

Bildschirmfoto 2023-12-10 um 17.26.22.png

Historical Bias

Everyday analogy: If you base your prediction of an issue on tendencies you have seen in the past, even though they might not reflect the contemporary situation.

Historical bias, as depicted in Figure 1 (1), is a concern in data-driven models where existing biases in the data are reflected in the model's outputs. It often occurs when historical data used for training already contains biases, which can lead the model to perpetuate stereotypes related to specific identity groups.

For example, in online translation, if historical data includes more instances of male doctors, the model is likely to translate "the doctor" into the masculine form "der Arzt" in German.

Representation Bias

Everyday analogy: If you only speak with your friends about politics, you might have difficulties in understanding other people's positions.

 Representation bias (2) arises when training data does not accurately represent the population, causing over- or underrepresentation of specific groups.

For instance, if a speech-to-text system is primarily trained on audiobooks narrated by white, middle-aged men, it may struggle to transcribe diverse speakers. Similarly, the absence of Swiss German in language options for speech recognition tools exemplifies representation bias.

Measurement Bias

Everyday analogy: If you use an old map to navigate, you might not be able to satisfyingly reach the destination, as the layout of the place might have changed.

Measurement bias (3) arises when training data features and labels do not align with what users are looking for or when data quality and labelling are inconsistent.

For example, using archaic Bible text for training translation models can lead to online translators producing results that do not match the style of contemporary language usage.

Aggregation Bias

Everyday analogy: If you use the same strategy for different games, even though they have different rules and challenges, it will not work out well for all of them.

Aggregation bias (4) is concerned with the model output and occurs when a single model cannot effectively serve diverse data groups.

An example of aggregation bias can come up in sentiment analysis, which aims to automatically determine the emotional tone or sentiment expressed in a text by using natural language processing. Bias in this context occurs when combining data from multiple languages without accounting for language-specific nuances, resulting in inaccurate emotional tone determinations due to cultural and linguistic differences.

Evaluation Bias

Everyday analogy: If you evaluate your soccer skills based on how many times you can juggle the ball with your feet, not bearing in mind that there are many other skills needed to be a successful player.

Evaluation bias (5) happens when the success of a model is tested on datasets that are not sufficiently representative of the real-world scenarios the model is intended to address.

An example of evaluation bias is evident in voice recognition models of conversational artificial intelligence models, such as Siri or Alexa. As they are primarily tested on standard accents, their performance for non-standard language is not tested adequately, which potentially leads to underperformance for regional dialect speakers.

Activity 1: Discussion

Your contact with bias

You can now discuss the following question in pairs  (3 minutes):

  • 1) Can you think of other examples for these types of bias?

Now, you can pair up with a different partner, summarise what you have discussed previously and address the question below (7 minutes):

  • 2) Why are these biases concerning?

Did you finish the exercise? 1) OTHER EXAMPLES FOR THESE TYPES OF BIAS Historical Bias - E.g. chatbots: could generate stereotypical/sexist/racist answers if already the training data embodied certain prejudices. Representation Bias - E.g. machine translation models: slang words could not be identified if the training data of model consisted mainly of standard language. Measurement Bias - E.g. language models: inconsistent annotation due to workers’ different contexts and backgrounds. Aggregation Bias - E.g. language learning software: could not work equally effective for all languages, if they fail to adjust for language-specific teaching methods and challenges. Evaluation Bias - E.g. chatbots: might not perform well if asked to replicate colloquial language, if they were evaluated on standard language (as their weakness was not detected). 2) REASONS WHY THESE BIASES ARE CONCERNING - Reproduction of stereotypes - Compromise of fairness and equity, bringing about of discrimination - Negative impact mainly on (already) vulnerable populations - Disapproval of and loss of trust in computational tools and technology in general - Inaccuracy and incorrectness of models’ output (that could remain unnoticed and lead to ill-advised decisions) - Innovational challenges (handling of bias is difficult and can make the model training very complex and time-consuming) - Legal challenges (tools could violate laws on privacy or discrimination)

UNIT 2: COMPUTATIONAL LINGUISTIC APPROACHES TO PREVENT BIASES

As we have seen in the first part of the module, the field of computational linguistics faces the challenge of reducing the biases in its practical applications. However, the discipline also offers promising approaches to detect and prevent bias. Linguistic theory can provide computational tools with knowledge of how biased language is realised in text or speech and how these biases can be removed.

Biased language is especially undesirable in reference sources, such as textbooks or encyclopaedia articles, through which we inform ourselves and acquire new knowledge. Therefore, the researchers Recasens, Danescu-Niculescu-Mizil and Jufarsky (2013) aimed to create a program, a linguistically-informed model, that can automatically detect biased language in text, applying it to Wikipedia. This research exemplifies how computational linguistics techniques are applied to real-world language-related challenges.

 

The research project

The first step of this project was to find biased sentences and then to identify the linguistic cues, such as words or phrases, that cause the bias. These cues can be used as indicators of bias. How did they go about that? As you probably know, people can make edits to Wikipedia articles and label why they corrected them. Thus, the authors searched the edits that were made to remove bias and analysed the changes linguistically. They found that biased language could be categorised into two main types; framing bias and epistemological bias.

Framing bias is characterised by the use of subjective or one-sided words or phrases. 

Everyday analogy: If you apply a filter to a photo, you can make it look more beautiful or more dramatic.

  • He is a perfectly/strictly organised colleague. (subjective intensifiers)

  • The eco-activists/eco-terrorists gathered in Berne.​ (one-sided terms)

Epistemological bias is a bit more subtle and is realised in words or phrases that make the assumption that something is true/false with various degrees of certainty, often via propositions.

Everyday analogy: If you hear a rumour, you might believe it or be sceptical, which leads you to speak of it in the respective manner.

  • Can you perceive how the word in italics affects the meaning of the sentence?

  • The results show/suggest that young people are interested in foreign languages. (factive verbs)

  • The politician stated/claimed that the new law will improve the conditions. (assertive verbs)

  • The kid gave in to/did not resist the temptations and ate the candies. (entailments)

  • She will probably not come back. (hedges)

Based on these findings, the researchers trained their computer program to recognise biased language by feeding the machine lists of the extracted biasing words. They further indicated the features of these expressions, such as to which the above-mentioned linguistic cues the word belongs, its position in a sentence or the grammatical relation.

When testing the model on new sentences, the program achieved 34.35% accuracy in identifying the most biased word. In the last step, the researchers compared the performance of the bias detector against human participants. It stood out that humans performed only slightly better at 37.39%, highlighting the difficulty of detecting bias for both humans and computer programs.  Automatic bias detectors can thus be a useful tool to editors of reference works, helping them to find biases that they did not notice themselves and also making their work more efficient.

Why is this research a relevant example of computational linguistics in action?

This research demonstrates how computational linguistics addresses real-world language bias, utilizing a for this field typical interdisciplinary approach that combines linguistic theories, data analysis, and machine learning. It explores various biases, shedding light on the complexity of language analysis and showcasing the field's challenges and opportunities.

Activity 2: Text analysis
Can you find the biases?

Now, it's your turn. Can you identify all instances of framing bias and epistemological bias in the following text?

(10 minutes)

Tipp: For both categories, note down the biased phrases, identify the words that introduce the bias and indicate the line number to facilitate the self-correction.  

Bildschirmfoto 2023-12-10 um 17.43.20.png

Did you finish the exercise? FRAMING BIAS: Positive framing bias: - The text positively frames computational linguistics, using words like "valuable" (line 1), "exciting" (l. 2), and "fascinating" (l. 21) to describe the field which introduces positive framing bias. - Moreover, there is an optimistic framing with phrases like "immense convenience”(l. 4), "boundless opportunities"(l. 12) and “tremendous potential”(l. 21). The words "immense", "boundless", “tremendous” introduce this positive framing bias. - There is a positive bias by suggesting that computational linguistics can "revolutionise healthcare" (l. 11) and "improve customer service" (l. 11). The words "revolutionise" and "improve" introduce this bias. - Furthermore, the phrase “attracting brilliant minds who eagerly explore its frontiers” (l. 18-19) is also positively biased. The expression “brilliant minds” is one-sided and portrays the academics in a favourable manner. Negative framing bias: - On the other hand, the text introduces negative framing bias by emphasising that computational linguistics might lead to “severe bias in language models” (l. 16), "the fatal erosion of privacy" (l. 16) and “threatening digital divide” (l. 16-17). The words “severe”, "fatal" and “threatening” in this context introduce negative framing bias. EPISTEMOLOGICAL BIAS: Epistemological bias presupposing truth/certainty: - The text argues that computational linguistics “proves to hold a central role in shaping the future of language processing” (l. 1-2), which presupposes the truth of this statement. The word “prove” introduces the bias. - Another instance of this bias can be seen in the phrase “computational linguistics is undoubtedly a fascinating field” (l. 21), where the word “undoubtedly” introduces the bias. Epistemological bias presupposing falsehood/uncertainty: - There is a bias that presupposes a certain falsehood and uncertainty in the phrases “[a]dvocates of this discipline argue that it opens up [...] possibilities” (l. 2-3) and “sceptics argue that while it offers certain advantages” (l. 8). The word “argue” introduces the bias. -Also in the phrase “[t]hey suggest that it can revolutionise healthcare” (l.11), there is an epistemological bias. It stems from the word “suggest”, which points out the uncertainty of the statement. -Moreover, there is also epistemological bias in the phrase “which could probably diminish our own linguistic capabilities” (l. 9), introduced by “probably”.

Final thought for this lesson

As we have seen, bias is a complex matter and it is hard to exclude in machine learning and difficult to detect in texts – for us humans but also for computational linguistic tools.

When is the next time, you will encounter bias in these contexts?

Sources used

TEXTS

Çıtak, E. (2023, July 27). Bias in Machine Learning: Concepts, Causes, and How to Fix It. Dataconomy. https://dataconomy.com/2023/07/27/how-to-fix-bias-in-machine-learning/

 

CrashCourse (Director). (2021, January 15). Computational Linguistics: Crash Course Linguistics #15. https://www.youtube.com/watch?v=3npuPXvA_g8

John-Mathews, J.-M. (2021, November 9). Measurement bias in Machine Learning. Giskard. https://www.giskard.ai/knowledge/where-do-biases-in-ml-come-from-3-n-measurement

Leidner, J. L., & Plachouras, V. (2017). Ethical by Design: Ethics Best Practices for Natural Language Processing. Proceedings of the First ACL Workshop on Ethics in Natural Language Processing, 30–40. https://doi.org/10.18653/v1/W17-1604

Recasens, M., Danescu-Niculescu-Mizil, C., & Jurafsky, D. (2013). Linguistic Models for Analyzing and Detecting Biased Language. Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, 1, 1650–1659.

 

Shee, E. (2021, October 12). 6 Types of AI Bias Everyone Should Know. Seldon. https://www.seldon.io/6-types-of-ai-bias

 

SuperAnnotate (2022, March 17). Bias in machine learning: Types and examples. https://www.superannotate.com/blog/bias-in-machine-learning

Suresh, H., & Guttag, J. V. (2021). A Framework for Understanding Sources of Harm throughout the Machine Learning Life Cycle. Equity and Access in Algorithms, Mechanisms, and Optimization, 1–9. https://doi.org/10.1145/3465416.3483305

 

TELUS International (2021, February 4). Seven Types Of Data Bias In Machine Learning. TELUS International. https://www.telusinternational.com/insights/ai-data/article/7-types-of-data-bias-in-machine-learning

 

Towards AI (2023, March 28). A Guide to Computational Linguistics and Conversational AI. Towards AI. https://towardsai.net/p/machine-learning/a-guide-to-computational-linguistics-and-conversational-ai

ILLUSTRATIONS

Dialani, P. (2019, January 10). Challenges Associated with AI Bias. Analytics Insight. https://www.analyticsinsight.net/challenges-associated-with-bias-ai/

 

Samuels, K. (2022, May 4). Women in AI: Breaking the Internet Glass Ceiling. Black Women Talk Tech. https://www.blackwomentalktech.com/women-in-ai

bottom of page