In the digital information age, detecting and mitigating linguistic biases, particularly in political discourse, presents a critical challenge. This study addresses this issue by developing a comprehensive pipeline for annotating and analyzing a dataset of news articles. The process involved expert-driven rule creation for identifying bias indicators, followed by annotation using a large language model (LLM). The integration of a Named Entity Recognition (NER) model is tailored to identify and categorize biases within the annotated dataset. The study's novelty lies in its blend of human expertise with LLM, enhancing the accuracy and relevance of bias detection. We conducted a comparative analysis of several language models. This research contributes significantly to both the theoretical and practical domains of bias detection in texts. The annotated dataset and the best model are made available for use and application.
CEUR ceur-ws.org
In recent decades, the domain of news reporting and content generation has undergone a
profound transformation [
The emergence of advanced artificial intelligence (AI) and Large Language Models (LLMs) [
The integration of AI in journalism has enhanced news production’s eficiency and scope
but also introduced complex challenges, particularly in bias detection. The term bias’ has
diverse definitions [
This paper addresses the challenge of detecting biases within political news available on web. We compiled an extensive dataset from diverse sources (CNN, BBC, Global etc.) via Google news feeds, consisting of articles selected over five months (May-October 2023). Post-collection, we conducted comprehensive preprocessing to structure and annotate the data for bias detection in both binary label and multi-class token classification tasks. This work focuses particularly on the multi-class token classification task and the methods used for annotating this data. The main contributions of this work are as follows:
Annotation Framework: We introduce an annotation framework wherein a team of expert
annotators first developed rules to identify indicators of bias in news content. These indicators
were then translated into structured prompts for analysis using OpenAI’s advanced LLM [
Named Entity Recognition (NER) Model Development: Following the identification of biased terms, we introduce a multi-label token classification task and develop a Named Entity Recognition (NER) model specifically tailored to this annotated dataset. This model is designed to focus on the multi-label classification of biased terms found in news articles. To establish benchmarks, we evaluated various state-of-the-art models for this purpose.
This work ofers a novel methodology for understanding the biases identified in political news content. This research is very timely for comprehending data construction and annotation, as well as establishing a baseline for bias detection. It also presents potential avenues for future work directed towards mitigating biases.
In its broadest definition, bias can be defined as a tendency or preference towards or against
certain groups or individuals [
A diverse collection of datasets has been assembled for various bias detection NLP tasks. For example, Toxigen [17] leverages GPT-3 with 274K prompts for implicit hate speech classification, while Stereotype [18] gathers 17K Wikidata entries for quantifying biases based on gender, profession, and race. HolisticBias [19] , MBIB [20], RedditBIAS [21], BIGNEWS [22] are among valuable resources for advancing research in the areas of bias detection, and linguistic analysis.
Research studies have underscored the inadvertent introduction of societal and cultural
biases into training data [
Our study introduces a novel approach, inspired by prior research, where we employ humandesigned rules to guide a LLM in annotating bias within news articles. We then use this data as a baseline for multi-label bias detection methods.
We propose a data annotation and model training pipeline in this work, shown in Figure 1. The initial stage consists of a data collection phase, where data is scraped from news websites and processed into a dataframe based on inclusion/exclusion criteria. The second stage consists of annotation process, which involves human-written rules, prompt construction, and the use of a LLM API for generating annotations. The third stage is the data quality stage, where experts assess the annotations, reach a consensus, and finalize the tags. The final stage is the data compliance and model building phase, the where annotated data is converted into a specific format (e.g., CoNLL), mapped, saved, and used to train a bias recognition . Next, we explain each phase of the pipeline.
Our dataset is focussed on racial slurs and biases in political speeches in North America. The data is curated using Google RSS by using ‘feedparser‘ and ‘newspaper.Article‘ API to extract article text, publication dates, news outlets, and URLs. These articles represent a wide range of political news sources, such as ’The Daily Chronicle’, ’Global Times News’, ’Liberty Voice’, ’New Age Politics’, and ’The Public Perspective’. The data collection strategy is as: • Time Frame: Five months from May to October 2023. • Inclusion/Exclusion Criteria: Relevant, accessible articles included; incomplete, non
English, redundant articles excluded. • Keywords: Political, economic, social, cultural, environmental, scientific, global afairs, historical, and emerging social issues. • Total Articles Scraped: 40,000. • Articles for Annotation: 4,000, covering a broad range of topics. • Annotation Method: Initial annotations with GPT-Turbo-3.5 API, followed by expert verification. • Basic Fields: Text, biased-words (n-gram words can be tokens, phrases or a pattern in a sentence), bias-label.
Rule Development : The initial phase of our dataset construction involves a collaborative efort from five experts across computer science, social science, political science, and media analysis. These experts developed a comprehensive set of rules to identify various forms of political bias in news narratives. The experts focussed on linguistic patterns and narrative structures indicative of biases like ideological leanings, emotive language, and selective reporting. An example of a rule is as: • Rule: Identification of Emotive Language • News Text: “The government’s disastrous policies have resulted in economic chaos.” • Annotation: Words such as “disastrous” and “chaos” are tagged as emotionally charged, which signifies a negative bias against the government’s policies.
Prompt Construction : Following the rule development, these guidelines are synthesized
into a single, comprehensive prompt designed for the OpenAI [
We obtained the dataframe with the text column, bias label and ‘bias-words’ identified in this step.
We implemented a human verification system to validate the annotations produced by the OpenAI model. This verification process was carried out over a two-month period by 5 experts and 2 students to ensure the uniformity and reliability of all annotations. Where necessary, adjustments were made based on a consensus among the team. To measure the success of this collaboration between our human experts and the AI model, we adopted an Inter-Rater Reliability Score. Currently, this metric stands at approximately 78%, which indicates a high level of concordance between human and AI evaluations.
We transformed our annotated dataset with biased words into the CONLL (Conference on Natural Language Learning) [37] format to align with standard NLP practices with FAIR (Findability, Accessibility, Interoperability, and Reusability) principles [38]. This data format involves structuring each word from our ‘text’ data as a separate entity, with corresponding ‘biased_words’ annotations, in a tabular layout for a sequence labeling task. Below is an example of a politically biased sentence formatted in the CONLL style, identifying specific biased words: • Word: The, government’s, reckless, policies, harm, the, economy; • Bias Tag: O, O, B-BIAS, I-BIAS, B-BIAS, I-BIAS, I-BIAS.
In this example, “B-BIAS” marks the beginning of a biased phrase, “I-BIAS” continues the phrase, and “O” indicates words that are not part of a biased phrase.
Annotated Dataset: The annotated dataset consists of 4,000 records, comprising 50.8k tokens in total, with 20% of the dataset specifically reserved for testing. The annotated dataset is available here 1.
We develop a multi-label token classification model, an Named Entity Recognition (NER) system, which is designed to identify linguistic biases in textual content. The proposed model innovates beyond traditional NER approaches by incorporating features sensitive to the manifestation of bias. It aims to show not only the presence of bias but also its scope and boundary within text.
The foundation of the model is a transformer-based neural network, employing the BERT [39] architecture for contextualized token embeddings. The classification layer of the model 1https://huggingface.co/datasets/newsmediabias/FAKE-NEWS-BIASES-LABELLED is a softmax layer, fine-tuned to distinguish between three classes: B-BIAS for the beginning of a biased entity, I-BIAS for tokens inside a biased entity, and O for tokens outside of biased entities. The classification of each token is computed as: bias = softmax ( ⋅
BERT + ) where bias signifies the probability distribution over the three classes for each token, is the weight matrix, is the bias vector, and BERT denotes the embeddings from BERT. The corpus used for training has been annotated with these bias-specific entity tags, as elaborated in Section 3.2. To optimize the parameters of our model, we employ the cross-entropy loss function: = −
∑ ⋅ log( ̂ ) ∈{, ,} where is the ground truth label, and ̂ is the predicted probability for each class. The summation runs over the classes (B-BIAS), (I-BIAS), and (Outside).
We present a comparative analysis of various LMs on the task of detecting linguistic biases within our annotated dataset. Each model is evaluated based on its ability to accurately classify the pre-defined bias entity categories: B-BIAS, I-BIAS, and O (non-biased).
In our study, we compare several baseline models, each fine-tuned on our annotated dataset for bias detection. We utilize BERT ‘bert-base-uncased‘ as a foundational model for deep bidirectional training, serving as our primary baseline. Additionally, we include DistilBERT ‘distilbert-base-uncased‘, a lighter version of BERT that ofers a balance between performance and eficiency. ALBERT ‘albert-base-v2‘ is employed for its parameter-reduction techniques, which enhance eficiency and highlight the trade-ofs in bias detection. Furthermore, we useRoBERTa ‘roberta-base-uncased‘, a BERT-alike model but trained on a larger corpus, to examine its efectiveness in bias detection [ 40]. We also use Llama-7b-chat [41] by Meta, which is a generative AI LLM and we tested its bias recognition ability with few-shot learning, where a limited number of examples, or shots, are provided to perform a new instance of the task.
The dataset, detailed in Section 3, was partitioned into three distinct subsets: training (80), validation (10), and testing (10). We use precision, recall, and the F1-score to evaluate the model’s performance on the bias entity recognition task.
Our experiments were conducted using Google Colab Pro, with GPUs such as the NVIDIA Tesla P100 and A100 (at times), along with a cloud-based environment. The software stack consisted of Python 3.9, PyTorch 1.7.1 as the primary deep learning framework, and the Hugging Face Transformers library version 4.36 for the implementation of transformer-based models. (1) (2)
Hyperparameter tuning played a vital role in optimizing the performance of each baseline model. For BERT, DistilBERT , and RoBERTa, we chose a learning rate of 2e-5, a batch size of 16 (smaller to save memory cycles), and a training duration of 4 epochs. For ALBERT, we opted for a learning rate of 1e-5, a smaller batch size of 16, and a similar epoch count of 4. Hyperparameters were specifically adjusted to suit the few-shot learning requirements of Llama 2, aiming to maximize its eficacy in learning from a limited number of examples. All models were fine-tuned on the same annotated dataset (except Llama 2), to ensure a basis for comparison across the diferent approaches.
The comparative analysis of baseline models, as detailed in Table 1. The results provides insightful findings regarding the efectiveness in the task of bias detection.
RoBERTa (Base-Uncased) showed best performance in our analysis, achieving the highest scores across precision (0.90), recall (0.89), and F1-score (0.89). This high performance can be attributed to RoBERTa’s enhanced training and larger model architecture. BERT (Base-Uncased) and ALBERT (Base-v2) display competitive results, with BERT marginally outperforming in precision and F1-score. This suggests that while ALBERT’s design optimizes memory eficiency, but it may slightly reduce efectiveness in complex tasks like bias detection.
DistilBERT (Base-Uncased) and Llama 2 with Few-Shot Prompts show decent, although at relatively lower performance. The reduced complexity of DistilBERT, designed for eficiency, likely impacts its ability to fully understand the intricacies of bias. Meanwhile, Llama 2’s performance suggests that while few-shot learning is good rapid adaptation, but it might efect the bias detection tasks.
Overall, the results highlight the efectiveness of comprehensive training and model complexity in bias detection. These results also pointing to the nuanced trade-ofs between model size, eficiency, and task-specific performance.
To provide a deeper understanding of each model’s ability in bias detection, we present comparing instances, in Table 2, where each model successfully detected bias against instances where it “This overtly discriminatory act has been condemned.“ - Successfully identified bias. “The policy blatantly favors the wealthy, ignoring others.“ - Correctly identified bias. “A clear case of nepotism has been observed in the selection process.“ - Accurately identified bias.
Missed Bias Detection “The policy has received mixed reactions from various groups.“ - Missed implicit bias. “Critics argue the decision could have unforeseen consequences.“ - Overlooked slight bias. “There are concerns about the efectiveness of the new regulation.“ - Failed. “Some experts are questioning the necessity of this policy.“ - Missed nuanced bias. “The project’s relevance is being debated in academic circles.“ - Failed to detect implicit bias. failed. For example, we observe that models like BERT and RoBERTa efectively identify explicit biases but sometimes miss more nuanced ones. DistilBERT and ALBERT tend to overlook subtle biases, which hints a need for refinement in detecting less apparent bias instances. The performance varies greatly with the context of the bias, as seen in Llama 2’s accurate detection in cases of ’nepotism’ but could not perform good in abstract scenarios.
This research presents the development and implementation of a comprehensive data annotation and model building pipeline. A key outcome of this research has been the successful annotation of a dataset comprising 4,000 records, alongside the identification of the most efective model for bias detection. We make both the annotated data and best model weights (RoBERTa) publicly available for further research and application. This study also acknowledges the dynamic nature of language, highlighted by the phenomenon of data drifts [ 27, 42]. This underscores the need for ongoing monitoring and updating of our data annotation protocols. The limited size of the dataset, at 4k records, poses potential limitations on the depth and diversity of biases that can be captured. Therefore, expanding the dataset is crucial for enhancing the robustness and generalizability of our models. The Llama 2 model [41], with its focus on few-shot learning, demonstrated promising results. However, to optimize its performance, we recommend finetuning it with more explicit instructions. This research lays a foundation for data annotation and model building in the evolving landscape of language and bias.
[17] T. Hartvigsen, S. Gabriel, H. Palangi, M. Sap, D. Ray, E. Kamar, ToxiGen: A Large-Scale Machine-Generated Dataset for Adversarial and Implicit Hate Speech Detection, in: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Association for Computational Linguistics, Dublin, Ireland, 2022, pp. 3309–3326. URL: https://aclanthology.org/2022.acl-long.234. doi:10.18653/v1/2022. acl-long.234. [18] M. Nadeem, A. Bethke, S. Reddy, StereoSet: Measuring stereotypical bias in pretrained language models, 2020. URL: http://arxiv.org/abs/2004.09456, arXiv:2004.09456 [cs]. [19] E. M. Smith, M. Hall, M. Kambadur, E. Presani, A. Williams, “I’m sorry to hear that”: Finding New Biases in Language Models with a Holistic Descriptor Dataset, in: Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguistics, Abu Dhabi, United Arab Emirates, 2022, pp. 9180–9211. URL: https://aclanthology.org/2022.emnlp-main.625. doi:10.18653/v1/2022. emnlp-main.625. [20] M. Wessel, T. Horych, T. Ruas, A. Aizawa, B. Gipp, T. Spinde, Introducing mbib-the first media bias identification benchmark task and dataset collection, in: Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2023, pp. 2765–2774. [21] S. Barikeri, A. Lauscher, I. Vulić, G. Glavaš, RedditBias: A Real-World Resource for Bias Evaluation and Debiasing of Conversational Language Models, in: C. Zong, F. Xia, W. Li, R. Navigli (Eds.), Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), Association for Computational Linguistics, Online, 2021, pp. 1941–1955. URL: https://aclanthology.org/2021.acl-long.151. doi:10.18653/v1/2021.acl-long.151. [22] Y. Liu, X. F. Zhang, D. Wegsman, N. Beauchamp, L. Wang, Politics: pretraining with same-story article comparison for ideology prediction and stance detection, arXiv preprint arXiv:2205.00619 (2022). [23] A. Caliskan, J. J. Bryson, A. Narayanan, Semantics derived automatically from language corpora contain human-like biases, Science 356 (2017) 183–186. [24] T. Spinde, M. Plank, J.-D. Krieger, T. Ruas, B. Gipp, A. Aizawa, Neural media bias detection using distant supervision with babe-bias annotations by experts, in: Findings of the Association for Computational Linguistics: EMNLP 2021, 2021, pp. 1166–1177. [25] M. Färber, V. Burkard, A. Jatowt, S. Lim, A multidimensional dataset based on crowdsourcing for analyzing and detecting news bias, in: Proceedings of the 29th ACM International Conference on Information & Knowledge Management, 2020, pp. 3007–3014. [26] T. Bolukbasi, K.-W. Chang, J. Y. Zou, V. Saligrama, A. T. Kalai, Man is to computer programmer as woman is to homemaker? debiasing word embeddings, Advances in neural information processing systems 29 (2016). [27] S. Raza, C. Ding, Fake news detection based on news content and social contexts: a transformer-based approach, International Journal of Data Science and Analytics 13 (2022) 335–362. [28] S. Raza, D. J. Reji, C. Ding, Dbias: detecting biases and ensuring fairness in news articles, International Journal of Data Science and Analytics (2022). doi:10.1007/ s41060-022-00359-4. [29] Y. Cai, A. Zimek, G. Wunder, E. Ntoutsi, Power of explanations: Towards automatic debiasing in hate speech detection, in: 2022 IEEE 9th International Conference on Data Science and Advanced Analytics (DSAA), IEEE, 2022, pp. 1–10. [30] Y. Wang, J. Mansurov, P. Ivanov, J. Su, A. Shelmanov, A. Tsvigun, C. Whitehouse, O. M.
Afzal, T. Mahmoud, A. F. Aji, et al., M4: Multi-generator, multi-domain, and multi-lingual black-box machine-generated text detection, arXiv preprint arXiv:2305.14902 (2023). [31] L. Cheng, S. Ge, H. Liu, Toward understanding bias correlations for mitigation in nlp, arXiv preprint arXiv:2205.12391 (2022). [32] V. S. Govindarajan, K. Atwell, B. Sinno, M. Alikhani, D. Beaver, J. J. Li, How people talk about each other: Modeling generalized intergroup bias and emotion, in: Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, 2023, pp. 2488–2498. [33] L. Bauer, K. Gopalakrishnan, S. Gella, Y. Liu, M. Bansal, D. Hakkani-Tur, Analyzing the limits of self-supervision in handling bias in language, arXiv preprint arXiv:2112.08637 (2021). [34] F. Zhou, Y. Mao, L. Yu, Y. Yang, T. Zhong, Causal-debias: Unifying debiasing in pretrained language models and fine-tuning via causal invariant learning, in: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2023, pp. 4227–4241. [35] Y. Lei, R. Huang, L. Wang, N. Beauchamp, Sentence-level media bias analysis informed by discourse structures, in: Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguistics, Abu Dhabi, United Arab Emirates, 2022, pp. 10040–10050. URL: https://aclanthology.org/2022.emnlp-main.682. doi:10.18653/v1/2022.emnlp-main.682. [36] S. Fu, D. Q. Calley, V. A. Rasmussen, M. D. Hamilton, C. K. Lee, A. Kalla, H. Liu, Genderbased language diferences in letters of recommendation, AMIA Summits on Translational Science Proceedings 2023 (2023) 196. [37] E. F. Sang, F. De Meulder, Introduction to the conll-2003 shared task: Language-independent named entity recognition, arXiv preprint cs/0306050 (2003). [38] M. D. Wilkinson, M. Dumontier, I. J. Aalbersberg, G. Appleton, M. Axton, A. Baak, N. Blomberg, J.-W. Boiten, L. B. da Silva Santos, P. E. Bourne, et al., The fair guiding principles for scientific data management and stewardship, Scientific data 3 (2016) 1–9. [39] J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, Bert: Pre-training of deep bidirectional transformers for language understanding, arXiv preprint arXiv:1810.04805 (2018). [40] Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, V. Stoyanov, Roberta: A robustly optimized bert pretraining approach, arXiv preprint arXiv:1907.11692 (2019). [41] H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y. Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosale, et al., Llama 2: Open foundation and fine-tuned chat models, arXiv preprint arXiv:2307.09288 (2023). [42] S. Raza, B. Schwartz, Entity and relation extraction from clinical case reports of COVID-19: a natural language processing approach, BMC Medical Informatics and Decision Making 23 (2023) 20. doi:10.1186/s12911-023-02117-3.