This is an internal, incomplete preview of a proposed change to the ACL Anthology.
For efficiency reasons, we generate only three BibTeX files per volume, and the preview may be incomplete in other ways, or contain mistakes.
Do not treat this content as an official publication.
Perceptions of hate can vary greatly across cultural contexts. Hate speech (HS) datasets, however, have traditionally been developed by language. This hides potential cultural biases, as one language may be spoken in different countries home to different cultures. In this work, we evaluate cultural bias in HS datasets by leveraging two interrelated cultural proxies: language and geography. We conduct a systematic survey of HS datasets in eight languages and confirm past findings on their English-language bias, but also show that this bias has been steadily decreasing in the past few years. For three geographically-widespread languages—English, Arabic and Spanish—we then leverage geographical metadata from tweets to approximate geo-cultural contexts by pairing language and country information. We find that HS datasets for these languages exhibit a strong geo-cultural bias, largely overrepresenting a handful of countries (e.g., US and UK for English) relative to their prominence in both the broader social media population and the general population speaking these languages. Based on these findings, we formulate recommendations for the creation of future HS datasets.
Public figures receive disproportionate levels of abuse on social media, impacting their active participation in public life. Automated systems can identify abuse at scale but labelling training data is expensive and potentially harmful. So, it is desirable that systems are efficient and generalisable, handling shared and specific aspects of abuse. We explore the dynamics of cross-group text classification in order to understand how well models trained on one domain or demographic can transfer to others, with a view to building more generalisable abuse classifiers. We fine-tune language models to classify tweets targeted at public figures using our novel DoDo dataset, containing 28,000 entries with fine-grained labels, split equally across four Domain-Demographic pairs (male and female footballers and politicians). We find that (i) small amounts of diverse data are hugely beneficial to generalisation and adaptation; (ii) models transfer more easily across demographics but cross-domain models are more generalisable; (iii) some groups contribute more to generalisability than others; and (iv) dataset similarity is a signal of transferability.
Human feedback is increasingly used to steer the behaviours of Large Language Models (LLMs). However, it is unclear how to collect and incorporate feedback in a way that is efficient, effective and unbiased, especially for highly subjective human preferences and values. In this paper, we survey existing approaches for learning from human feedback, drawing on 95 papers primarily from the ACL and arXiv repositories. First, we summarise the past, pre-LLM trends for integrating human feedback into language models. Second, we give an overview of present techniques and practices, as well as the motivations for using feedback; conceptual frameworks for defining values and preferences; and how feedback is collected and from whom. Finally, we encourage a better future of feedback learning in LLMs by raising five unresolved conceptual and practical challenges.
Thousands of new news articles appear daily in outlets in different languages. Understanding which articles refer to the same story can not only improve applications like news aggregation but enable cross-linguistic analysis of media consumption and attention. However, assessing the similarity of stories in news articles is challenging due to the different dimensions in which a story might vary, e.g., two articles may have substantial textual overlap but describe similar events that happened years apart. To address this challenge, we introduce a new dataset of nearly 10,000 news article pairs spanning 18 language combinations annotated for seven dimensions of similarity as SemEval 2022 Task 8. Here, we present an overview of the task, the best performing submissions, and the frontiers and challenges for measuring multilingual news article similarity. While the participants of this SemEval task contributed very strong models, achieving up to 0.818 correlation with gold standard labels across languages, human annotators are capable of reaching higher correlations, suggesting space for further progress.
Detecting online hate is a complex task, and low-performing models have harmful consequences when used for sensitive applications such as content moderation. Emoji-based hate is an emerging challenge for automated detection. We present HatemojiCheck, a test suite of 3,930 short-form statements that allows us to evaluate performance on hateful language expressed with emoji. Using the test suite, we expose weaknesses in existing hate detection models. To address these weaknesses, we create the HatemojiBuild dataset using a human-and-model-in-the-loop approach. Models built with these 5,912 adversarial examples perform substantially better at detecting emoji-based hate, while retaining strong performance on text-only hate. Both HatemojiCheck and HatemojiBuild are made publicly available.
Annotating abusive language is expensive, logistically complex and creates a risk of psychological harm. However, most machine learning research has prioritized maximizing effectiveness (i.e., F1 or accuracy score) rather than data efficiency (i.e., minimizing the amount of data that is annotated). In this paper, we use simulated experiments over two datasets at varying percentages of abuse to demonstrate that transformers-based active learning is a promising approach to substantially raise efficiency whilst still maintaining high effectiveness, especially when abusive content is a smaller percentage of the dataset. This approach requires a fraction of labeled data to reach performance equivalent to training over the full dataset.
Manual fact-checking does not scale well to serve the needs of the internet. This issue is further compounded in non-English contexts. In this paper, we discuss claim matching as a possible solution to scale fact-checking. We define claim matching as the task of identifying pairs of textual messages containing claims that can be served with one fact-check. We construct a novel dataset of WhatsApp tipline and public group messages alongside fact-checked claims that are first annotated for containing “claim-like statements” and then matched with potentially similar items and annotated for claim matching. Our dataset contains content in high-resource (English, Hindi) and lower-resource (Bengali, Malayalam, Tamil) languages. We train our own embedding model using knowledge distillation and a high-quality “teacher” model in order to address the imbalance in embedding quality between the low- and high-resource languages in our dataset. We provide evaluations on the performance of our solution and compare with baselines and existing state-of-the-art multilingual embedding models, namely LASER and LaBSE. We demonstrate that our performance exceeds LASER and LaBSE in all settings. We release our annotated datasets, codebooks, and trained embedding model to allow for further research.
During COVID-19 concerns have heightened about the spread of aggressive and hateful language online, especially hostility directed against East Asia and East Asian people. We report on a new dataset and the creation of a machine learning classifier that categorizes social media posts from Twitter into four classes: Hostility against East Asia, Criticism of East Asia, Meta-discussions of East Asian prejudice, and a neutral class. The classifier achieves a macro-F1 score of 0.83. We then conduct an in-depth ground-up error analysis and show that the model struggles with edge cases and ambiguous content. We provide the 20,000 tweet training dataset (annotated by experienced analysts), which also contains several secondary categories and additional flags. We also provide the 40,000 original annotations (before adjudication), the full codebook, annotations for COVID-19 relevance and East Asian relevance and stance for 1,000 hashtags, and the final model.
We investigate the use of machine learning classifiers for detecting online abuse in empirical research. We show that uncalibrated classifiers (i.e. where the ‘raw’ scores are used) align poorly with human evaluations. This limits their use for understanding the dynamics, patterns and prevalence of online abuse. We examine two widely used classifiers (created by Perspective and Davidson et al.) on a dataset of tweets directed against candidates in the UK’s 2017 general election. A Bayesian approach is presented to recalibrate the raw scores from the classifiers, using probabilistic programming and newly annotated data. We argue that interpretability evaluation and recalibration is integral to the application of abusive content classifiers.
Online abusive content detection is an inherently difficult task. It has received considerable attention from academia, particularly within the computational linguistics community, and performance appears to have improved as the field has matured. However, considerable challenges and unaddressed frontiers remain, spanning technical, social and ethical dimensions. These issues constrain the performance, efficiency and generalizability of abusive content detection systems. In this article we delineate and clarify the main challenges and frontiers in the field, critically evaluate their implications and discuss potential solutions. We also highlight ways in which social scientific insights can advance research. We discuss the lack of support given to researchers working with abusive content and provide guidelines for ethical research.
Word embeddings are increasingly used for the automatic detection of semantic change; yet, a robust evaluation and systematic comparison of the choices involved has been lacking. We propose a new evaluation framework for semantic change detection and find that (i) using the whole time series is preferable over only comparing between the first and last time points; (ii) independently trained and aligned embeddings perform better than continuously trained embeddings for long time periods; and (iii) that the reference point for comparison matters. We also present an analysis of the changes detected on a large Twitter dataset spanning 5.5 years.