Leon Derczynski

Automatic language identification is a challenging problem. Discriminating between closely related languages is especially difficult. This paper presents a machine learning approach for automatic language identification for the Nordic languages, which often suffer miscategorisation by existing state-of-the-art tools. Concretely we will focus on discrimination between six Nordic languages: Danish, Swedish, Norwegian (Nynorsk), Norwegian (Bokm{\aa}l), Faroese and Icelandic.

Measuring inter-annotator agreement is important for annotation tasks, but many metrics require a fully-annotated set of data, where all annotators annotate all samples. We define Sparse Probability of Agreement, SPA, which estimates the probability of agreement when not all annotator-item-pairs are available. We show that under certain conditions, SPA is an unbiased estimator, and we provide multiple weighing schemes for handling data with various degrees of annotation.

Machine learning on sets towards sequential output is an important and ubiquitous task, with applications ranging from language modeling and meta-learning to multi-agent strategy games and power grid optimization. Combining elements of representation learning and structured prediction, its two primary challenges include obtaining a meaningful, permutation invariant set representation and subsequently utilizing this representation to output a complex target permutation. This paper provides a comprehensive introduction to the field as well as an overview of important machine learning methods tackling both of these key challenges, with a detailed qualitative comparison of selected model architectures.

Machine learning on sets towards sequential output is an important and ubiquitous task, with applications ranging from language modelling and meta-learning to multi-agent strategy games and power grid optimization. Combining elements of representation learning and structured prediction, its two primary challenges include obtaining a meaningful, permutation invariant set representation and subsequently utilizing this representation to output a complex target permutation. This paper provides a comprehensive introduction to the _eld as well as an overview of important machine learning methods tackling both of these key challenges, with a detailed qualitative comparison of selected model architectures.

Improvement in machine learning-based NLP performance are often presented with bigger models and more complex code. This presents a trade-off: better scores come at the cost of larger tools; bigger models tend to require more during training and inference time. We present multiple methods for measuring the size of a model, and for comparing this with the model's performance.In a case study over part-of-speech tagging, we then apply these techniques to taggers for eight languages and present a novel analysis identifying which taggers are size-performance optimal. Results indicate that some classical taggers place on the size-performance skyline across languages. Further, although the deep models have highest performance for multiple scores, it is often not the most complex of these that reach peak performance.

Data-driven and machine learning based approaches for detecting, categorising and measuring abusive content such as hate speech and harassment have gained traction due to their scalability, robustness and increasingly high performance. Making effective detection systems for abusive content relies on having the right training datasets, reflecting a widely accepted mantra in computer science: Garbage In, Garbage Out. However, creating training datasets which are large, varied, theoretically-informed and that minimize biases is difficult, laborious and requires deep expertise. This paper systematically reviews 63 publicly available training datasets which have been created to train abusive language classifiers. It also reports on creation of a dedicated website for cataloguing abusive language datahatespeechdata.com. We discuss the challenges and opportunities of open science in this field, and argue that although more dataset sharing would bring many benefits it also poses social and ethical risks which need careful consideration. Finally, we provide evidence-based recommendations for practitioners creating new abusive content training datasets.

We present the results and main findings of SemEval-2020 Task 12 on Multilingual Offensive Language Identification in Social Media (OffensEval 2020). The task involves three subtasks corresponding to the hierarchical taxonomy of the OLID schema (Zampieri et al., 2019a) from OffensEval 2019. The task featured five languages: English, Arabic, Danish, Greek, and Turkish for Subtask A. In addition, English also featured Subtasks B and C. OffensEval 2020 was one of the most popular tasks at SemEval-2020 attracting a large number of participants across all subtasks and also across all languages. A total of 528 teams signed up to participate in the task, 145 teams submitted systems during the evaluation period, and 70 submitted system description papers.

The power that machine learning models consume when making predictions can be affected by a model's architecture. This paper presents various estimates of power consumption for a range of different activation functions, a core factor in neural network model architecture design. Substantial differences in hardware performance exist between activation functions. This difference informs how power consumption in machine learning models can be reduced.

Misinformation spread presents a technological and social threat to society. With the advance of AI-based language models, automatically generated texts have become difficult to identify and easy to create at scale. We present "The Rumour Mill", a playful art piece, designed as a commentary on the spread of rumours and automatically-generated misinformation. The mill is a tabletop interactive machine, which invites a user to experience the process of creating believable text by interacting with different tangible controls on the mill. The user manipulates visible parameters to adjust the genre and type of an automatically generated text rumour. The Rumour Mill is a physical demonstration of the state of current technology and its ability to generate and manipulate natural language text, and of the act of starting and spreading rumours.

We aimed to investigate whether daily fluctuations in mental health-relevant Twitter posts are associated with daily fluctuations in mental health crisis episodes. We conducted a primary and replicated time-series analysis of retrospectively collected data from Twitter and two London mental healthcare providers. Daily numbers of ‘crisis episodes’ were defined as incident inpatient, home treatment team and crisis house referrals between 2010 and 2014. Higher volumes of depression and schizophrenia tweets were associated with higher numbers of same-day crisis episodes for both sites. After adjusting for temporal trends, seven-day lagged analyses showed significant positive associations on day 1, changing to negative associations by day 4 and reverting to positive associations by day 7. There was a 15% increase in crisis episodes on days with above-median schizophrenia-related Twitter posts. A temporal association was thus found between Twitter-wide mental health-related social media content and crisis episodes in mental healthcare replicated across two services. Seven-day associations are consistent with both precipitating and longer-term risk associations. Sizes of effects were large enough to have potential local and national relevance and further research is needed to evaluate how services might better anticipate times of higher risk and identify the most vulnerable groups.