The communities that we live in affect our health in ways that are complex and hard to define. Moreover, our understanding of the place-based processes affecting health and inequalities is limited. This undermines the development of robust policy interventions to improve local health and well-being. News media provides social and community information that may be useful in health studies. Here we propose a methodology for characterising neighbourhoods by using local news articles. More specifically, we show how we can use Natural Language Processing (NLP) to unlock further information about neighbourhoods by analysing, geoparsing and clustering news articles. Our work is novel because we combine street-level geoparsing tailored to the locality with clustering of full news articles, enabling a more detailed examination of neighbourhood characteristics. We evaluate our outputs and show via a confluence of evidence, both from a qualitative and a quantitative perspective, that the themes we extract from news articles are sensible and reflect many characteristics of the real world. This is significant because it allows us to better understand the effects of neighbourhoods on health. Our findings on neighbourhood characterisation using news data will support a new generation of place-based research which examines a wider set of spatial processes and how they affect health, enabling new epidemiological research. • Novel methodology for characterising neighbourhoods based on local news. • Natural Language Processing can create meaningful neighbourhood characterisations from local news articles. • Extensive quantitative and qualitative evaluation show analysis is sound.

Sigmoid output layers are widely used in multi-label classification (MLC) tasks, in which multiple labels can be assigned to any input. In many practical MLC tasks, the number of possible labels is in the thousands, often exceeding the number of input features and resulting in a low-rank output layer. In multi-class classification, it is known that such a low-rank output layer is a bottleneck that can result in unargmaxable classes: classes which cannot be predicted for any input. In this paper, we show that for MLC tasks, the analogous sigmoid bottleneck results in exponentially many unargmaxable label combinations. We explain how to detect these unargmaxable outputs and demonstrate their presence in three widely used MLC datasets. We then show that they can be prevented in practice by introducing a Discrete Fourier Transform (DFT) output layer, which guarantees that all sparse label combinations with up to k active labels are argmaxable. Our DFT layer trains faster and is more parameter efficient, matching the F1@k score of a sigmoid layer while using up to 50% fewer trainable parameters. Our code is publicly available at https://github.com/andreasgrv/sigmoid-bottleneck.

Classifiers in natural language processing (NLP) often have a large number of output classes.For example, neural language models (LMs) and machine translation (MT) models both predict tokens from a vocabulary of thousands.The Softmax output layer of these models typically receives as input a dense feature representation, which has much lower dimensionality than the output.In theory, the result is some words may be impossible to be predicted via argmax, irrespective of input features, and empirically, there is evidence this happens in small language models (Demeter et al., 2020).In this paper we ask whether it can happen in practical large language models and translation models.To do so, we develop algorithms to detect such unargmaxable tokens in public models.We find that 13 out of 150 models do indeed have such tokens; however, they are very infrequent and unlikely to impact model quality.We release our algorithms and code so that others can test their models.1

Abstract Using natural language processing, it is possible to extract structured information from raw text in the electronic health record (EHR) at reasonably high accuracy. However, the accurate distinction between negated and non-negated mentions of clinical terms remains a challenge. EHR text includes cases where diseases are stated not to be present or only hypothesised, meaning a disease can be mentioned in a report when it is not being reported as present. This makes tasks such as document classification and summarisation more difficult. We have developed the rule-based EdIE-R-Neg, part of an existing text mining pipeline called EdIE-R (Edinburgh Information Extraction for Radiology reports), developed to process brain imaging reports, ( https://www.ltg.ed.ac.uk/software/edie-r/ ) and two machine learning approaches; one using a bidirectional long short-term memory network and another using a feedforward neural network. These were developed on data from the Edinburgh Stroke Study (ESS) and tested on data from routine reports from NHS Tayside (Tayside). Both datasets consist of written reports from medical scans. These models are compared with two existing rule-based models: pyConText (Harkema et al. 2009. Journal of Biomedical Informatics42(5), 839–851), a python implementation of a generalisation of NegEx, and NegBio (Peng et al. 2017. NegBio: A high-performance tool for negation and uncertainty detection in radiology reports. arXiv e-prints, p. arXiv:1712.05898 ), which identifies negation scopes through patterns applied to a syntactic representation of the sentence. On both the test set of the dataset from which our models were developed, as well as the largely similar Tayside test set, the neural network models and our custom-built rule-based system outperformed the existing methods. EdIE-R-Neg scored highest on F1 score, particularly on the test set of the Tayside dataset, from which no development data were used in these experiments, showing the power of custom-built rule-based systems for negation detection on datasets of this size. The performance gap of the machine learning models to EdIE-R-Neg on the Tayside test set was reduced through adding development Tayside data into the ESS training set, demonstrating the adaptability of the neural network models.

When parsing morphologically-rich languages with neural models, it is beneficial to model input at the character level, and it has been claimed that this is because character-level models learn morphology.We test these claims by comparing character-level models to an oracle with access to explicit morphological analysis on twelve languages with varying morphological typologies.Our results highlight many strengths of character-level models, but also show that they are poor at disambiguating some words, particularly in the face of case syncretism.We then demonstrate that explicitly modeling morphological case improves our best model, showing that characterlevel models can benefit from targeted forms of explicit morphological modeling.