This paper introduces topic modelling, a machine learning technique that automatically identifies ‘topics’ in a given corpus. The paper illustrates its use in the exploration of a corpus of academic English. It first offers the intuitive explanation of the underlying mechanism of topic modelling and describes the procedure for building a model, including the decisions involved in the model-building process. The paper then explores the model. A topic in topic models is characterised by a set of co-occurring words, and we will demonstrate that such topics bring us rich insights into the nature of a corpus. As exemplary tasks, this paper identifies the prominent topics in different parts of papers, investigates the chronological change of a journal, and reveals different types of papers in the journal. The paper further compares topic modelling to two more traditional techniques in corpus linguistics, semantic annotation and keywords analysis, and highlights the strengths of topic modelling. We believe that topic modelling is particularly useful in the initial exploration of a corpus.

In this study, I investigate the patterns of verbal agreement of twenty-three singular collective nouns in English which take plural of-dependents (e.g., a group of children or a bunch of people). The main aim is to explore the extent to which this constituent may influence the number of the main verb, by considering factors such as the role of overt morphological markers for number in the oblique noun and the influence of syntactic distance on the patterns of verbal agreement. The data will show strong conditioning effects of the plural of-dependent on the verb, in triggering plural verbal number, for example. Contrary to what has been claimed hitherto, syntactic distance does not result in an increase of semantic (i.e., plural) agreement in this type of construction. In fact, the non-overtly marked plural noun people, whose influence on verbal agreement is strong in local syntactic domains, shows a considerable decrease in the proportion of plural verbal forms which it takes with increasing distance, contrary to overtly marked plural nouns. Finally, in light of the grammaticalisation suffered by homologous constructions such as a lot of, I will also consider the role of verbal agreement in the possible idiomatisation of the structures that I scrutinise here.

In this paper, I argue that the study of the linguistic aspects of epistemology has become unhelpfully concentrated on the corpus-based study of hedging and that a corpus-driven approach can help to improve upon this. Through focussing on a corpus of texts from one discourse community (that of genetics) and identifying frequent tri-lexical clusters containing highly frequent lexical items that are identified as keywords, I undertake an inductive analysis identifying patterns of epistemic significance. Several of these patterns are shown to be hedging devices and the whole corpus frequencies of the most salient of these, candidate and putative, are then compared to the whole corpus frequencies for comparable wordforms and clusters of epistemic significance. Finally I interviewed a ‘friendly geneticist’ in order to check my interpretation of some of the terms used and to get an expert interpretation of the overall findings. I argue that the highly unexpected patterns of hedging found in genetics demonstrate the value of adopting a corpus-driven approach and constitute an advance in our current understanding of how to approach the relationship between language and epistemology.

Automatically time-aligning utterances at the segmental level is increasingly common practice in phonetic and sociophonetic work because of the obvious benefits it brings in allowing the efficient scaling up of the amount of speech data that can be analysed. The field is arriving at a set of recommended practices for improving alignment accuracy, but methodological differences across studies (e.g., the use of different languages and different measures of accuracy) often mean that direct comparison of the factors which facilitate or hinder alignment can be difficult. In this paper, following a review of the state of the art in automatic segmental alignment, we test the effects of a number of factors on its accuracy. Namely, we test the effects of: (1) the presence or absence of pause markers in the training data, (2) the presence of overlapping speech or other noise, (3) using training data from single or multiple speakers, (4) using different sampling rates, (5) using pre-trained acoustic models versus models trained ‘from scratch’, and (6) using different amounts of training data. For each test, we examine three different varieties of English, from New Zealand, the USA and the UK. The paper concludes with some recommendations for automatic segmental alignment in general.

In this paper, I discuss the theoretical and practical issues raised in the development of Tibidabo, a syntactically and semantically annotated corpus of Spanish, which has been developed using the DELPH-IN corpus annotation environment and the HPSG-based grammar, SRG. I describe in detail the linguistic analysis that Tibidabo provides, focussing on various syntactic phenomena that are frequently discussed in the literature and consider alternative analyses proposed in the HPSG literature and in AnCora, another Spanish Treebank. 2 2 This work was funded by the Ramón y Cajal programme of the Spanish Ministerio de Ciencia e Innovación.

Manual corpus annotation facilitates exhaustive and detailed corpus-based analyses of evaluation that would not be possible with purely automatic techniques. However, manual annotation is a complex and subjective process. Most studies adopting this approach have paid insufficient attention to the methodological challenges involved in manually annotating evaluation – especially concerning transparency, reliability and replicability. This article illustrates a procedure for annotating evaluative expressions in text that facilitates more transparent, reliable and replicable analyses. The method is demonstrated through a case study analysis of appraisal ( Martin and White, 2005 ) in a small-size specialised corpus of CEO letters published by the British energy company, BP, and four competitors before and after the Deepwater Horizon oil spill of 2010. Drawing on Fuoli and Paradis's (2014) model of trust–repair discourse, we examine how attitude and engagement resources are strategically deployed by BP's CEO in the attempt to repair stakeholders’ trust after the accident.

Threats made by stalkers have tremendous negative effects on the more than one million victims who receive them every year – from experiencing ‘psychological terrorism’ to undergoing physical harm ( Burgess and Marchetti, 2009 ). However, stalking, a criminal offence within the law, is ill-defined and difficult to prosecute, since the victim of stalking must demonstrate the stalker's intent to intimidate or cause ‘substantial emotional distress’ ( Black et al., 1990 : 717).Linguistically, such indicators of emotion and intent are manifested by markers of ‘stance’, a speaker/writer's culturally organised feelings, judgments or assessments about a recipient or proposition ( Biber et al., 1999 ). Through a corpus analysis of 397 authentic threats, I examine variation in the manifestation and function of overt grammatical markers of stance (adverbials, modals, and that- and to-complement clauses) between threats to stalk, harass and defame. Specifically, certainty verbs + that-complement clause constructions and prediction modals occurred at a significantly higher rate (p<0.001) in stalking threats – those in which the victims reported feeling intense fear or distress. In addition, strong co-occurrence patterns were found between these stance markers and pronouns. And, when using a qualitative social constructionist approach to examine distinctions in stance meaning and reveal underlying functional patterns, ( Precht, 2003 : 255), three distinctive functional patterns were revealed with the trigrams ‘I will be’, ‘I will have’ and ‘you know that’, wherein the stalking threatener is in complete volitional control of his or her own actions, demonstrates possession over the victim or an object related to the threat, and accuses the victim of a behaviour thought to be wrong, respectively.Therefore, while it has been demonstrated that there is not a one-to-one correspondence between linguistic markers and threateners' actions ( Gales, 2010 ; and Lord et al., 2008 ), through a closer examination of overt grammatical markers of stance – in combination with social psychology theories connecting speaker role, power and fear (e.g., Burgess and Marchetti, 2009 ) – linguists can help hone an understanding of stance in stalking threats and potentially contribute to the demonstration of a victim's claim of feeling fear.