Word Frequency: Use or misuse? John M. Kirk, Queen’s University Belfast, Northern Ireland. This paper will not be concerned with statistical treatments of word frequency beyond percentage distributions and relativised frequencies per thousands or millions of words. Its primary concern will be with frequency as a property of data, adopting a critical look at statements such as ‘each text comprises 2,000 words’. It will be concerned with words as tokens, types and lemmatised types; the range of functions and meanings of words; and words and lexemes. It will consider words of low frequency as well as of high frequency. In its critical section, it will ask whether word frequencies are self-explanatory or need explanation, and whether approximation is as useful as precision. It refers to a range of well-known corpora of English as well as the three corpora which I have compiled: Corpus of Dramatic Texts in Scots, the Northern Ireland Transcribed Corpus of Speech (NITCS), and the Irish component of the International Corpus of English (ICE-Ireland). Word Frequency, Statistical Stylistics, and Authorship Attribution. David Hoover, New York University, USA. The availability of large corpora and of electronic texts has renewed interest in the venerable topic of word frequency. Innovations in analytic techniques and in the ways word frequencies are selected for analysis have also been instrumental. It is these upon which I will focus here. Authorship attribution and statistical stylistics have, until recently, typically been based upon fewer than the 100 most frequent words of a corpus. These words – almost exclusively function words – are attractive because they are so frequent that they account for most of the running words of a text, and because such words have been assumed to be especially resistant to intentional manipulation by an author, so that their frequencies should reveal authorial habits that remain relatively constant across a variety of texts. Recent work on style variation has suggested that selecting words because of their frequency in sections of texts rather than in the entire corpus is more effective in capturing stylistic shifts. Removing words that are frequent overall only because they are very frequent in a single text has also been shown to dramatically improve the accuracy of an analysis. Another trend has been to increase the number of words analyzed to as many as the 6000 most frequent words, a point at which almost all the words of the text are included and almost all are content words. Finally, following the innovative work of John Burrows, there is a great deal of current interest in Delta, a new measure of the differences between texts that is based upon comparing how different the texts are from the mean for the entire corpus. Further refinements in the selection of words for analysis and in alternative formulas for calculating Delta suggest that further improvements in the accuracy may be possible and that we may be nearing a theoretical explanation of how and why word frequency analysis is able to capture authorship and style. I will be discussing these issues mainly with reference to a 2,000,000 word corpus of contemporary American poetry and a much larger corpus of 46 Victorian novels. Word frequency in Context: Alternative architectures to examine related words, register variation, and historical change. Mark Davies, Brigham Young University, USA. The use of relational databases that are composed of the frequency of n-gram in a given corpus allows users to quickly and easily examine word frequency. Perhaps the first large corpus to use such an approach was the 100 million word Corpus del Español, which was created in 2002 (www.corpusdelespanol.org/). This was followed by two BNC-based 100 million word corpora that were modelled on the same architecture: Phrases in English (pie.usna.edu) and Variation in English Words and Phrases (VIEW; view.byu.edu), as well as a 40 million word Corpus of Historical English (view.byu.edu/che) The relational database/n-grams architecture allows simple word frequency queries such as the following (all of which can be carried out on a 100 million word corpus in 1-2 seconds): • Overall frequency of a given word, set of words, phrase, or substring in the corpus • “Slot-based” queries; e.g. the most common nouns one “slot” after mysterious, or z-score ranked words immediately preceding chair • Wide-range collocates; e.g. the most common nouns within a ten word window (left or right) of string or broken In addition, however, the architecture that we have used for VIEW and the Corpus of Historical English allows several other types of queries that cannot be carried out directly with competing architectures (e.g. SARA/XARA, the IMS Corpus Workbench, or the Phrases in English architecture), including the following: • Comparison of frequency with related words; e.g. nouns occurring immediately after utter but not after complete or sheer, or adjectives within ten words of woman but not man • One simple query to find the frequency of words in separate databases, such as user-defined, customized lists (clothing, emotions, technology terms, etc) or synsets from WordNet • Register variation; e.g. all verbs or all words ending in *ble or all three-word lexical bundles that are more common in academic texts than in fiction, or in legal or medical texts • Historical variation; e.g. words, phrases, or collocates of a given word or part of speech, which are more common in the 1900s than in the 1800s Finally, even within the relational database/n-grams approach, there are competing architectures that favour certain types of queries over others, and we will briefly consider some of these issues. Issues for Historical Corpora: First catch your word. Christian Kay, University of Glasgow, Scotland. The Historical Thesaurus of English (THE; www.arts.gla.ac.uk/sesll/englang/thesaur/homepage.htm) is a semantic index to the Oxford English Dictionary (OED 1884-; OED Online 2000-) supplemented by Old English materials published separately in A Thesaurus of Old English (Roberts, Kay, Grundy, 2000). Word senses are organised in a hierarchy of categories and subcategories, with up to fourteen levels of delicacy. The material is held in a database and first steps towards Internet publication are being taken by an AHRC-ICT Strategy Project creating searches for use in a range of humanities disciplines (Smith, Horobin, Kay, n.d.). The main problem which besets searching historical texts is that of variable spelling – the further one goes back in time, the worse it gets. Similar problems affect texts in non-standard varieties, as experience of the Scottish Corpus of Texts and Speech (SCOTS; www.scottishcorpus.ac.uk/) and the Dictionary of the Scots Language (DSL; www.dsl.ac.uk/dsl/) demonstrate. Dictionary headwords lemmatize common variants but are by no means comprehensive; an alternative may be a rule-based system which predicts possibilities. Corpora have further problems in that lemmatization may not solve problems of homonymy and polysemy. The paper will suggest ways of addressing these problems using the resources described above. References The Oxford English Dictionary, 1884-1933 ed. by Murray, Sir James A. H.; Bradley, Henry; Craigie, Sir William A. and Onions, Charles T.; Supplement, 1972-1986 ed. by Burchfield, Robert W., 2nd edn, 1989, ed. by Simpson, John A. and Weiner, Edmund S. C.; Additions Series, 1993-1997, ed. by Simpson, John A.; Weiner, Edmund S. C. and Proffitt, Michael; 3rd edn (in progress) OED Online, March 2000-, ed. by Simpson, John A. (Oxford: Oxford University Press). Roberts, Jane and Kay, Christian with Grundy, Lynne, A Thesaurus of Old English, King’s College London Medieval Studies XI, 2 vols (London, 1995), Second impression (Amsterdam: Rodopi, 2000). An electronic version, supported by British Academy LRG-37362, may be seen at www.arts.gla.ac.uk/ sesll/englang/thesaur/toe1.htm. Smith, Jeremy; Horobin, Simon; and Kay, Christian, Lexical Searches for the Arts and Humanities, AR112456. In Search of a Bad Reference Corpus Mike Scott, University of Liverpool, UK. What are the tolerable limits of similarity between a reference corpus and a node text for the generation of a useful set of keywords? There is of course considerable subjectivity in the notion of usefulness, which will vary according to research goals which cannot in general be predicted with certainty. Nevertheless, the aim here is to explore the ways in which the similarity between reference corpus and node text vary on various important dimensions, such as size in tokens, similarity of texttype, similarity of historical period, similarity of subject-matter. This presentation starts from the formula proposed by Berber Sardinha (2004, 101-3) which suggests that the larger the reference corpus, the more keywords will be detected, and his formula for predicting the number of keywords produced with a given text and reference corpus. It also considers his recommendation that a reference corpus should be about five times the size of the node text. Using a series of reference corpora, the paper compares keyword results in relation to specific texts. The aim is to identify not, as one might imagine, the characteristics of the good reference corpus, but the limits defining a poor one, since in many cases, e.g. the analysis of a dead language or a restricted corpus, the chance of accessing a good re
