Pilar Durán

Lexical Diversity and Lexical Sophistication in First Language Writing

The research reported in this chapter has several aims. First, we wished to extend the exploration of vocabulary richness, particularly lexical diversity to the medium of writing. It was important to assess the extent to which vocabulary diversity as measured by D, and word rarity continued to be valid developmental measures in older children producing samples of writing. That is to say, do these measures continue to improve in line with increasing age and developing proficiency in the written modality? In respect of D, Ruth Berman’s project ‘Developing literacy in different contexts and in different languages’ (Berman, 2000) that we referred to in Chapter 1, has already applied D to both the speech and writing of children in three age groups (9–10, 12–13, 16–17) and adults across two genres in seven languages. Results show main effects for age, language, and genre but not for modality, that is to say, speech versus writing (Berman and Verhoeven, 2002). In an analysis that focused exclusively on the results for Swedish, on the other hand, differences were found between modalities at all four ages, and an overall effect of age on writing (Stromqvist et al., 2002). However, differences between individual age groups at the older and younger ages were not significant. This study gives further support to D as a developmental measure, but it is not possible to scrutinise the results for writing within a single language. In addition, there is no single independent measure of the quality of the writing produced that can be used for validation purposes, and this is one of the strengths of the research to be described in this chapter.

Early Child Language 2: the Bristol Corpus

The previous chapter used an American corpus of language data from children aged between 27 and 32 months to carry out an initial validation of procedures used in the vocd software and to assess the criterion-related validity of the measure D. We also examined the relationship between sample size and D-values and tested the reliability of D.

Comparing the Diversity of Lexical Categories: the Type-Type Ratio and Related Measures

The aim of this chapter is to provide an empirical demonstration of the sensitivity of type-type ratios to sample size and to outline and demonstrate a solution derived from the mathematical modelling procedures described in earlier chapters. However, because researchers working in language development and education are probably less aware of the problems related to sample size for type-type ratios than they are for type-token ratios, and because of the significance for theory and practice of recent research that has used such measures, we will begin with an extensive review of the issues. This will encompass analyses of rare words or lexical sophistication, studies of vocabulary composition and early lexical style, and of the ‘noun bias’ issue.

A New Measure of Inflectional Diversity and its Application to English and Spanish Data Sets

It will be recalled that in earlier chapters we compared different versions of D that systematically varied the definition of a word type. In Chapter 4 we analysed transcripts from the 38 children in the 32-month directory of the New England Corpus (Snow, 1989; Dale et al., 1989) in the CHILDES database. We compared D calculated from inflected forms (‘go’, ‘goes’, ‘going’ as three types) with D calculated from stems (‘go’, ‘goes’, ‘going’ as one type). Clearly, inflected forms will always give higher values on any lexical diversity measure, provided that subjects are actually using inflectional morphology, and the difference between these two versions of D proved to be highly reliable in the New England data even though the two measures were correlated at a level approaching unity.

Lexical Diversity and the Investigation of Accommodation in Foreign Language Proficiency Interviews

In this chapter we continue to investigate the validity of D as a measure of lexical diversity and will demonstrate its application as a solution to a measurement problem in a new context: a study of oral language assessment procedures using data from teenagers learning a foreign language. Previous research into language proficiency interviews by Richards and Malvern (2000) found that the aspect of teachers’ language in oral interviews that adapted most to the ability of their students was lexical diversity. The analysis reported here focuses on this finding in greater depth using the new measure, D, rather than the mean segmental type-token ratio (MSTTR) that had been used in the previous analysis. We will investigate the relationship between D and other measures of foreign language proficiency and compare the Ds of students and teachers as well as computing correlations between teachers’ D and measures of the students’ proficiency to judge the extent to which the teachers finely tune their own language to the language level of individual students.

Early Child Language 1: the New England Corpus

Chapters 4 and 5 provide examples of the application of vocd to two corpora of early child language, beginning here with American transcripts obtained from the CHILDES database (MacWhinney, 2000b) and followed up in Chapter 5 with a British dataset. The analyses of the corpus reported in this chapter aimed initially to try out the software on a real data set, as opposed to artificial transcripts specially constructed for debugging, and to see whether the use of switches to remove repeated words and phrases (‘retracings’) and to strip words of regular inflections would make the predicted differences to the D-values obtained.

Overview and Conclusions

At the beginning of this book, we compared the search for a single, all-embracing, measure of the quality of vocabulary deployment with ‘a quest for the Holy Grail’. As we have shown, even for the attempt to define a valid measure of lexical diversity independent of text length, the journey, like all such quests, has been full of wrong turnings, false dawns, and confusion. That there has been confusion is in no doubt (see, for example, Malvern and Richards, 1997; Richards and Malvern, 1997), and research results have frequently been uninterpretable, or possibly wrong, because the authors are unaware of, or simply ignore, the sensitivity to sample size of simplistic measures, such as NDW, TTR, and related indices.

A Mathematical Model of Lexical Diversity

As well as being a function of text length, there is another limitation to using raw TTR for the whole of a language sample — it does not take into account the frequency with which types are repeated. As a simple illustration of the point, we can consider three imaginary texts of 40 tokens containing 20 types. In Text A, each of the 20 types is repeated twice; in Text B ten of the 20 types appear three times and the remaining ten once; Text C is more natural, with ten types occurring once, four twice, three occuring three times, two four times, and one type occurring five times. Neither raw TTR nor any of its simple transformations, such as RootTTR or LogTTR, will distinguish among these three transcripts as they all have the same overall numbers of types and of tokens producing a TTR of 0.5, but they are manifestly different in how they deploy the same vocabulary — they have different frequency distributions.

Traditional Approaches to Measuring Lexical Diversity

As we have indicated in the previous chapter, measuring lexical diversity is not as straightforward as it may seem, and to demonstrate the issues involved in its quantification we first need to consider in some detail an apparently simple approach and then discuss the most influential metric, which we shall show is fatally flawed.