Can Type-Token Ratio be Used to Show Morphological Complexity of Languages?

ABSTRACTType-token ratio (TTR), also known as vocabulary size divided by text length (V/N) is a simple measure of lexical diversity. It has been used in literary studies, studies in child language and even psychiatry. The basic problem of TTR is that it is affected by the length of the text sample. Several suggestions for improving this fault have been given, including standardizing the length of text samples, using logarithms in the basic formula, etc. We show in this paper that simple TTR and its more elaborate calculation MATTR can be used for approximation of morphological complexity of languages. This usage of TTR has been notified by Juola with analysis of six languages. We analyse text material with TTR and MATTR from two differing sources: firstly, text of the EU constitution with 21 languages and secondly with 16 of the same languages with available non-parallel random data from the Leipzig corpus. We compare the automatic analysis results to two independent linguistic measures of morphological complexity. Firstly, we use number of non-homographic noun forms in a language's inflectional paradigms, the paradigm size. Secondly we use available inflectional synthesis figures of verbs produced by the AUTOTYP project. We enrich our corpus findings with data from information retrieval (IR) results. It has been suggested that improvements in achieved IR effectiveness with usage of word form variation management depend on the morphological complexity of the languages. Thus this IR gain data can be used to give independent evidence to evaluation of morphological complexity. Our results show that earlier Juola complexity figures and TTR and MATTR calculations correlate moderately in the EU constitution data. Figures given by TTR and MATTR correlate highly with each other in both corpora, and they also correlate highly with the number of non-homographic noun forms in a language. Correlation to inflectional synthesis of the verbs was found weakly positive in most cases, but the data was scarce. All the three computed measures are able to order the languages quite meaningfully in a morphological complexity order that at least groups most of the languages with same kind of languages and the most and least complex languages are clearly separated. It seems also that TTR and MATTR order the languages quite consistently with both corpora. In the conclusion we discuss how the complexity figures can be utilized. ACKNOWLEDGEMENTSMost of this paper was finished while the author was visiting UFAM, Universidad Federal do Amazonas, Institute of Computing, and funded by FAPEAM, Fundação de Amparo à Pesquisa do Estado do Amazonas (http://www.fapeam.am.gov.br/) with grant number 159/2012. The author wishes to thank the anonymous referee of the Journal of Quantitative Linguistics for useful comments when preparing the final version of the paper. Prof. Johanna Nichols kindly provided the AUTOTYP verbal inflection synthesis database for use.Notes* Address correspondence to: Kimmo Kettunen, Kiviteltankatu 6 A 11, FIN-00710 Helsinki. Tel: +358 50 5710859. E-mail: kkettun4@welho.com1 Strictly speaking Ehret and Szmrecsanyi use a variant of the Juola method, where 10% of the characters in the samples are removed randomly and the resulting file is compressed. This approach is introduced in Juola (Citation2008) and evidently it is comparable to the distortion method of Juola (Citation1998).