Communicative efficiency and syntactic predictability: A cross- linguistic study based on the Universal Dependencies corpora

There is ample evidence that human communication is organized efficiently: more predictable information is usually encoded by shorter linguistic forms and less predictable information is represented by longer forms. The present study, which is based on the Universal Dependencies corpora, investigates if the length of words can be predicted from the average syntactic information content, which is defined as the average information content of a word given its counterpart in a dyadic syntactic relationship. The effect of this variable is tested on the data from nine typologically diverse languages while controlling for a number of other well-known parameters: word frequency and average word predictability based on the preceding and following words. Poisson generalized linear models and conditional random forests show that the words with higher average syntactic informativity are usually longer in most languages, although this effect is often found in interactions with average information content based on the neighbouring words. The results of this study demonstrate that syntactic predictability should be considered as a separate factor in future work on communicative efficiency. 1 Research hypothesis It is well known that more predictable information tends to be presented by shorter forms and less coding material, whereas less predictable information is expressed by longer forms and more coding material. This formfunction mapping allows for efficient communication. A famous example is the inverse correlation between the frequency of a linguistic unit and its length discovered by Zipf (1935[1968]). The main cause is an underlying law of economy, saving time and effort (Ibid: 38). In the domain of grammar, Greenberg (1966) provided substantial cross-linguistic evidence that relative frequencies of unmarked members of grammatical categories (e.g. singular number or present tense) are more frequent than their marked counterparts (e.g. dual/plural or future/past, respectively). This idea has been developed further by Haspelmath (2008), who provides numerous examples of coding asymmetries in which the more frequent morphosyntactic forms are shorter than the functionally comparable less frequent ones. These asymmetries can be explained by the tendency of language users to make communication efficient: “The overall number of formal units that speakers need to produce in communication is reduced when the more frequent and expected property values are assigned zero” (Hawkins, 2014: 16). While the accounts mentioned above are based on context-free probability of linguistic units, some other approaches, which go back to Shannon’s (1948) information theory, take into consideration the conditional probability of a unit given its context. The measures computed from these conditional probabilities are often called information content, surprisal, or informativity. There is ample evidence of ‘online’ word reduction in speech production based on contextual predictability (e.g. Aylett and Turk, 2004; Bell et al., 2009). In addition, one has found ‘offline’ effects of average informativity on formal length in written corpora: the more predictable a word is on average, the shorter it is (Piatandosi et al., 2011). One of the explanations of such correlations is known as the hypothesis of Uniform Information Density (Levy and Jaeger, 2007), which says that information tends to be distributed uniformly across the speech