A large body of research over the past two decades has demonstrated that children and adults are equipped with statistical learning mechanisms that facilitate their language processing and boost their acquisition. However, this research has been conducted primarily using artificial languages that are highly simplified relative to real language input. Here, we aimed to determine to what extent adult native and non-native speakers show sensitivity to real-life language statistics obtained from large-scale analyses of authentic language use. Through a within-subject design, we conducted a series of behavioral experiments geared towards assessing the sensitivity to two types of distributional statistics (frequency and entropy) during online processing of multiword sequences across four registers of English (spoken, fiction, news and academic language). Our results show that both native and non-native speakers are able to `tune to' multiple distributional statistics inherent in different types of real language input.

Language production is incremental in nature; we tend to plan linguistic chunks prior to articulating the first word of the utterance. Researchers have acquired knowledge about how far ahead sentences are generally planned, but mostly in monolinguals or the speaker's first language (Allum & Wheeldon, 2007; Martin, Crowther, Knight, Tamborello II, & Yang, 2010; Wagner, Jescheniak, & Schriefers, 2010). It is unclear whether the scope of planning is the same in bilinguals, or the speaker's second language. Here, we examined planning scope in Dutch–English bilinguals' sentence production using a paradigm that elicits descriptions of short animations. Analyses of speech onset times and articulation durations suggest that, on the surface, bilinguals have comparable planning scope in L1 and L2. However, in their L2, bilinguals extended their articulation duration, suggesting that they committed early to the initial noun phrase, but produced it more slowly to buy time to plan the next noun phrase.