CNL home CNL home
Home People Research Publications Resources Contact

Creating Language across Multiple Time-Scales

Language is a hallmark of the human species; the flexibility and unbounded expressivity of our linguistic abilities is unparalleled in the biological world. As such, language poses fundamental questions for the cognitive sciences: How did language evolve in response to environmental and biological forces? How is language acquired by each new generation? And how is language processed 'on-line' in everyday social interactions? These questions have typically been treated as separate topics, to be addressed more or less independently. The work done in the Cognitive Neuroscience Lab suggests this tendency is misguided — there are strong constraints across domains, allowing each to shed light on one another.

CNL research is therefore conducted within a unified framework for understanding language across multiple time-scales: evolution, acquisition and processing. The theoretical foundations for this framework are laid out in three key publications, each relating one of the time-scales to the others (evolution: Christiansen & Chater, 2008; acquisition: Chater & Christiansen, 2010; processing: Christiansen & Chater, 2016). A comprehensive account of the CNL research can be found in the book Creating language: Integrating evolution, acquisition, and processing by Christiansen & Chater (2016).

An overview of our evolutionary account of language can be found in the streaming video of Dr. Christiansen's keynote, Language Evolution: Complexity across Multiple Timescales, at the 2013 European Conference on Complex Systems (ECCS’13), Barcelona, Spain.

To get an idea of our perspective on language processing and its importance for acquisition, see the streaming video of Dr. Christiansen's plenary, Language Acquisition as Learning to Process, at the 2014 International Congress for the Study of Child Language, Amsterdam, the Netherlands.

Language as Shaped by the Brain

Our research on language evolution has shown that traditional notions of universal grammar as a biological endowment of abstract linguistic constraints can be ruled out on evolutionary grounds (Chater, Reali & Christiansen, 2009; Christiansen, Reali & Chater, 2011; Baronchelli, Chater, Pastor-Satorras & Christiansen, 2012; Baronchelli, Chater, Christiansen & Pasto- Satorras, 2013). Instead, the fit between the mechanisms employed for language and the way in which language is acquired and used can be explained by processes of cultural evolution shaped by the human brain (Christiansen & Chater, 2008; Chater & Christiansen, 2010). On this account, language evolved by 'piggy-backing' on pre-existing neural mechanisms, constrained by socio-pragmatic considerations, the nature of our thought processes, perceptuo-motor factors, and cognitive limitations on learning, memory and processing. Using a variety of different methods, the CNL Lab has been exploring how one of these constraints -- the ability to learn and process sequentially presented information -- might have played an important role in shaping language through cultural evolution (Reali & Christiansen, 2009). Together, the results from this and other cross-disciplinary research points to a key role for sequence learning in the evolution and acquisition of language (see also Frank & Christiansen, in press; Milne, Wilson & Christiansen, 2018).

Language Acquisition as Multiple-Cue Integration

The idea of language as shaped by the brain suggests that much of the neural hardware involved in language may not be specific to it. This means that language has to be acquired largely by mechanisms that are not uniquely dedicated for this purpose. A substantial amount of CNL Lab research is therefore dedicated to testing the hypothesis that language has evolved to rely on a constellation of probabilistic information sources, or 'cues', for its acquisition (Christiansen, 2013). Although each cue may be only partially reliable in isolation, when integrated they allow language to be as expressive as it is while still being learnable by domain-general learning mechanisms. Using corpus analyses, we have revealed that there are phonological cues in words that indicate their meaning probabilistically (Monaghan, Shillcock, Christiansen & Kirby, 2014; Blasi, Wichmann, Hammarström, Stadler & Christiansen, 2016). We have further shown that a probabilistic relationship exists between the sound of a word and how it is used: nouns tend to sound like other nouns and verbs like other verbs (e.g., Monaghan, Christiansen & Chater, 2007). Importantly, these phonological cues not only play a fundamental role in language acquisition (Fitneva, Christiansen & Monaghan, 2009) but also affect syntactic processing in adulthood (Farmer, Christiansen & Monaghan, 2006; Monaghan, Christiansen, Farmer & Fitneva, 2010). These results suggest that the integration of phonological cues with other types of information is integral to the computational architecture of our language system (Christiansen & Monaghan, 2016). Such multiple-cue integration, in turn, is what makes language learnable without universal grammar given the rich sources of information available in the input (Reali & Christiansen, 2005; Van den Bos, Christiansen & Misyak, 2012).

Language Processing as a Usage-Based Skill

The multiple-cue integration perspective on language acquisition highlights the rich nature of the input. In combination with the emphasis on cultural evolution of language, this points to a usage-based account of language processing in which linguistic experience plays a crucial role in determining language ability. A third line of CNL research thus investigates the impact of experience on language processing. Much of this work has focused on the processing of relative clauses as an example. Evidence from corpus analyses and on-line sentence processing experiments has shown that variations in the distribution of different relative clause types are directly reflected in the ease with which adults process such constructions (Reali & Christiansen, 2007). Further differences in relative clause processing are hypothesized to emerge from variation across individuals in their experience with language. Predictions from this account are supported by studies manipulating language exposure in both connectionist networks (MacDonald & Christiansen, 2002) and human subjects (Wells, Christiansen, Race, Acheson & MacDonald, 2009). Experimental data from ongoing work moreover suggest that individual differences in basic abilities for statistical learning, in turn, may affect individuals’ ability to learn from experience (Misyak, Christiansen & Tomblin, 2010; Misyak & Christiansen, 2012). Additional results from computational modeling (Christiansen & Chater, 1999; Christiansen & MacDonald, 2009) and behavioral experimentation (de Vries, Geukes, Zwitserlood, Petersson & Christiansen, 2012) indicate that such sequential learning also may support the processing of recursive sentence constructions (Christiansen & Chater, 2015). Seen together, this research indicates that language processing is best construed as a usage-based skill, relying on the integration of multiple constraints.

The Now-or-Never Bottleneck

A recent line of CNL research -- that cuts across evolution, acquisition and processing —- is investigating a fundamental constraint on language: the Now-or-Never bottleneck (Christiansen & Chater, 2016). During normal linguistic interaction, we are faced with an immense challenge by the combined effects of rapid input, short-lived sensory memory, and severely limited sequence memory. To overcome this bottleneck, language users must learn to chunk language input as rapidly as possible into increasingly more abstract levels of linguistic representation. Using a lab-based cultural transmission experiment, combined with network analyses (Baronchelli, Ferrer-i-Cancho, Pastor-Satorras, Chater & Christiansen, 2013), we showed that repeated processes of chunking across “generations” of human learners can result in language-like structural reuse (Cornish, Dale, Kirby & Christiansen, 2017). Focusing on language acquisition, we have developed a computational model that learns in a purely incremental fashion, through on-line processing of simple statistics, and offers broad, cross-linguistic coverage while uniting comprehension and production within a single framework (McCauley & Christiansen, 2011, submitted; Chater, McCauley & Christiansen, 2016). The model achieves strong performance across over 200 single-child corpora representing 29 different languages from the CHILDES database, simulates data from developmental psycholinguistic experiments (McCauley & Christiansen, 2014) and captures differences between first and second-language learners in chunk use (McCauley & Christiansen, 2017). Predictions from this computational modeling and the theoretical framework concerning the role of multiword units as building blocks of language (see also Arnon & Christiansen, 2017) have been confirmed experimentally (Arnon, McCauley & Christiansen, 2017; Jolsvai, McCauley & Christiansen, 2013; McCauley & Christiansen, 2015; McCauley, Isbilen & Christiansen, 2017).

Statistical Learning and Language

Our ability to learn aspects of linguistic structure from distributional regularities in the input is another line of CNL research that straddles the evolution, acquisition and processing of language. Accordingly, we have investigated some of the basic limitations on human statistical learning of sequential structure (Conway & Christiansen, 2005, 2006, 2009; for reviews, see Frost, Armstrong, Siegelman & Christiansen, 2015; Frost, Armstrong & Christiansen, submitted). Evidence from behavioral studies has demonstrated that individual differences in human sequence learning predict variations in syntactic processing abilities (Misyak & Christiansen, 2012; Misyak, Christiansen & Tomblin, 2010 -- for reviews of individual differences, see Siegelman, Bogaerts, Christiansen & Frost, 2017 [statistical learning]; Kidd, Donelley & Christiansen, 2018 [language]). Moreover, neuroimaging data revealed similar neural correlates of both sequence learning and syntactic processing (Christiansen, Conway & Onnis, 2012; Weber, Christiansen, Petersson, Indefrey & Hagoort, 2016) and results from a neuropsychological study showed that break-down of syntactic language in agrammatic aphasia is associated with a sequence learning deficit (Christiansen, Kelly, Shillcock & Greenfield, 2010). To determine whether humans may have evolved more complex sequence learning skills, we have been involved in the development of a new observer-independent joystick-based experimental paradigm for studying sequence learning in nonhuman primates (Heimbauer, Conway, Christiansen, Beran & Owren, 2012, in press). We have also developed a memory-based perspective on statistical learning, with a focus on the chunking of sequential information (Christiansen, in press), resulting in a new, more reliable method for measuring individual differences in statistical learning (Isbilen, McCauley, Kidd & Christiansen, 2017). Together, this part of my work has pointed to a key role for statistically-based chunking in the evolution (Isbilen & Christiansen, submitted), acquisition (Christiansen, in press), and processing (McCauley et al., 2017) of language.

The Puzzle of Danish

A final line of CNL research is focusing on the acquisition and processing of Danish. A key prediction of the cultural evolution of language perspective is that because of historical trajectories, not all languages may be equally easy to learn and use. Danish appears to provide an important example in this regard. Previous work has shown that compared to other comparable languages, Danish children are behind on vocabulary and past tense acquisition. Danish phonology is very complex, comprising about 40 different vowels, combined with a tendency for speakers to either delete consonants or turn them into vowel-like sounds (semi-vowels), resulting in long vowel sequences (as confirmed by corpus analyses; Trecca, McCauley, Andersen, Bleses, Basbøll, Højen, Madsen, Ribu & Christiansen, submitted). Preliminary work suggests that these phonological properties make it harder for Danish toddlers to learn new words (Trecca, Bleses, Madsen & Christiansen, 2018). At Aarhus University in Denmark, we are pursuing a project funded by the Danish Council for Independent Research to investigate the implications of the opaque Danish phonology on how adult Danes process language. This project combines computational analyses and behavioral experimentation to look at language processing at the phonological, sentential, and dialogue level, comparing Danish with Norwegian (a closely related language that has a more transparent phonology).

Research funded by

Cornell University