The notion of prediction is central to the understanding of language and cognition. In the most general view, it is assumed that the brain constantly creates explanations for its sensory input by generating internal models of the world that are tested against the actual input by the sensory systems. Mismatches between the model and the sensory input are reflected in prediction errors that lead to model updating (Friston 2005, 2010). By acknowledging these basic insights into the role of prediction in cognition, we assume that there are detectable traces of prediction in the way how language is used, processed and structured.

This idea is even more intriguing when we consider the manifold intuitions and notions of predictions, expectations and anticipation that can be found not only in the area of language processing but also in typological, historical, acquisition, and cross-modal linguistic research. All these notions share certain commonalities and there are clear associations with other cognitively motivated mechanisms such as feedback in communication and evaluative processes related to prominence in language.

The main purpose of the LingCologne 2023 conference is to identify and present the variety of notions related to prediction in language in order to identify and work out potential links and points of contact between distinct concepts and applications of prediction to the investigation of language. It is not the goal to develop a unitary concept of prediction but to highlight commonalities and shared traits. For this purpose, LingCologne2023 offers a forum to present the state of the art of distinct accounts to identify potential targets for a future scientific programme developing unifying frameworks to capture these accounts from a broader perspective.

In order to provide an initial “classification”, we identified four subject areas in which the notion of prediction plays a vital role with respect to language and related domains. This conference brings together eight of the world’s leading researchers working in these four areas as a first step for future research that focuses more directly and collaboratively on the role of prediction in language. Due to the interdisciplinary perspectiv, this conference also contributes to research relating prediction in language to prediction in cognition and ultimately the brain.

PROCESSING | Prediction in language processing

In the field of psycho- and neurolinguistics, it is widely acknowledged that “language users predict upcoming language input” (Huettig, 2015). The current debate of prediction in language processing largely revolves around the question to what extent the predicted material ranges – from the anticipation of specific elements (e.g., lexical content) to the constant expectation of features and sub-components in the course of processing.

Earlier research mainly focused on the anticipation of lexical elements such as “sugar” in the sentence “I like coffee with milk and __”. Kutas & Hillyard (1984) were the first to identify a particular electrophysiological component related to the violation of such lexical anticipations in form of an N400 effect. Subsequent research showed that the N400 is not sensitive to semantic violations alone but can also be elicited by expectation violations at different linguistic levels (such as morphosyntax or discourse). Bornkessel-Schlesewsky & Schlesewsky (2019, p. 1) argued more recently that the amplitude of the N400 directly “reflects […] prediction errors weighted by the relevance of the information source leading to the error”, relating linguistic research more directly to the advances in the field of cognitive neurosciences by implementing the generation of forward models as comprehensive representations that support the prediction of upcoming material in a more general sense.

As a general mechanism of cognition, the interaction of predictions does not only play a role in isolated words or controlled spoken sentences, but also in the whole complex of multi-modal face-to-face communication. This also includes predictions on facial expressions and co-speech gestures indicating that language comprehension uses predictions on forthcoming sensory patterns also beyond the auditory signal (Zhang et al. 2021).

STRUCTURE | Prediction in language structure, typology, and acquisition

Despite being a prominent topic in current research, there is no unitary framework capturing different aspects of prediction in language in a straightforward way. Even more so, “this line of psycholinguistic research has so far hardly been complemented by theoretical linguistic analysis investigating the phenomena in question from the perspective of anticipation.” (Solstad, Daskalaki & Järvikivi, 2021, p. 320). We assume that a closer investigation of the role of prediction also at the systems level of language can contribute and complement the understanding of the role of prediction in language in general. We explicitly assume that prediction is not a phenomenon of processing alone but rather reflected in the structures and principles of a language system itself.

In typology, it is often assumed that language users build up certain preferences based on the frequency of using particular structures (Comrie, 1989; Haspelmath, 2021). In addition, there are certain generalized principles that may derive from cognitive demands or other non-linguistic constraints (e.g., the S/A preference as a species-specific property, see Bickel et al., 2015). Language systems are structured in line with these preferences and many linguistic constructions are often described as being organized alongside certain rankings or hierarchies (e.g. the accessibility hierarchy, prominence rankings). Deviations from such preferred patterns and or rankings entail a certain notion of lower predictability and the lower ranked or “unexpected” elements are typically marked by particular encoding strategies. Assuming that humans tend to avoid prediction errors, this could explain why particular (or simply longer) encoding mechanisms are used for elements that violate structure-based predictions.

Predictions also play a role in language acquisition and here the link to predictability in other cognitive domains is most striking. Drawing on predictive processing accounts, some authors state that language acquisition is directly associated with the learning to predict (Mani & Huettig, 2012; Molinaro et al., 2017). These accounts correlate the competence of a language learner to the ability to predict upcoming language input (with an acclaimed mechanism of anticipatory language processing) and explicitly point at the role of generalized predictive processes at different levels beyond the anticipation of only a particular set of elements.

MODELS | Prediction in computational models

Assessing the role of prediction in language requires the joint collaboration from different areas of language research and the combination of different data types, such as corpus, behavioural and neurophysiological data. Integrating data within a systematic theoretically and empirically plausible framework requires powerful computational models that capture the high level of variation typically found in language and communication.

For instance, Rabovsky and colleagues (e.g., Rabovsky, 2020; Rabovsky et al., 2018) developed neural network models for the aforementioned N400 effect capturing changes in the probabilistic representation of meaning associated with this effect. In their research, they attempt at developing models that bring together semantic theory, models of probabilistic distribution and neural correlates, thereby offering a blueprint for integrative prediction-oriented research of other linguistic levels.

Other researchers develop models that capture the integration of sensory input and internal expectations related to communication. For instance, Blank and colleagues investigate how prior expectations influence speech and face perception during human communication by combining behavioural, neuroimaging, and computational methods (e.g., Blank & von Kriegstein, 2013; Blank & Davis, 2016). These thorough analyses of the role of speech and face perception in communication and the investigation of signal integration with respect to prediction errors at these levels provide the foundation for future investigations of higher level aspects of language and communication.

COGNITION | Prediction in other cognitive domains

Recent accounts in cognitive science as well as philosophy of mind assign predictive mechanisms a central role in human cognition. The basic assumption is that our “mind is shaped by how we manage these predictive efforts” (Hohwy, 2013, p. 258). In these accounts, basically nothing is ever perceived and no information is processed without the constant matching of the input to a pre-generated model of how the input will most likely look like. This does not only affect external sensations but also internal information processing in the mind. Related to the working of the brain, these accounts “captured a growing consensus that one could understand the brain as a statistical organ” (Friston, 2018, p. 1019).

In philosophy, some proponents of a strong prediction account suggest that predictive coding as a basic underlying mechanism allows for the implementation of a single unifying framework of cognition in the future (Clark, 2013; Hohwy, 2013). In psychology, there is a broad body of research investigating the links of predictive processes to other cognitive domains, such as memory, attention, and decision making. Particularly interesting in this regard is the investigation of (domain-general) feedback mechanisms and error-related negativities associated with adaptive behaviour and selective attention following behavioural errors (Maier et al., 2011; Yeung et al., 2004). Finally, the human capacity of learning in general is argued to be intimately linked to the ability to predict (Frömer et al., 2021), complementing the recent claims in language acquisition research mentioned above.