LingLunch: Philippe Blache (Aix-Marseille Université/ LPL)

Philippe Blache (Aix-Marseille Université/ LPL): Bridging Hierarchical and Parallel Perspectives: A Two-Level Model for Language Processing

Language processing has long been conceptualized as a hierarchical system. Throughout the 20th century, many linguistic theories—whether explicitly or implicitly—organized language into a sequence of ordered modules (Chomsky, 1957; Tesnière, 1959). More recently, this view has been revisited in both artificial intelligence and neurolinguistics (Tenney et al., 2019; Friederici, 2017), offering renewed support for hierarchical organization. This perspective has clear advantages, particularly from a computational standpoint, as it neatly separates linguistic categories and structures at each level. However, it faces difficulties in accounting for phenomena such as non-compositionality, incomplete information, or ill-formed inputs, all of which are common in natural communication.

In the late 1990s, an alternative perspective emerged, challenging this view by proposing a parallel architecture (Jackendoff, 1997, 2002). In this framework, different components of language processing operate independently rather than sequentially. Their interactions are managed through interface links that align representations across levels, making a fully parallel system theoretically possible. Approaches such as Construction Grammar (Fillmore, 1988; Goldberg, 2006) are consistent with this view, as they integrate multiple sources of linguistic information into unified representations, effectively implementing these interface connections.

Despite its conceptual appeal, this parallel approach remains largely theoretical in practical terms. Although each level independently generates information that can contribute to identifying constructions combining multiple sources, the model does not fully explain how constructions of varying granularity—some highly lexicalized, others more abstract—are produced. Nor does it clarify how these constructions are ultimately combined into a coherent, global representation of language.

In this presentation, I propose an approach that bridges these two perspectives. Grounded in a two-level framework, this model addresses how basic units are constructed at varying levels of granularity and how they are subsequently integrated. At the first level, parallelism is implemented, enabling the formation of multimodal units—ranging from simple to complex—while allowing, when possible, direct and non-compositional access to partial meaning. The second, higher level focuses on integrating these units into a cohesive, global semantic representation.