Composing or Not Composing? Towards Distributional Construction Grammar
Philippe Blache, Giulia Rambelli, Emmanuele Chersoni, Alessandro Lenci · 2024
Read Composing or Not Composing? Towards Distributional Construction Grammar by Philippe Blache, Giulia Rambelli, Emmanuele Chersoni, Alessandro Lenci (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-07419.txt.
Introduction
Philippe Blache and co-authors ask how listeners build meaning from speech when not every phrase composes neatly, in a 2024 paper. Classic accounts assemble meaning step by step, yet idioms and fixed phrases refuse that treatment. The team extends construction grammar, where stored form-meaning pairs do the work, with tools from distributional semantics.
Their formal frame joins constructions, frames, and events, and reads understanding as an activation contest settled by similarity and unification. Word-usage statistics feed the construction store, so frequent patterns glow brighter. The proposal aims at language technology that handles both the rule-like and the rote parts of real speech.
Catalog record
| CREATOR | Philippe Blache, Giulia Rambelli, Emmanuele Chersoni, Alessandro Lenci |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 68 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-07419
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Composing or Not Composing? Towards Distributional Construction Grammar Philippe Blache, Giulia Rambelli, Emmanuele Chersoni, Alessandro Lenci Abstract The mechanisms underlying language comprehension remain an open question. Classically, building the meaning of a linguistic utterance is said to be incremental, step-by-step, based on a compositional process. However, several works have long demonstrated that non-compositional phenomena are also at work. It is therefore necessary to implement a unified approach that integrates both approaches. In this paper, we propose a framework based on Construction Grammar that accounts for these different access mechanisms. First, we provide a formal description of this framework by extending the feature structure representation proposed in Sign-Based Construction Grammar . We present a general representation of the meaning based on the interaction of three semantic components: constructions, frames and events. Secondly, we describe how to implement a processing mechanism for building the meaning based on the notion of activation evaluated in terms of similarity and unification. Keywords: Compositionality, Construction Grammar, Distributional Semantics, comprehension, hybrid access to the meaning 1 Introduction This paper addresses a long-standing and still unresolved question: What are the cognitive and linguistic principles that underlie natural language comprehension? While most existing research has focused on written texts and formal language use, we still lack a clear understanding of how meaning is accessed in more natural settings such as everyday conversation where speakers often produce incomplete, noisy, or non-canonical utterances. The classical view is grounded in the principle of compositionality (see szabo2004compositionality for an overview) which holds that the meaning of a sentence or more generally a message, whether written or spoken, is a function of the meanings of its constituent parts. This process is typically understood to be incremental: individual word meanings are combined into increasingly complex structures in a stepwise, hierarchical fashion until a full interpretation is achieved. However, research in theoretical, descriptive, and experimental linguistics has shown that meaning can often be derived more holistically from global patterns or conventionalized structures without relying solely on compositional mechanisms. The large number of facilitation effects observed in processing idiomatic expressions and other multiword expressions ( contreras2022models ) , as well as compositional frequent expressions ( rambelli2023frequent ) , goes against a strong version of compositionality. Indeed, we may still recognize the parts composing the meaning of well-known expressions (i.e., their compositionality), without this entailing that we “build them from scratch” each time we use them. At the same time, several studies demonstrate that online processing may occur in a good-enough manner, relying on expectations, and potentially resulting in shallow interpretation or misinterpretation, such as for garden-path sentences ( christianson2001thematic ; ferreira2003misinterpretation ) or semantic attraction ( kim2005independence ; Brouwer17 ; cong2023language ) . Moreover, efficient comprehension usually adopts contextual constraints to anticipate or predict upcoming input, leading to facilitated processing once the expected component is encountered ( levy2008 ; huettig2015 ; Ferreira2016-ve ; pickering2018predicting ; pickering2013integrated ; goldberg2022good ; Blache24 ; britton2024influence ) . Taken together, this body of evidence suggests that two distinct but concurrent mechanisms are involved in language comprehension. The first is a compositional route, in which meaning is constructed incrementally, word by word, as each new lexical item is integrated into an evolving syntactic and semantic structure. The second is a non-compositional route, which relies on the recognition of familiar patterns or constructions, enabling more direct access to meaning without full syntactic parsing. A central challenge, then, is to develop a unified formalism that can account for the interaction between these two modes of processing. Additionally, when trying to explain the understanding process, a second important question is the influence of the context. More precisely, several approaches in sociolinguistics, corpus linguistics and cognitive linguistics ( Langacker00 ; Tomasello05 ; Bybee10 ) have underlined the fact that meaning emerges from language use. These so-called usage-based theories argue that every linguistic structure is shaped by repeated exposure and usage patterns. In particular, at the lexical level, the distribution of the words constitutes one of the most important sources of information for accessing their meaning: usage specifies the semantic relations between words. To summarize, explaining language comprehension needs to account for the fact that 1) meaning can be accessed following two different routes: incrementally, in a compositional way, but also globally, based on a direct access and 2) meaning depends on use and context. While traditional theories do not account for these phenomena, these observations are at the basis of Construction Grammar ( Fillmore88 ; Goldberg03 ; HoffmanTrousdale2013 ; Hilpert2019 ; Hoffmann2022 ) . In this theoretical paradigm, linguistic objects are organized in terms of form/meaning pairs called constructions (hereafter Cx ; Goldberg03 ; Goldberg06 ). While there is no converging definition and operationalization of Cxs ( ungerer2023constructionist ) , what is interesting is that constructions can be elementary objects (morpheme, words) as well as complex ones ( Goldberg03 ; Goldberg06 ; Sag12 ) : multiword expressions, idioms, syntactic constructions (e.g., ditransitive, passive, covariational conditional etc.). Moreover, constructions differ in their degree of schematicity, ranging from fully lexicalized and fixed expressions (e.g., kick the bucket ) to highly schematic and fully productive constructions (e.g., the ditransitive construction ). In all cases, the form-meaning pairing is at the core of the definition of a Cx : recognizing the form makes it possible to access its meaning directly. For example, in the case of idioms, as soon as the construction is identified (typically after 2 or 3 words), the complete meaning is accessed directly, without any compositional mechanism. However, most constructionist approaches have primarily focused on formalizing representations rather than modeling the dynamic processes of language comprehension. To bridge this gap, a more comprehensive framework is needed that 1) operationalizes how Cxs are integrated during processing by means of compositional and noncompositional mechanisms, and 2) includes usage-based semantic information into structured linguistic descriptions. This paper addresses both issues with the final aim to build a constructionist model of language processing. Considering the first point, it is necessary to identify in which situation one route is used instead of the other. Our hypothesis is that, by default, a pattern-based mechanism is applied, making it possible to recognize constructions and explaining different shallow processing effects ( Ferreira07 ) . In this case, constructions are recognized on the basis of different cues (morphological, lexical, distributional, etc.). This hypothesis relies on the notion of activation ( Anderson04 ; Lewis05 ; Blache24 ) , considering that a construction can be recognized above a certain activation threshold. Considering the second issue, most examples of constructional meaning are hand-coded, leaving the semantics of constructions unclear ( feldman2020advances ; ungerer2023constructionist ) . To overcome this limit, we introduce a novel semantic representation of Construction Grammar, named Distributional Construction Grammar (DCxG). The idea is to incorporate distributional vectors into the construction representation, thereby accommodating the more usage-based aspects of meaning. To summarize, the goal of this paper is twofold: • Specifying the nature and the representation of constructions, integrating usage-based models of meaning (i.e., the Distributional Semantics framework); • Defining an activation function and the different mechanisms making it possible to identify when constructions are recognized. The perspective of this paper is mainly theoretical, with the aim of laying down the foundations of a general model of