Compositionality and the Crosslinguistic Variation of Complement Coercion

September 7-8, 2023, Scandic Hotel Nidelven, Trondheim

An international workshop supported by a grant from the Van Riemsdijk Foundation and by strategic funds from the Department of Language and Literature, Faculty of Humanities, Norwegian University of Science and Technology

Thursday, September 7

9:00-9:15 — Introduction, Giosuè Baggio

9:15-10:00 — The Active Lexicon

Gillian Ramchand (UiT The Arctic University of Norway)

In this talk, I explore the role of the Lexicon in compositional processes. I embrace two core innovations which depart from the assumptions usually found in the linguistic literature. Instead of approaching the issue from the monolithic notions of ‘recursion’ focused on the sentence as a whole while abstracting away from lexical contents, I explicitly centre the role of the Lexical item and its properties in memory as the nexus of combinatoric action. Secondly, instead of seeing combinatorics through the lens of function-argument composition, I argue for partitioning primitive composition into two main types: referential narrowing vs conceptual combination. In this sense, I endorse a version of a two-tiered theory of meaning composition (Pelletier 2017), which interleaves conceptual and referential affordances (Asher 2011; McNally & Boleda 2017).

10:00-10:15 — Coffee break

10:15-11:00 — A Meaningful Relation Between Time and Space: Potentially Universal Properties of Natural Language (and Hence the Syntactic Spine)

Kristin Melum Eide (Norwegian University of Science and Technology) 

This talk explores the idea that all categories projecting to the syntactic spine are time-related, denoting events (encompassing states) which by their very nature exist in time. A predication occurs when this abstract time line is crossed by an anchoring entity (an object/location) existing in space. This fundamental time-space relation underlies the universality of nouns vs verbs, of entities vs properties/predicates, and the EPP (including expletives). Coercion occurs e.g. in complementation as negotiations between otherwise rigid syntax and the conceptually much more flexible semantic representation. I present illustrating data from domains of predication, deixis, tense, aspect, modality and (hopefully) extraction.

11:00-11:30 — Discussion 

11:30-13:00 — Lunch break

13:00-13:45 — Logical Metonymy: A Rational Speech Act Approach

Alessandro Lenci (University of Pisa)

The Rational Speech Act (RSA) framework (Goodman & Frank 2016, Franke & Jäger 2016) is a Bayesian model that formalizes the Gricean approach to communication based on the recursive reasoning of speakers' and listeners' about each other mutual knowledge and expectations. RSA has been applied to address various pragmatic phenomena (e.g. implicatures) and referential games. In this talk, I will present an incremental version of RSA, arguing that it can provide a computational model of enriched compositionality (Jackendoff 1997). The proposed approach integrates RSA with neural language models used to generate dynamic expectations. I will showcase the RSA model to explain behavioral data about logical metonymy.

13:45-14:30 — What do Brain Data Tell us about Coercion

Lia Călinescu (Norwegian University of Science and Technology)

Compositionality is a central principle in linguistic theory and has been seen by many as a reasonable working hypothesis for understanding the neural computations involved in online meaning comprehension. Studying purely semantic composition has however proven challenging due to the difficulty of teasing apart processes related to semantics from those that reflect syntactic structure building. Although at its basis the principle of compositionality has been taken to indicate a tight correspondence between syntactic and semantic levels of representation (Partee 1997), it has also been shown that the basic operations of simple composition are not enough to account for the range of meanings speakers are able to derive (Jackendoff 1997, Pylkkänen and McElree 2006). Thus, some researchers propose an enriched compositional apparatus, where additional semantic processes are needed to derive meaning, and which are not necessarily mirrored by the syntax. Phenomena involving enriched composition have been seen by the neurolinguistic community as a unique opportunity of probing purely semantic processes in the brain. Few other constructions have been investigated so extensively in the experimental literature as complement coercion. This phenomenon arises when an aspectual verb such as begin, finish, continue, combines with an entity denoting complement manuscript (e.g., Pedro finished the manuscript), giving rise to an additional inference (i.e., Pedro finished writing the manuscript) which is not encoded in either the lexical items alone or the syntactic configuration of the sentence. For over 20 years, numerous experiments using almost all methods available to date (eye-tracking, self paced reading, EEG/MEG, fMRI) have consistently shown processing costs related to the processing of an entity denoting noun preceded by a coercion verb (Traxler et al 2005, Pylkaanen and McElree 2007, Baggio et al 2010). Such results have been taken to index processes involved in the construction of an event sense for the complement. In a new EEG study, we asked whether previous results generalize to the Norwegian language. In Norwegian, aspectual verbs are very rarely used in combination with entity denoting nouns, so we wanted to see if Norwegian speakers are still able to derive the eventive interpretation of the complement and fully and successfully comprehend coercion sentences. In addition, we used a novel paradigm, in an attempt to independently isolate processes related to both simple and enriched composition. Our ERP results suggest that what has been previously attributed to processing costs incurred by the need to construct an event interpretation for the complement might reflect the degree of semantic relatedness between verb and noun. This leaves open the question as to whether complement coercion is actually costly for the language processor. Time-frequency data additionally suggest an intricate pattern of results, compatible with a flexible and complex composition apparatus.

14:30-14:45 — Coffee break

14:45-16:00 — Discussion 

19:00 — Social dinner

Friday, September 8

9:00-9:15 — Introduction, Giosuè Baggio

9:15-10:00 — Mass-Count Coercion in Distributional Semantic Spaces

Emmanuele Chersoni (The Hong Kong Polytechnic University)

The distinction between count nouns and mass nouns has often been described in the formal semantics literature as the opposition between discrete, countable objects and substances that cannot be divided into sub-units. However, cases like the following are extremely frequent in natural language:
     (1) There is rabbit in my soup. (count noun in mass context)
     (2) Two wines at table four! (mass noun in count context)
Both cases are examples of coercion, a phenomenon occurring when the standard interpretation of an expression yields an impossible conceptual representation, and consequently, a more plausible interpretation is built by enriching the semantic representation with concepts that are associated with the standard meaning of the target expression (enriched composition; Jackendoff 1997). In real language use, it is rare to find nouns occurring exclusively in mass or in count contexts, so the distinction between mass and count nouns can be conceived as a continuum. We present a first study in which we extracted a set of candidate mass and count nouns from natural language corpora, dividing the sentence contexts into mass and count ones; next, we used static and contextualized distributional semantic models to measure the semantic shift of nouns when they are coerced into a different type (Liu & Chersoni 2022). The issue of mass and count syntax has also been investigated in experimental studies on light verb constructions, and in particular, it has been argued that the choice of syntactic construction affects duration in event construal (Wittenberg & Levy 2017): for example, punctive events in count syntax are conceptualized as individual sub-events (e.g. to kiss > to give a kiss) and are perceived by humans as taking less time than in the transitive frame. We propose a second study in which we use the BERT language model (Devlin et al. 2019) to generate the distributional representations of verbs and corresponding light verb constructions, and we apply the method of semantic projection (Grand et al. 2022) to check if a shift in temporal duration can be observed in the dimensions of the vector space model.

10:00-10:15 — Coffee break

10:15-11:00 — Lexical Creativity in Compositionality: What is the Role and Place of Coercion?

Alexandra Anna Spalek (University of Oslo)

This talk reflects about the nature of the lexicon with a particular focus on its interpretational flexibility. I will focus on the verbal domain using data from Spanish, German and English. The description of the empirical landscape will illustrate that the meaning of verbs strongly depends on their arguments. In a short survey of some central models of the lexicon, I will address the need for coercion, understood as a compositional clash. The contribution of my talk will be twofold: 1) it calls for the need of underspecification and coercion, and 2) it shows that coercion is a broader phenomenon than ‘complement coercion’ and ‘aspectual coercion’.

11:00-11:30 — Discussion

11:30-13:00 — Lunch break

13:00-13:45 — On the Infrequency of Complement Coercion with Norwegian Aspectual Verbs: A Corpus Study

Matteo Radaelli (Norwegian University of Science and Technology)

Combinations of an aspectual verb and an entity denoting complement (e.g., 'begin the book') have been often assumed to involve a mismatch between the verb’s requirements (an event) and the complement’s denotation (an entity; Pustejovsky 1991, 1995; Jackendoff 1997). The interpretation of the entire verb phrase is possible through reconstruction of the covert event (e.g., reading). In this talk, I will present a corpus-based analysis of complement coercion in Norwegian. The objective of this study is to explore this phenomenon in a language that has been subjected to little empirical and theoretical research, with the intention of gaining deeper insight into the structural variations of the complement coercion construction in Norwegian, especially focusing on their syntactic patterns, variation in composition and frequency. The corpus analysis was performed on four corpora, one web-based and three speech-based, taking into account the main aspectual verbs in Norwegian (begynne ‘begin’, starte ‘start’, ende ‘end’, avslutte ‘end/conclude’, fortsette ‘continue’). From preliminary results, our investigations have interestingly revealed a low occurrence frequency of complement coercion, with only 56 instances identified among 37730 sentences. These results suggest that Norwegian speakers do not tend to use such constructions to integrate aspectual information of an expressed action. They would rather either use full forms with non-finite clauses (e.g., infinitive clauses) or fully specific event verbs (eg., spise opp 'eat up') that implicitly integrate aspectual information of the expressed action.

13:45-14:30 — LLM Learnable and Unlearnable Linguistic Meanings

Nicholas Asher (Université Paul Sabatier)

Coercion makes the simple function application model of compositionality unworkable or at least very unhelpful. Asher (2011) makes a valiant attempt to formalize coercion within the lambda calculus, but that is something that seems almost effortless in large language models. My talk begins by drawing comparisons between LLM semantics and more standard truth conditional semantics using continuations. I then turn to the thorny question of what LLMs can learn given their distributional semantic underpinnings. I rehearse a couple of claims or rather theoretical proofs (Asher et al. 2023) that certain "sharp concepts" are unlearnable by LLMs, even if we abstract away from physical and computational limitations like say context window size. These results, I believe, shed light on exactly how coercion and flexible, contextually aware composition should be understood.

14:30-14:45 — Coffee break

14:45-16:00 — Discussion and conclusions

Organisation and contact: Giosuè Baggio (giosue.baggio@ntnu.no)