Flexible Meaning Adjustments in Context: Underspecification vs. Reinterpretation

How rare is complement coercion? Converging evidence from corpora, language models, and the language of science
Giosuè Baggio (Norwegian University of Science and Technology)

Theoretical and experimental research on complement coercion over the last three decades has largely assumed that the phenomenon is common across languages and registers, without testing whether this assumption holds. Recent results from our lab suggest otherwise: under a reasonably specific definition of complement coercion, canonical constructions (e.g., ‘The boy began the book’) turn out to be surprisingly infrequent. First, I support this claim with a corpus study of written Norwegian. Second, I turn to the consequences for machine learning: 17 Norwegian language models prove unable to reliably retrieve covert events, and additional linguistic context provides only limited benefits. Third, I show that coercion is similarly rare in the language of science, based on data from biomedical English. Finally, in search for an explanation, I speculate on the cognitive and computational origins of this infrequency. I argue that, in partial analogy with the evolution of programming languages, some natural languages and registers may avoid certain forms of type polymorphism, precisely for the computational (and therefore cognitive) costs that they incur.

A clozer look at possible and impossible interpretations in complement coercion: First insights from a two-step task
Oliver Bott (Bielefeld University)

Complement coercion as exemplified in Mia began the pizza has been extensively discussed in semantics as well as in psycho- and neurolinguistics. However, experimental evidence regarding the interpretation possibilities of complement coercion sentences have been scarce - especially when considering the limits of their semantics and pragmatics. In our talk we will present first results from a series of studies currently under way in project B01 of CRC 1646 trying to systematically investigate the space of admissable readings complement coercion sentences allow for. In doing so, we will contrast coercion sentences with aspectual verbs versus different kinds of psychological verbs which have often been discussed together in the coercion literature. In addition, we will focus on their combinatorics with complement noun phrases of different kinds. Our discussion will highlight the specific restrictions observed in complement coercion sentences in relation to overt realizations of event arguments as in Mia began to slice the pizza. The reported findings heavily make use of a specific task dedicated to comparing the constructional variants with each other opening a window into their interpretational possibilities.

A graph-based analysis of semantic types and coercion in
contextualized word embeddings

Complement coercion is characterized by the semantic type mismatch between a noun and its context. This talk introduces a graph-based method to examine how lexical and contextual type information is reflected in contextualized word embeddings. We select nouns from ten semantic types, annotate corpus instances for type matching (matching vs. coercion vs. other mismatch vs. unrestricted), and construct graphs using BERT and sense-enhanced BERT embeddings. Two metrics--Neighbor Type Probability (NTP) and Neighbor Type Entropy (NTE)--are proposed to analyze neighborhood type distributions. Results show that graphs constructed with sense-enhanced BERT embeddings reflect semantic type information better, and matching and mismatch sentences
can be distinguished through the proposed metrics. The method helps determine some tricky cases related to semantic types, and further inspires the identification and definition of coercion. It also reveals potential coercion instances that might be ignored in linguistic research.

Ad hoc creation of object mass nouns in Spanish and French – Productivity and meaning adjustment in context
Désirée Kleineberg (Bielefeld University)

In languages overtly marking the mass-count distinction, object mass nouns like Engl. furniture show syntactic properties of typical mass nouns in, for instance, not being compatible with the indefinite singular article (*a furniture). Unlike typical mass nouns such as water, they refer, however, not to homogeneous substances, but to a plurality of discrete entities. Consequently, they may be combined with stubbornly distributive predicates (round furniture vs. *round water). While being a recurrent topic in research on the mass-count distinction, also cross-linguistically, their morphological properties and particularly the possibility of their spontaneous formation have not been studied empirically to date. This contribution addresses this research gap and presents two experimental studies on the ad hoc creation of OMNs in French and Spanish. The results provide, on the one hand, empirical evidence for the general unproductivity of an OMN-schema and to their necessary long pathways of change. On the other hand, the experiments point to speaker’s ability of flexible meaning adjustment. Although they were shown ad hoc creations in simple syntactic contexts, they could always decipher the creative constructs presented in the experiment.
 

Framing the words unsaid: A frame-based approach to Complement Coercion operations, mechanisms and restrictions
Natalja Peiseler (Bielefeld University)

In this talk, we will present our approach to model meaning aspects of German complement coercion constructions such as Lisa begann das Buch, Lisa began the book. Our considerations focus on lexical selectional restriction of “coercion verbs”, combinatorial processes with regard to incremental sentence processing, and mechanisms and limitations of event inferences. We will explore the applicability of Düsseldorf-Frames (e.g. Gamerschlag et al. 2013; Kallmeyer,  Osswald & Pogodalla, 2017), or rather a slightly extended, psycholinguistically inspired version (Peiseler, 2024; Beckmann, Indefrey & Petersen, 2018) as a format to model processes of online-comprehension that involve mismatch or gap detection and mechanism of pragmatic enrichment.

Revisiting the processing cost of complement coercion: Pitting representational against information-theoretic accounts
Panagiota Rassia (Bielefeld University)

Sentences featuring complement coercion (e.g., Ilsa began the novel) have been typically analysed as involving a semantic type mismatch between an event‑selecting verb (i.e., began) and an entity‑denoting complement noun (i.e., novel). According to type‑shifting theory (Pustejovsky, 1991, i.a.), one of the most prominent and widely‑investigated theoretical accounts of complement coercion, this type mismatch is remedied by reinter‑preting the entity‑denoting noun as an event involving that noun (e.g., Ilsa began reading / writing / … the novel). Experimental work investigating the phenomenon within the type‑shifting framework has provided evidence that sentences of this kind are cognitively taxing above and beyond sentence plausibility (cf. McElree, Traxler, Pickering, Seely, & Jackendoff, 2001, i.a.). Over the years, different theories have been advanced in order to account for effects (or the lack thereof) associated with the processing of such sentences (de Almeida, 2004; de Almeida & Dwivedi, 2008; Lai, Lacadie, Constable, Deo, & Piñango, 2017; McElree et al., 2001; Piñango & Deo,
2016, i.a.). Despite their distinct theoretical motivations, all the proposed accounts posit a coercion‑specific mechanism. More recently, Delogu, Crocker, and Drenhaus (2017) challenged the claims of traditional repre‑sentational accounts and approached the phenomenon from an information‑theoretic angle. Premising on empirical evidence from an eye‑tracking‑while‑reading and an EEG experiment conducted in German, Deloguand colleagues argue that the processing effort observed for coerced expressions is due to the high surprisal (i.e., unexpectedness) of the complement noun. Despite the compelling evidence, Delogu et al. (2017) do not report acceptability norming results for their experimental items. An acceptability judgement experiment we conducted using Delogu et al.’s items revealed a decrease in acceptability when additional linguistic material followed the complement noun region. In this talk, I present ongoing experimental work on a conceptual repli‑cation of Delogu et al. (2017) as part of my PhD project.

Missive nouns, polysemy and coercion
Peter Sutton (University of Potsdam)

Noun-related complementiser clauses (nccs) which are formed with that in English have a limited distribution. They are found with a wide variety of informational nouns such as those in (1a). However, there is a class of informational nouns that have been observed not to license nccs, namely informational–physical entity ‘dot object’ nouns such as those in (1b) (see e.g., Elliott, 2020). Moreover, nccs cannot be used with nouns that straightforwardly denote physical objects (1c) or events (1d).

(1) 

a. the allegation/belief/evidence/fact/information/report/statement that Bilbo found the ring
b. #the book/brochure/newsletter/newspaper that Bilbo found the ring
c. #the house that Bilbo met Gandalf
d. #the event/celebration that Gandalf set off fireworks

This talk presents data, derived from corpus studies on English and Finnish, that are surprising, given the distribution of nccs with nouns such as book and newspaper in (1b). We show that a subset of informational–physical entity ‘dot object’ nouns in fact do license nccs, a class I label missive nouns. For instance, in (2) and (3) letter and artikkeli (‘article’, Finnish) occur with nccs.

(2) Well that old lady got a letter that she had to bring all her valuables and she was sent accordingly to an old people’s home. (UkWaC)

(3) Luin artikkelin, että vieä tänä vuonna tullan myymään  (FiTenTen24) read.1 pst article. ACC COMP still this. ESS year.ESS come.PASS sell.INF.ILL enemmän dvd-soittimia, kuin Blu-ray soittimia. more DVD-Player.PL.PAR than Blu-Ray player .PL.PAR "I read a article that, this year, there will still be more DVD-Player than Blu-Rayplayers sold."

These and similar data are challenging for analyses of nccs. On ratzer’s approach, nccs select for nouns that denote ‘contentful entities’ Bondarenko, 2022; Elliott, 2020; Kratzer, 2006; Moulton, 2015, 2009). However, the notion of contentful entities becomes ad hoc if, for instance, newspapers are not contentful entities, but newspaper articles are. An alternative to the Kratzerian analysis is to restrict the distribution of nccs in terms of the polysemy class of the head noun (Sutton, 2026, manuscript). However, this also cannot straightforwardly account for the data in (2) and (3). The goal of this talk is to present and assess both the alternative semantic and pragmatic solutions to this puzzle. Using a combination of corpus data and native speaker judgements for English and Finnish, I show that argue that, based on the restricted distribution of missive nouns with nccs, that the solution must be at least in part pragmatic, of which one possible analysis is based upon coercion.