Composing the PSP of Free Choice
Sigrid Beck (University of Tübingen)
Expressions of free choice like 'whatever', 'any' and Free Choice Relative clauses have been argued to give rise to a counterfactual presupposition PSP (e.g. Dayal (1998), von Fintel (2000)). This PSP has received some limited attention in the semantic literature, but its composition with the free choice element is not yet clear. The talk brings together some recent contributions, striving for a synthesis.
Representing Implicit Objects in Discourse: a Semantics of Generalized Result Nominals
James Pustejovsky (Brandeis University)
In this talk, I discuss the relationship between language as used in narrative and dialogue and how this language refers to a speaker's dynamically changing world. More specifically, I discuss the phenomenon in language of semantically implicit arguments, those objects that are not expressed linguistically in the discourse, but which are created as a result of event predicates bringing about a change of state. This includes both abstract objects and concrete objects introduced into the discourse through linguistic or nonverbal communicative acts, as well as situated actions and events. To this end, I introduce the class of "Generalized Result Nominals (GRN"), a notion that incorporates both linguistically realized surface expressions in language, and a formal representation of the creation resulting from any state-changing predicate in language.
Working within Generative Lexicon Theory, we adopt the Dynamic Event Semantics of Pustejovsky and Moszkowicz (2011), where we define a Generalized Result Nominal as the object-oriented reification of the change brought about by any change of state predicate. Developing the theory of result nominals presented in Mellon (2011) and the Dynamic Argument Structure of Jezek and Pustejovsky (2019), we show both how ubiquituous Generalized Result Nominals are in language, and how they play a crucial role in discourse and narrative coherence. All transformation predicates, not just creation verbs, output a new entity into the discourse. When these are not expressed explicitly in the syntax, they are represented as implicit arguments.
This allows us to model narratives as both state-changes as well as input-output process graphs over entities. This in turn provides a rich representation with which to perform logical reasoning and inference in computational tasks.
A bigger picture: The graphics of lexical decomposition
Joost Zwaarts (Utrecht University)
Decompositions of word meanings come in many visual shapes: formulas, trees, tables, diagrams, graphs, pictures, ... This talk explores this graphical diversity in order to get a bigger picture of the what, how and why of lexical decompositions, in formal and cognitive linguistics, onomasiological and semasiological, data-driven and theory-driven, symbolic and non-symbolic. It contributes to a much-needed reflection in semantics and linguistics about its variety of meta-languages, against the background of the growing attention for visualization and diagramming.
A Decompositional Semantics for Responsive Attitude Predicates: The Case of 'to know'
Edgar Onea & Malte Zimmermann (University of Graz, University of Potsdam, resp.)
Putting the discourse on hold with 'denn'
Doris Penka (University of Konstanz)
The German discourse particle 'denn' is restricted to occur (mainly) in questions, e.g. 'Wie spät ist es denn'? In the spirit of much recent work on discourse particles, Theiler (2021) proposed an account of 'denn' in terms of Conventional Implicatures (CIs): 'denn' is felicitous in a question Q iff the speaker S requires an answer to Q to proceed in the discourse.
We implement Theiler’s analysis in an extended Scoreboard model of discourse, where a context structure comprises (at least) the Common Ground, a stack of questions to be addressed and a To-do List for each interlocutor. In the spirit of Farkas & Bruce (2010), we assume an actual and a projected version of each of these components. Under this view, an utterance is a proposal to update the respective component. A denn-Q uttered as reaction to a preceding utterance is analyzed as an intermediate discourse move between a proposal to update and the (intended) actual update. More concretely, a denn-Q stops the projected context from becoming the actual context and interleaves a new question that needs to be dealt with first. We define a “stopping” operation Hold. The contribution of a denn-Q can then be stated as combining its CI content in terms of Hold with the usual contribution of a questioning speech act. Thus, a denn-Q at the same time posits a novel question and, via denn/Hold, signals acknowledgment of the previous move.
We also show that cases where a denn-Q doesn’t react to a previous utterance can be subsumed under the proposed account, if it is assumed that update proposals are driven not just by the corresponding explicit discourse moves but also by implicit (though still public) self-driven moves triggered by non-verbal evidence. We thus use 'denn' as a window to investigate the properties of discourse dynamics.
Process-oriented event model for coreference resolution
Kyeongmin Rim (Brandeis University)
In this work, we demonstrate an annotation scheme and its computational implementation to annotate coreference relations between text mentions based on process-oriented event model that integrates argument structures, subevent decompositions, and event temporality. The annotation process is made up of three phases: 1) mention/span recognition and temporal ordering, 2) event-participant linking and coreference linking, and 3) coreference classification based on subevent semantics. The results of the annotation are visualized into unified semantic graphs that represent all aspects of annotation phases. With the proposed unified graphical representation of event semantics, we argue that coreference resolution can be significantly improved, especially with spatial, temporal and transformation-based entity linking. Specifically, we propose a coreference typology (used in the third phase of the annotation) that is expanding previous study on "near-identity"-based coreference relations with linkages directly inferred from subevent structures derived from existing lexical resources such as VerbNet. By treating entity state changes rooted in event semantics (spatial, temporal and transformational) as coreferential (near-identity) expressions, we add the spatio-temporal dimensions (that comes free with dynamic event structures) to the coreference resolution task. This allows us to build timeline-like unified graphs that illustrate entity relations and coreference chains in a text document simultaneously.
We conducted manual annotation (by linguistics experts) based on the scheme and tool we developed, using different genres of text as source data. We report annotation reliability measures and highlight problems and issues we faced using the scheme on different genres. The annotation environment and data are publicly available as open-source projects.