Distinguishing fair from unfair compositional generalization tasks

2025-01-01

Compositional generalization benchmarks seek to assess whether learning agents can successfully combine familiar concepts in novel ways.COGS (Kim and Linzen, 2020) provides a suite of such tasks in the area of interpretive semantics (mapping sentences to logical forms).A noteworthy finding for COGS is that model performance varies widely across tasks.In this paper, we argue that these performance differences reflect deep properties of these tasks.We focus on two COGS tasks: an easy task (models are generally successful) and a hard task (no present-day models get any traction).Using both experiments and conceptual analysis, we argue that the easy task requires only a single distributional generalization that is wellsupported by the training data, whereas the hard task involves a learning target that is ambiguous or even contradicted by the training data.We additionally argue that pretraining can disambiguate the hard task without compromising the goal of testing compositional generalization.Overall, our findings offer practical guidance to designers of compositional generalization benchmarks and also yield new insights into the nature of compositionality itself.

<div> Reading Law with Linguistics: How Linguistic Theory and Data Inform Statutory Interpretation of Artifact Nouns</div>

2024

• Linguistics
• Computer Science
• Sociology

On Kratzer’s “The Notional Category of Modality”

Studies in linguistics and philosophy · 2022

• Computer Science
• Artificial Intelligence
• Linguistics

Counterfactuals to the Rescue

Oxford University Press eBooks · 2021 · 1 citations

• Epistemology
• Psychology
• Linguistics

A central question for any theory of the interpretation of conditionals is what can be held constant and what must bce given up on the face of a counterfactual supposition. This chapter brings grammatical evidence to bear on the question from polarity reversal, the phenomenon where positive polarity items can exceptionally, albeit systematically, appear in the scope of negation in the antecedent of counterfactual conditionals. Taking polarity sensitive expressions to be associated with alternatives and to give rise to scalar assertions, it shows that polarity reversal can result in scalar assertions because in making a counterfactual assumption any contextual entailments are given up once the information that gives rise to them is revised. The analysis reveals the role that contextual information tied to presuppositions plays in determining a particular type of dependency between facts.

Endorsement of inconsistent imperatives

Proceedings of the Linguistic Society of America · 2019-03-30 · 1 citations

There is an ongoing debate regarding how imperatives convey speaker endorsement. One line of approach builds it into the imperative meaning. Another posits weaker meanings. Indifference uses, like 'Go right! Go left! I don't care!', pose a challenge to the endorsement account. We reconcile the endorsement approach with such uses and argue that they can reduce to the speaker endorsing disjunctive prejacents, which results from one imperative operator taking a list of prejacents under its scope. This analysis predicts that intonational patterns that signal lists will facilitate disjunctive interpretations. We test and confirm this prediction in an experimental study.

Deontic Logic and Normative Systems - 14th International Conference, {DEON} 2018, Utrecht, The Netherlands, July 3-6, 2018.

2018-01-01 · 1 citations

Imperatives and Intonation: The Case of the Down-Stepped Level Terminal Contour

2018-01-01 · 5 citations

Anankastic conditionals are just conditionals

Semantics and Pragmatics · 2016-05-10 · 93 citations

Since Sæbø (1985, 2001) drew the attention of formal semanticists to the compositionality problems raised by anankastic conditionals like If you want to go to Harlem, you have to take the A train, a number of authors have proposed analyses tailor-made for such conditionals. We demonstrate that the seemingly puzzling properties of anankastic conditionals in fact show up independently from each other within a wider range of conditionals, which we call ‘near-anankastic’. While they do not have the means-of implication typically associated with anankastics, near-anankastics give rise to their own special additional implications. As a crucial ingredient for a unified account, we provide a new analysis of the semantics of the desire predicate in the antecedent — an issue that has not been adequately pursued in the previous literature. We claim that want has an independently motivated reading on which it predicates the existence of an action-relevant preference (Condoravdi & Lauer 2011, 2012, Lauer 2013). We then show that the semantically determined interpretation of anankastic and near-anankastic conditionals arises, predictably and compositionally, from a range of interacting factors that are at play in the interpretation of conditional sentences more generally. The special implications associated with each kind of conditional arise pragmatically. Anankastic and near-anankastic conditionals alike turn out to be just what they seem: regular, hypothetical, indicative conditionals. http://dx.doi.org/10.3765/sp.9.8 BibTeX info

Distinguishing Past, On-going, and Future Events: The EventStatus Corpus

2016-01-01 · 15 citations

Determining whether a major societal event has already happened, is still on-going, or may occur in the future is crucial for event prediction, timeline generation, and news summarization. We introduce a new task and a new corpus, EventStatus, which has 4500 English and Spanish articles about civil unrest events labeled as PAST, ON-GOING, or FU-TURE. We show that the temporal status of these events is difficult to classify because local tense and aspect cues are often lacking, time expressions are insufficient, and the linguistic contexts have rich semantic compositionality. We explore two approaches for event status classification: (1) a feature-based SVM classifier augmented with a novel induced lexicon of future-oriented verbs, such as "threatened" and "planned", and (2) a convolutional neural net. Both types of classifiers improve event status recognition over a state-of-the-art TempEval model, and our analysis offers linguistic insights into the semantic compositionality challenges for this new task.

Ignorance, indifference, and individuation with<i>wh-ever</i>

2015-04-30 · 12 citations

Abstract Wh-ever phrases give rise to implications of ignorance or indifference. Previous analyses fold these implications into the truth-conditional or presuppositional content of wh-ever and posit a variation condition requiring non-rigid reference of the wh-ever description across possibilities in a modal domain, such as the speaker’s epistemic state. This chapter shows first that variation based on non-rigidity is not strong enough to account for the precise form of the ignorance implication. Secondly, the ignorance or indifference implication is obligatory only with singular wh-ever phrases. The chapter proposes an analysis which unifies the different uses and solves both these problems. It relies on property-based individuation and associates contextually determined alternatives to wh-ever phrases which yield more specific descriptions. The ignorance and indifference implications result from the enriched meaning contributed by the alternatives and their locus of discharge.