Experimental Detection of Embedded Implicatures

Emmanuel Chemla, Benjamin Spector

TL;DR

This paper presents experimental evidence for "embedded implicatures," a phenomenon where pragmatic inferences (like "some but not all") can be triggered by linguistic operators, challenging traditional Gricean views. The authors introduce a novel experimental paradigm using graded judgments and carefully designed stimuli to distinguish between local and global interpretations of scalar implicatures, providing crucial data for the ongoing debate between localist and globalist theories.

What Problem Does This Paper Address?

The paper tackles a long-standing debate in linguistics and philosophy of language concerning the nature of "scalar implicatures." These are inferences we make beyond the literal meaning of words, such as inferring "some but not all" from the statement "some students passed." Traditionally, these inferences were thought to arise from general pragmatic reasoning (globalism). However, a newer "grammatical" view suggests these inferences can be generated locally, meaning they can be influenced by linguistic operators within a sentence, similar to how negation or quantifiers work.

The core problem is distinguishing between these two theories. While simple sentences don't offer clear evidence, complex sentences with embedded operators are key. If implicatures are purely pragmatic (globalist), they should only apply to the entire sentence. If they can be grammatical (localist), they can be triggered under an operator. Previous experimental work by Geurts and Pouscoulous (2009) claimed to find no evidence for these "embedded" or "local" implicatures, suggesting globalism. Chemla and Spector aim to refute this by providing new experimental data and a more sensitive methodology.

Methods

The authors employ a sentence-picture matching task, a standard method in experimental semantics. However, they introduce two key innovations:

  • Graded Judgments: Instead of asking participants to label a sentence as strictly "true" or "false," they used a continuous scale (a cursor moving from "false" to "true"). This allows for more nuanced data, capturing degrees of truth or acceptability, which they hypothesize can reveal subtle readings that might be missed with binary judgments. This is particularly important given the "charity principle," where subjects might default to the weakest reading if multiple are available.
  • Carefully Designed Stimuli: They created specific image-picture pairs designed to isolate different readings (literal, global, and local). Crucially, in their second experiment, they focused on "non-monotonic environments" (e.g., "Exactly one letter is connected to some of its circles"). In such environments, the local reading does not entail the global or literal reading, making it uniquely detectable. This contrasts with earlier experiments where the local reading was a subset of the global reading, making it harder to isolate.

Datasets

The paper does not describe training datasets in the machine learning sense. Instead, it details two experiments using novel sets of stimuli created by the authors:

  • Experiment 1: Used sentence-picture pairs to test scalar items ('some', 'or') embedded under universal quantifiers. The stimuli were designed to create four conditions: "false" (no reading true), "literal" (only literal reading true), "weak" (literal and global true, local false), and "strong" (all readings true).
  • Experiment 2: Focused on non-monotonic environments. The stimuli were designed to create four conditions: "false" (no reading true), "literal" (only literal true), "local" (only local true), and "all" (all three readings true).

These stimuli were presented to native French speakers.

Key Insights and Ideas

  • Local Implicatures Exist: The most significant finding is that participants do assign higher truth values to sentences when the local reading is true, even when it's not entailed by the global or literal reading. This directly contradicts the claims of Geurts and Pouscoulous (2009).
  • Graded Judgments are Powerful: The use of a continuous scale was crucial for detecting these subtle readings. The authors argue that binary judgments might have masked the existence of local implicatures due to the charity principle.
  • Non-Monotonic Environments are Key: The second experiment, focusing on non-monotonic environments, provided the strongest evidence. By creating conditions where the local reading was the only true reading, they could isolate its contribution.
  • Support for Grammatical Theories: The existence of genuinely local implicatures, especially in non-monotonic contexts, is difficult to explain with purely pragmatic (globalist) theories. It lends strong support to grammatical theories that allow for covert operators to generate these inferences locally.

Key Results

  • Experiment 1: Participants rated sentences significantly higher when the local reading was true (strong condition) compared to when it was false (weak condition), even though the only difference was the truth of the local implicature. This suggests the local reading is available.
  • Experiment 2:
    • In the "local" condition (where only the local reading was true), participants gave significantly higher ratings than in the "false" condition.
  • Crucially, the "local" condition was rated significantly higher than the "literal" condition, indicating that the local reading is not just an alternative but a distinct and often preferred interpretation.
  • The "all" condition (where literal, global, and local were true) received the highest ratings, supporting the idea that more true readings lead to higher acceptability.
  • Control Conditions: In downward-entailing (DE) environments, where local implicatures are generally disfavored, participants showed much lower ratings for the "local" condition compared to the "both" condition (where literal and local were true), suggesting they could distinguish between contexts where local implicatures are expected and those where they are not.

Limitations and Future Work

  • Charity Principle: While the graded judgment method aims to mitigate the charity principle, the authors acknowledge that it might still play a role in influencing how easily local readings are detected.
  • Generalizability: The experiments focused on specific scalar items ('some', 'or') and universal quantifiers. Future work could explore other scalar items and different types of embedding environments.
  • Cross-Linguistic Variation: The experiments were conducted with French speakers. Further research is needed to determine if these findings generalize to other languages.
  • Acquisition: The paper doesn't discuss how children acquire these embedded implicatures, which is a significant area for future research.

The authors suggest that their findings provide strong evidence for the existence of embedded scalar implicatures, which are best explained by grammatical theories of implicature.

Summary updated

Abstract

According to the Gricean approach to scalar implicatures (SIs for short), SIs are pragmatic inferences that result from a reasoning about the speaker’s communicative intentions. In recent years, an alternative view of SIs (let us call it the ‘grammatical view’ of SIs) has been put forward, according to which they result from the optional presence of a covert so-called exhaustivity operator in the logical form of the relevant sentences and are thus reducible to standard semantic entailment (see Chierchia 2006, Fox 2007, Chierchia, Fox, and Spector in press, a.o, building on earlier grammatical approaches by, e.g., Landman 1998, Chierchia 2004).
BibTeX
@inproceedings{chemla2009experimental,
  author = {Emmanuel Chemla and Benjamin Spector},
  title = {{Experimental Detection of Embedded Implicatures}},
  booktitle = {Amsterdam Colloquium on Logic, Language and Meaning},
  year = {2009},
  doi = {10.1007/978-3-642-14287-1\_6},
}
Figure