The challenge of modeling scalar implicatures computationally is that:
AImplicatures are purely subjective and cannot be formalized
BComputing implicatures requires representing and reasoning about alternatives, speaker rationality, and listener expectations — all context-dependent and computationally complex
CImplicatures appear only in speech, not written language
DImplicatures are not meaningful phenomena
Scalar implicatures (e.g., 'some' implicating 'not all') require reasoning about alternatives and rational speaker behavior. Computational models must represent all logically stronger alternatives, compute listener beliefs about speaker rationality, and infer what the speaker intended to communicate. This is formally tractable but computationally complex, especially with multi-level reasoning.
Question 2 Multiple Choice
Why is modeling common ground (shared knowledge between speaker and listener) critical for computational pragmatics?
ABecause pragmatics is about grammar, not context
BBecause implicatures, reference resolution, and contextual interpretation all depend on what speaker and listener know and believe is known in common
CBecause common ground is unchanging and predetermined
DBecause pragmatics is irrelevant to understanding meaning
Common ground (or context) affects interpretation profoundly. 'It's raining' is a casual observation if the listener sees the rain; it's a warning if the listener doesn't. Pronouns, demonstratives, indirect speech acts, and implicatures all depend on common ground. Systems must track and update common ground dynamically.
Question 3 True / False
Language models like GPT demonstrate human-level pragmatic competence because they match human judgments on pragmatic inference tasks.
TTrue
FFalse
Answer: False
Language models show impressive performance on some pragmatic tasks (sarcasm detection, indirect request recognition) but often lack deep pragmatic understanding. They capture surface patterns from training data but may not truly compute pragmatic inferences. When context is novel or reasoning is multi-step, models often fail. They're useful tools but not full solutions.
Question 4 True / False
Modeling sarcasm computationally is fundamentally impossible because sarcasm is fundamentally subjective and context-dependent.
TTrue
FFalse
Answer: False
While sarcasm is context-dependent, computational models have made progress. Sarcasm often involves a contrast between literal and expected meaning; models can learn these patterns. Detection accuracy is lower than literal speech, but not chance. Computational approaches to sarcasm are imperfect but meaningful and improving.
Question 5 Short Answer
Explain why the Rational Speech Acts framework is useful for computational pragmatics and what it models.
Think about your answer, then reveal below.
Model answer: The RSA framework models pragmatic meaning through recursive reasoning: the speaker chooses utterances that are informative and relevant given rational listener inference; the listener infers the speaker's intent given rational speaker behavior. Computationally, this involves computing alternatives, reasoning about listener beliefs, and iteratively refining both. It provides a formal, tractable framework for implicature and context effects.
RSA bridges pragmatic theory and computation. It makes formal assumptions explicit and tractable, enabling modeling of empirical phenomena. Iterations of reasoning (speaker reasoning about listener expectations, listener reasoning about speaker rationality) capture how pragmatic meaning emerges.