Computational Pragmatics

Core Idea

Computational pragmatics applies computational methods to pragmatic phenomena: modeling how context determines meaning, how implicatures are computed, how speakers vary utterances relative to audience and context, and how irony, metaphor, and indirect speech acts are processed. This bridges formal pragmatics, cognitive modeling, and NLP. Systems must model shared knowledge, discourse structure, and common ground — challenging problems in AI because they require representing complex, dynamic context.

How It's Best Learned

Study computational models of implicature computation (e.g., Rational Speech Acts framework). Examine NLP systems for indirect request recognition, sarcasm detection, and pragmatic inference. Learn how language models capture pragmatic intuitions. Understand limitations of current systems in context-dependent tasks. Explore questions: How are alternatives modeled? How do systems represent common ground? What pragmatic phenomena are computational tractable vs. intractable?

Common Misconceptions

• Thinking pragmatics is too context-dependent for formal modeling; computational frameworks show surprising tractability.
• Assuming language models automatically capture pragmatics; most successful pragmatic phenomena modeling requires explicit context mechanisms.

Explainer

Pragmatics studies how context determines meaning — how the same utterance "It's cold" means different things depending on whether you're in a freezing car or a cool room, whether you're asking for a sweater or complaining about the air conditioning. Traditional linguistics has often sidelined pragmatics as too context-dependent for formal study, but computational pragmatics shows that context effects are partially formalizable and computationally tractable.

Several core problems in computational pragmatics:

Implicature computation: When a speaker says "Some students passed," listeners infer "Not all students passed" (scalar implicature). Computationally, this requires enumerating alternatives ("Some," "All," "None"), reasoning about why the speaker chose the weaker alternative, and inferring the stronger meaning. This requires models of rationality and information structure — not trivial computationally.

Reference resolution and common ground: Pronouns and definite descriptions refer based on context. "It's raining; you should bring an umbrella" — the "it" refers to weather because context makes that salient. Computationally, systems must track discourse entities, their salience, and mutual knowledge. Systems that don't model common ground fail at reference.

Indirect speech acts and context-dependent interpretation: "Can you pass the salt?" is not a question about ability but a polite request. Interpretation depends on social context (formality, relationship), physical context (is the salt nearby?), and pragmatic reasoning (why would the speaker ask this?). Computational models must represent these contexts.

Irony and sarcasm: "Great job," said when someone makes a mistake, is sarcastic — it means the opposite. Detection requires recognizing that literal meaning contradicts expected context. Models can learn patterns (certain words + negative context → likely sarcasm) but real pragmatic understanding is deeper.

The Rational Speech Acts (RSA) framework provides one formalizable approach. The basic idea:

• The speaker chooses utterances that informatively communicate the intended meaning, given what they believe the listener will infer
• The listener interprets utterances by reasoning about what rational speaker would say given their beliefs
• This creates a loop of reasoning that resolves implicatures and contextual effects

Computationally, RSA requires:

1. Enumerate alternatives to the utterance

2. For each alternative, compute the probability a rational speaker would choose it

3. For each interpretation, compute how likely it is given speaker rationality

4. Iterate reasoning (listener reasons about speaker's reasoning about listener's reasoning...)

Modern language models (like GPT) learn pragmatic patterns from massive text, and they often perform well on pragmatic tasks. But there are limits: multi-step reasoning, novel contexts, and deep pragmatic understanding remain challenging. Models capture surface patterns but may not compute pragmatic meaning the way humans do.

The future of computational pragmatics involves:

• Better models of common ground and context
• Integration of pragmatic reasoning with formal semantics
• Handling of under-specified, ambiguous utterances
• Modeling of dialectal and register variation in pragmatic norms

Computational pragmatics shows that while pragmatics is context-dependent, it's not entirely intractable. Systematic models can formalize important aspects and make progress on real language understanding.