Implementation:Pyro ppl Pyro RSA Hyperbole

Property	Value
Implementation Type	Pattern Doc
Source File	examples/rsa/hyperbole.py
Module	examples.rsa
Pyro Features	pyro.sample, pyro.factor, Search, HashingMarginal, memoize, poutine.scale, nested inference, QUD (Question Under Discussion) reasoning
Reference	Adapted from https://gscontras.github.io/probLang/chapters/03-nonliteral.html

Overview

This file implements an RSA model for interpreting hyperbole, explaining how listeners can understand non-literal numerical utterances like "It cost $10,000" (when it actually cost much less). The model captures the pragmatic reasoning that allows hyperbolic interpretation through Question Under Discussion (QUD) uncertainty.

The key innovation is that the pragmatic listener reasons about:

• The actual price (from a log-linear prior over price ranges)
• The valence (positive or negative feeling, correlated with price)
• The QUD being addressed (price, valence, both, approximate price, etc.)

Five QUD functions determine what information is relevant: exact price, valence, price+valence, approximate price, and approximate price+valence. The speaker chooses utterances with costs (precise numbers are costlier than round numbers), explaining why round numbers are used for approximate communication.

Code Reference

State = collections.namedtuple("State", ["price", "valence"])

qud_fns = {
    "price": lambda state: State(price=state.price, valence=None),
    "valence": lambda state: State(price=None, valence=state.valence),
    "priceValence": lambda state: State(price=state.price, valence=state.valence),
    "approxPrice": lambda state: State(price=approx(state.price), valence=None),
    "approxPriceValence": lambda state: State(price=approx(state.price), valence=state.valence),
}

@Marginal
def pragmatic_listener(utterance):
    price = price_prior()
    valence = valence_prior(price)
    qud = qud_prior()
    state = State(price=price, valence=valence)
    qudValue = qud_fns[qud](state)
    speaker_marginal = speaker(qudValue, qud)
    pyro.sample("speaker", speaker_marginal, obs=utterance)
    return state

I/O Contract

Parameter	Type	Description
--price	int	Utterance price value (default: 10000)
-n / --num-samples	int	Number of samples (default: 10)

Output:

• Joint distribution over (price, valence) states given the utterance
• Probabilities for each possible interpretation

State space:

• Prices: [50, 51, 500, 501, 1000, 1001, 5000, 5001, 10000, 10001]
• Valence: True (negative feeling) or False

Usage Examples

# Interpret "It cost $10,000"
# python hyperbole.py --price 10000

# The pragmatic listener will assign high probability to
# states where the actual price is much lower than $10,000
# but the valence is negative (expensive feeling)
pragmatic_marginal = pragmatic_listener(10000)
for s in pragmatic_marginal.enumerate_support():
    print(s, pragmatic_marginal.log_prob(s).exp().item())

Related Pages

• Pyro_ppl_Pyro_SearchInference - Search inference utilities
• Pyro_ppl_Pyro_RSA_Generics - RSA model for generic statements
• Pyro_ppl_Pyro_RSA_Schelling - Schelling coordination game
• Pyro_ppl_Pyro_RSA_SemanticParsing - RSA with compositional semantics