Implementation:Pyro ppl Pyro RSA SemanticParsing

Property	Value
Implementation Type	Pattern Doc
Source File	examples/rsa/semantic_parsing.py
Module	examples.rsa
Pyro Features	pyro.sample, pyro.factor, BestFirstSearch, HashingMarginal, memoize, nested inference
Reference	Adapted from http://dippl.org/examples/zSemanticPragmaticMashup.html

Overview

This file implements a Rational Speech Act (RSA) model that combines pragmatic reasoning with CCG-based compositional semantics. It demonstrates how Pyro can be used for cognitive science models of language understanding, where speakers and listeners reason about each other recursively.

The implementation includes:

• Lexical semantics: Word meanings are represented as classes (BlondMeaning, NiceMeaning, etc.) with sem() (semantic function) and syn() (syntactic category) methods following CCG (Combinatory Categorial Grammar).
• Compositional semantics: Words are combined using CCG rules -- function application based on syntactic types (left/right application).
• World model: A world consists of objects with probabilistic properties (blond, nice, tall), sampled via pyro.sample.
• Pragmatic reasoning: A literal listener interprets utterances, a speaker chooses utterances based on the literal listener, and an RSA listener reasons about the speaker.

The inference uses BestFirstSearch, which enumerates program executions ordered by probability, combined with HashingMarginal to create marginal distributions.

Code Reference

@Marginal(num_samples=100)
def literal_listener(utterance):
    m = meaning(utterance)
    world = world_prior(2, m)
    pyro.factor("world_constraint", heuristic(m(world)) * 1000)
    return world

@Marginal(num_samples=100)
def speaker(world):
    utterance = utterance_prior()
    L = literal_listener(utterance)
    pyro.sample("speaker_constraint", L, obs=world)
    return utterance

def rsa_listener(utterance, qud):
    world = world_prior(2, meaning(utterance))
    S = speaker(world)
    pyro.sample("listener_constraint", S, obs=utterance)
    return qud(world)

I/O Contract

Parameter	Type	Description
utterance	str	Natural language sentence (e.g., "all blond people are nice")
qud	callable	Question Under Discussion function mapping world to answer
-n / --num-samples	int	Number of samples for BestFirstSearch (default: 10)

Output:

• HashingMarginal distribution over world states or QUD answers

Key sample sites:

• {name}_blond, {name}_nice, {name}_tall: Object properties (Bernoulli)
• ix_{c}: CCG combination choices (Categorical)
• utterance: Speaker utterance choice (Categorical)

Usage Examples

# Compute literal listener interpretation
mll = Marginal(literal_listener_raw, num_samples=100)

def is_any_qud(world):
    return any(map(lambda obj: obj.nice, world))

result = mll("all blond people are nice", is_any_qud)
print(result())

# Compute RSA pragmatic listener interpretation
rsa = Marginal(rsa_listener, num_samples=100)
result = rsa("some of the blond people are nice", is_all_qud)
print(result())

Related Pages

• Pyro_ppl_Pyro_SearchInference - Search inference utilities used by RSA models
• Pyro_ppl_Pyro_RSA_Generics - RSA model for generic statements
• Pyro_ppl_Pyro_RSA_Hyperbole - RSA model for hyperbole
• Pyro_ppl_Pyro_RSA_Schelling - Schelling coordination game