Home

Awesome

Bringing machine learning and compositional semantics together

Demonstration code for the paper

Liang, Percy and Christopher Potts. 2014. Bringing machine learning and compositional semantics together. Annual Review of Linguistics 1(1): 355–376.

The purpose of the code is just to illustrate how the algorithms work, as an aid to understanding the paper and developing new models that synthesize compositionality and machine learning.

All of the files contain detailed explanations and documentation, with cross references to the paper. evenodd.py, grammar.py, synthesis.py run demos corresponding to examples and discussions from the paper.

The current verson has been tested with Python 2.7 and Python 3.5. The only other requirement is Numpy 1.10 or greater, and that dependency is only for the neural network in distributed.py.

Code snippets

Build the gold-standard grammar and use it for parsing:

from grammar import Grammar, gold_lexicon, rules, functions
gram = Grammar(gold_lexicon, rules, functions)
gram.gen('minus two plus three')

And interpretation:

for lf in gram.gen('minus two plus three'):
    print gram.sem(lf)

Check out the crazy things that the crude grammar does with the simple example (it creates 486 logical forms):

from synthesis import crude_lexicon
crude_gram = Grammar(crude_lexicon, rules, functions)
crude_gram.gen('minus two plus three')

Train a semantic parsing model using the default training set from semdata.py and the crude grammar as a starting point:

from semdata import sem_train
from synthesis import phi_sem
from learning import SGD
# For semantic parsing, the denotations are ignored:
semparse_train = [[x,y] for x, y, d in sem_train]

# The space of output classes is determined by GEN:
weights = SGD(D=semparse_train, phi=phi_sem, classes=crude_gram.gen)

And now see that the crude grammar plus the learned weights deliver a right result for 'minus two plus three' (the training set happens to favor the second parse in the list returned by the gold grammar gram):

from learning import predict
predict(x='minus two plus three', w=weights, phi=phi_sem, classes=crude_gram.gen)

For semantic parsing with the trees/derivations as latent variables, and for learning from denotations, see synthesis.evaluate_interpretive and synthesis.evaluate_latent_semparse.

grammar.py

Implements a simple interpreted context-free grammar formalism in which each nonterminal node is a tuple (s, r) where s is the syntactic category and r is the logical form representation. The user supplies a lexicon, a rule-set, and possibly a set of functions that, together with Python itself, make the logical forms interpretable as Python code.

semdata.py

Uses grammar.py to create training and testing data for the semantic models.

learning.py

The core learning framework from section 3.2 of the paper.

evenodd.py

A simple supervised learning example using learning.py. The examples correspond to table 3 in the paper.

synthesis.py

Implements three different theories for learning compositional semantics. All are illustrated with the same feature function and train/test data.