def test(grammarText, sentences):
"""Test the coverage of a CFG grammar.
grammarText -- the grammar string
sentences -- a list of sentences to test, with invalid ones prefixed with '*'
"""
valid_sentences = [s for s in sentences if s[0] != '*']
invalid_sentences = [s[1:] for s in sentences if s[0] == '*']
parser = ChartParser(parse_cfg(grammarText))
for sentence in valid_sentences:
parses = parser.nbest_parse(sentence.split())
print(sentence + "\n" + "\n".join(map(str, parses)) + "\n")
assert parses, "Valid sentence failed to parse."
for sentence in invalid_sentences:
print("*" + sentence)
parses = parser.nbest_parse(sentence.split())
assert parses == [], "Invalid sentence parsed successfully."
def parse(sentence, cfg):
grammar = parse_cfg(cfg)
parser = ChartParser(grammar, TD_STRATEGY)
words = sentence.split()
draw.draw_trees(*parser.get_parse_list(words))
def guess(self, verbose=None):
"""
Makes a guess based on the next observation.
Updates self._curr_guess.
:rtype: CFG
:returns: The next guess
"""
if verbose is not None:
self._verbose = verbose
sentence = Sentence(next(self._text))
self._num_steps += 1
self._log("String {}: {}".format(self._num_steps, sentence))
if sentence in self._data:
self._log("String already seen")
return self._curr_guess
# Info from previous guess
num_contexts = len(self._contexts)
num_subs = len(self._substrings)
if self._curr_guess is not None:
num_nts = len(set(p.lhs()
for p in self._curr_guess.productions())) - 1
else:
num_nts = 0
total_timer = Timer()
total_timer.start()
# Update data and terminals
words = sentence.get_words()
self._data.add(sentence)
self._terminals.update(set(words))
# Update contexts
self._log("Updating contexts...")
inds = range(0, len(words) + 1)
contexts = [
Context(words[:i], words[j:]) for i in inds for j in inds[i:]
]
self._contexts.update(ContextSet(contexts))
self._log(
"{} new contexts added".format(len(self._contexts) - num_contexts))
# Update substrings
self._log("Updating substrings...")
is_new_sentence = True
if self._curr_guess_parser is not None:
try:
parses = self._curr_guess_parser.parse(words)
is_new_sentence = len(list(parses)) == 0
except:
is_new_sentence = True
if is_new_sentence:
subs = [Sentence(words[i:j]) for i in inds for j in inds[i:]]
self._substrings.update(SentenceSet(subs))
self._log("{} new substrings added".format(
len(self._substrings) - num_subs))
else:
self._log("Sentence already generated by current guess")
# Construct the nonterminals
self._log("Constructing nonterminals...")
kernels = set()
for i in range(1, self._k + 1):
subsets = [
SentenceSet(j) for j in combinations(self._substrings, i)
]
kernels.update(subsets)
for kernel in kernels:
if kernel not in self._nonterminals:
nt_name = self._new_name()
contexts = self._oracle.restr_right_triangle(
kernel, self._contexts)
nt = Nonterminal(nt_name)
self._nonterminals[kernel] = nt
self._nt_contexts[nt] = contexts
# Get a set of nonterminals with unique contexts
self._log("Removing equivalent nonterminals...")
context_nts = {con: nt for nt, con in self._nt_contexts.iteritems()}
self._log(
"{} nonterminals removed".format(len(kernels) - len(context_nts)))
self._log("{} new nonterminals constructed".format(
len(context_nts) - num_nts))
# Construct the rules
self._log("Constructing rules...")
self._productions = set()
timer = Timer()
# Lexical rules
timer.start()
for t in self._terminals:
t_kernel = SentenceSet([Sentence([t])])
t_nt = self._nonterminals[t_kernel]
t_contexts = self._nt_contexts[t_nt]
for contexts, nt in context_nts.iteritems():
rule = Production(nt, [t])
if rule in self._productions:
continue
if rule in self._eliminated_rules:
continue
if contexts.issubset(t_contexts):
self._productions.add(rule)
else:
self._eliminated_rules.add(rule)
timer.stop()
num_lex = len(self._productions)
self._log("{} lexical rules ({:.2f} secs)".format(
num_lex, timer.elapsed()))
# Binary rules
timer.reset()
timer.start()
for kernel_l in self._nonterminals:
for kernel_r in self._nonterminals:
kernel_rhs = kernel_l + kernel_r
sents_rhs = list(kernel_rhs.intersection(self._substrings))
inds = range(len(sents_rhs) / self._k + 1)
kers_rhs = [
sents_rhs[self._k * i:self._k * (i + 1)] for i in inds
]
kers_rhs = [SentenceSet(k) for k in kers_rhs if len(k) > 0]
nts_rhs = [self._nonterminals[k] for k in kers_rhs]
contexts_nts_rhs = [self._nt_contexts[nt] for nt in nts_rhs]
if len(contexts_nts_rhs) > 0:
contexts_rhs = contexts_nts_rhs[0].intersection(
*contexts_nts_rhs)
else:
contexts_rhs = self._contexts
# Membership queries
new_strs_rhs = kernel_rhs.difference(SentenceSet(sents_rhs))
new_contexts_rhs = self._oracle.restr_right_triangle(
new_strs_rhs, contexts_rhs)
contexts_rhs.intersection_update(new_contexts_rhs)
# Building the rules
for contexts, nt in context_nts.iteritems():
nt_l = context_nts[self._nt_contexts[
self._nonterminals[kernel_l]]]
nt_r = context_nts[self._nt_contexts[
self._nonterminals[kernel_r]]]
rule = Production(nt, [nt_l, nt_r])
if rule in self._productions:
continue
if rule in self._eliminated_rules:
continue
if contexts.issubset(contexts_rhs):
self._productions.add(rule)
else:
self._eliminated_rules.add(rule)
timer.stop()
num_bin = len(self._productions) - num_lex
self._log("{} binary rules ({:.2f} secs)".format(
num_bin, timer.elapsed()))
# Start rules
timer.reset()
timer.start()
for contexts, nt in context_nts.iteritems():
rule = Production(self._start_symbol, [nt])
if rule in self._productions:
continue
if rule in self._eliminated_rules:
continue
if Context([], []) in contexts:
self._productions.add(rule)
else:
self._eliminated_rules.add(rule)
timer.stop()
num_start = len(self._productions) - num_lex - num_bin
self._log("{} start rules ({:.2f} secs)".format(
num_start, timer.elapsed()))
# Construct the grammar
self._curr_guess = CFG(self._start_symbol, self._productions)
self._curr_guess_parser = ChartParser(self._curr_guess)
total_timer.stop()
elapsed = total_timer.elapsed()
num_rules = len(self._curr_guess.productions())
self._log("Constructed grammar with {} rules ({:.2f} secs)".format(
num_rules, elapsed))
return self._curr_guess
