# Natural Language Toolkit: Viterbi Probabilistic Parser # # Copyright (C) 2001-2010 NLTK Project # Author: Edward Loper # Steven Bird # URL: # For license information, see LICENSE.TXT from nltk.tree import Tree, ProbabilisticTree from api import * ##////////////////////////////////////////////////////// ## Viterbi PCFG Parser ##////////////////////////////////////////////////////// class ViterbiParser(ParserI): """ A bottom-up C{PCFG} parser that uses dynamic programming to find the single most likely parse for a text. The C{ViterbiParser} parser parses texts by filling in a X{most likely constituent table}. This table records the most probable tree representation for any given span and node value. In particular, it has an entry for every start index, end index, and node value, recording the most likely subtree that spans from the start index to the end index, and has the given node value. The C{ViterbiParser} parser fills in this table incrementally. It starts by filling in all entries for constituents that span one element of text (i.e., entries where the end index is one greater than the start index). After it has filled in all table entries for constituents that span one element of text, it fills in the entries for constitutants that span two elements of text. It continues filling in the entries for constituents spanning larger and larger portions of the text, until the entire table has been filled. Finally, it returns the table entry for a constituent spanning the entire text, whose node value is the grammar's start symbol. In order to find the most likely constituent with a given span and node value, the C{ViterbiParser} parser considers all productions that could produce that node value. For each production, it finds all children that collectively cover the span and have the node values specified by the production's right hand side. If the probability of the tree formed by applying the production to the children is greater than the probability of the current entry in the table, then the table is updated with this new tree. A pseudo-code description of the algorithm used by C{ViterbiParser} is: - Create an empty most likely constituent table, M{MLC}. - For M{width} in 1...len(M{text}): - For M{start} in 1...len(M{text})-M{width}: - For M{prod} in grammar.productions: - For each sequence of subtrees [M{t[1]}, M{t[2]}, ..., M{t[n]}] in M{MLC}, where M{t[i]}.node==M{prod}.rhs[i], and the sequence covers [M{start}:M{start}+M{width}]: - M{old_p} = M{MLC}[M{start}, M{start+width}, M{prod}.lhs] - M{new_p} = P(M{t[1]})*P(M{t[1]})*...*P(M{t[n]})*P(M{prod}) - if M{new_p} > M{old_p}: - M{new_tree} = Tree(M{prod}.lhs, M{t[1]}, M{t[2]}, ..., M{t[n]}) - M{MLC}[M{start}, M{start+width}, M{prod}.lhs] = M{new_tree} - Return M{MLC}[0, len(M{text}), M{start_symbol}] @type _grammar: C{WeightedGrammar} @ivar _grammar: The grammar used to parse sentences. @type _trace: C{int} @ivar _trace: The level of tracing output that should be generated when parsing a text. """ def __init__(self, grammar, trace=0): """ Create a new C{ViterbiParser} parser, that uses {grammar} to parse texts. @type grammar: C{WeightedGrammar} @param grammar: The grammar used to parse texts. @type trace: C{int} @param trace: The level of tracing that should be used when parsing a text. C{0} will generate no tracing output; and higher numbers will produce more verbose tracing output. """ self._grammar = grammar self._trace = trace def grammar(self): return self._grammar def trace(self, trace=2): """ Set the level of tracing output that should be generated when parsing a text. @type trace: C{int} @param trace: The trace level. A trace level of C{0} will generate no tracing output; and higher trace levels will produce more verbose tracing output. @rtype: C{None} """ self._trace = trace def nbest_parse(self, tokens, n=None): # Inherit docs from ParserI tokens = list(tokens) self._grammar.check_coverage(tokens) # The most likely constituent table. This table specifies the # most likely constituent for a given span and type. # Constituents can be either Trees or tokens. For # Trees, the "type" is the Nonterminal for the tree's # root node value. For Tokens, the "type" is the token's # type. The table is stored as a dictionary, since it is # sparse. constituents = {} # Initialize the constituents dictionary with the words from # the text. if self._trace: print ('Inserting tokens into the most likely'+ ' constituents table...') for index in range(len(tokens)): token = tokens[index] constituents[index,index+1,token] = token if self._trace > 1: self._trace_lexical_insertion(token, index, len(tokens)) # Consider each span of length 1, 2, ..., n; and add any trees # that might cover that span to the constituents dictionary. for length in range(1, len(tokens)+1): if self._trace: print ('Finding the most likely constituents'+ ' spanning %d text elements...' % length) #print constituents for start in range(len(tokens)-length+1): span = (start, start+length) self._add_constituents_spanning(span, constituents, tokens) # Find all trees that span the entire text & have the right cat trees = [constituents.get((0, len(tokens), self._grammar.start()), [])] # Sort the trees, and return the requested number of them. trees.sort(lambda t1,t2: cmp(t2.prob(), t1.prob())) return trees[:n] def _add_constituents_spanning(self, span, constituents, tokens): """ Find any constituents that might cover C{span}, and add them to the most likely constituents table. @rtype: C{None} @type span: C{(int, int)} @param span: The section of the text for which we are trying to find possible constituents. The span is specified as a pair of integers, where the first integer is the index of the first token that should be included in the constituent; and the second integer is the index of the first token that should not be included in the constituent. I.e., the constituent should cover C{M{text}[span[0]:span[1]]}, where C{M{text}} is the text that we are parsing. @type constituents: C{dictionary} from C{(int,int,Nonterminal)} to (C{ProbabilisticToken} or C{ProbabilisticTree}). @param constituents: The most likely constituents table. This table records the most probable tree representation for any given span and node value. In particular, C{constituents(M{s},M{e},M{nv})} is the most likely C{ProbabilisticTree} that covers C{M{text}[M{s}:M{e}]} and has a node value C{M{nv}.symbol()}, where C{M{text}} is the text that we are parsing. When C{_add_constituents_spanning} is called, C{constituents} should contain all possible constituents that are shorter than C{span}. @type tokens: C{list} of tokens @param tokens: The text we are parsing. This is only used for trace output. """ # Since some of the grammar productions may be unary, we need to # repeatedly try all of the productions until none of them add any # new constituents. changed = 1 while changed: changed = 0 # Find all ways instantiations of the grammar productions that # cover the span. instantiations = self._find_instantiations(span, constituents) # For each production instantiation, add a new # ProbabilisticTree whose probability is the product # of the childrens' probabilities and the production's # probability. for (production, children) in instantiations: subtrees = [c for c in children if isinstance(c, Tree)] p = reduce(lambda pr,t:pr*t.prob(), subtrees, production.prob()) node = production.lhs().symbol() tree = ProbabilisticTree(node, children, prob=p) # If it's new a constituent, then add it to the # constituents dictionary. c = constituents.get((span[0], span[1], production.lhs()), None) if self._trace > 1: if c is None or c != tree: if c is None or c.prob() < tree.prob(): print ' Insert:', else: print ' Discard:', self._trace_production(production, p, span, len(tokens)) if c is None or c.prob() < tree.prob(): constituents[span[0], span[1], production.lhs()] = tree changed = 1 def _find_instantiations(self, span, constituents): """ @return: a list of the production instantiations that cover a given span of the text. A X{production instantiation} is a tuple containing a production and a list of children, where the production's right hand side matches the list of children; and the children cover C{span}. @rtype: C{list} of C{pair} of C{Production}, (C{list} of (C{ProbabilisticTree} or token. @type span: C{(int, int)} @param span: The section of the text for which we are trying to find production instantiations. The span is specified as a pair of integers, where the first integer is the index of the first token that should be covered by the production instantiation; and the second integer is the index of the first token that should not be covered by the production instantiation. @type constituents: C{dictionary} from C{(int,int,Nonterminal)} to (C{ProbabilisticToken} or C{ProbabilisticTree}). @param constituents: The most likely constituents table. This table records the most probable tree representation for any given span and node value. See the module documentation for more information. """ rv = [] for production in self._grammar.productions(): childlists = self._match_rhs(production.rhs(), span, constituents) for childlist in childlists: rv.append( (production, childlist) ) return rv def _match_rhs(self, rhs, span, constituents): """ @return: a set of all the lists of children that cover C{span} and that match C{rhs}. @rtype: C{list} of (C{list} of C{ProbabilisticTree} or C{Token}) @type rhs: C{list} of C{Nonterminal} or (any) @param rhs: The list specifying what kinds of children need to cover C{span}. Each nonterminal in C{rhs} specifies that the corresponding child should be a tree whose node value is that nonterminal's symbol. Each terminal in C{rhs} specifies that the corresponding child should be a token whose type is that terminal. @type span: C{(int, int)} @param span: The section of the text for which we are trying to find child lists. The span is specified as a pair of integers, where the first integer is the index of the first token that should be covered by the child list; and the second integer is the index of the first token that should not be covered by the child list. @type constituents: C{dictionary} from C{(int,int,Nonterminal)} to (C{ProbabilisticToken} or C{ProbabilisticTree}). @param constituents: The most likely constituents table. This table records the most probable tree representation for any given span and node value. See the module documentation for more information. """ (start, end) = span # Base case if start >= end and rhs == (): return [[]] if start >= end or rhs == (): return [] # Find everything that matches the 1st symbol of the RHS childlists = [] for split in range(start, end+1): l=constituents.get((start,split,rhs[0])) if l is not None: rights = self._match_rhs(rhs[1:], (split,end), constituents) childlists += [[l]+r for r in rights] return childlists def _trace_production(self, production, p, span, width): """ Print trace output indicating that a given production has been applied at a given location. @param production: The production that has been applied @type production: C{Production} @param p: The probability of the tree produced by the production. @type p: C{float} @param span: The span of the production @type span: C{tuple} @rtype: C{None} """ str = '|' + '.' * span[0] str += '=' * (span[1] - span[0]) str += '.' * (width - span[1]) + '| ' str += '%s' % production if self._trace > 2: str = '%-40s %12.10f ' % (str, p) print str def _trace_lexical_insertion(self, token, index, width): str = ' Insert: |' + '.' * index + '=' + '.' * (width-index-1) + '| ' str += '%s' % (token,) print str def __repr__(self): return '' % self._grammar ##////////////////////////////////////////////////////// ## Test Code ##////////////////////////////////////////////////////// def demo(): """ A demonstration of the probabilistic parsers. The user is prompted to select which demo to run, and how many parses should be found; and then each parser is run on the same demo, and a summary of the results are displayed. """ import sys, time import nltk from nltk import tokenize from nltk.parse import ViterbiParser # Define two demos. Each demo has a sentence and a grammar. demos = [('I saw the man with my telescope', nltk.toy_pcfg1), ('the boy saw Jack with Bob under the table with a telescope', nltk.toy_pcfg2)] # Ask the user which demo they want to use. print for i in range(len(demos)): print '%3s: %s' % (i+1, demos[i][0]) print ' %r' % demos[i][1] print print 'Which demo (%d-%d)? ' % (1, len(demos)), try: snum = int(sys.stdin.readline().strip())-1 sent, grammar = demos[snum] except: print 'Bad sentence number' return # Tokenize the sentence. tokens = sent.split() parser = ViterbiParser(grammar) all_parses = {} print '\nsent: %s\nparser: %s\ngrammar: %s' % (sent,parser,grammar) parser.trace(3) t = time.time() parses = parser.nbest_parse(tokens) time = time.time()-t if parses: average = reduce(lambda a,b:a+b.prob(), parses, 0)/len(parses) else: average = 0 num_parses = len(parses) for p in parses: all_parses[p.freeze()] = 1 # Print some summary statistics print print 'Time (secs) # Parses Average P(parse)' print '-----------------------------------------' print '%11.4f%11d%19.14f' % (time, num_parses, average) parses = all_parses.keys() if parses: p = reduce(lambda a,b:a+b.prob(), parses, 0)/len(parses) else: p = 0 print '------------------------------------------' print '%11s%11d%19.14f' % ('n/a', len(parses), p) # Ask the user if we should draw the parses. print print 'Draw parses (y/n)? ', if sys.stdin.readline().strip().lower().startswith('y'): from nltk.draw.tree import draw_trees print ' please wait...' draw_trees(*parses) # Ask the user if we should print the parses. print print 'Print parses (y/n)? ', if sys.stdin.readline().strip().lower().startswith('y'): for parse in parses: print parse if __name__ == '__main__': demo()