# Natural Language Toolkit: Dependency Grammars # # Copyright (C) 2001-2023 NLTK Project # Author: Jason Narad # # URL: # For license information, see LICENSE.TXT # import logging import math from nltk.parse.dependencygraph import DependencyGraph logger = logging.getLogger(__name__) ################################################################# # DependencyScorerI - Interface for Graph-Edge Weight Calculation ################################################################# class DependencyScorerI: """ A scorer for calculated the weights on the edges of a weighted dependency graph. This is used by a ``ProbabilisticNonprojectiveParser`` to initialize the edge weights of a ``DependencyGraph``. While typically this would be done by training a binary classifier, any class that can return a multidimensional list representation of the edge weights can implement this interface. As such, it has no necessary fields. """ def __init__(self): if self.__class__ == DependencyScorerI: raise TypeError("DependencyScorerI is an abstract interface") def train(self, graphs): """ :type graphs: list(DependencyGraph) :param graphs: A list of dependency graphs to train the scorer. Typically the edges present in the graphs can be used as positive training examples, and the edges not present as negative examples. """ raise NotImplementedError() def score(self, graph): """ :type graph: DependencyGraph :param graph: A dependency graph whose set of edges need to be scored. :rtype: A three-dimensional list of numbers. :return: The score is returned in a multidimensional(3) list, such that the outer-dimension refers to the head, and the inner-dimension refers to the dependencies. For instance, scores[0][1] would reference the list of scores corresponding to arcs from node 0 to node 1. The node's 'address' field can be used to determine its number identification. For further illustration, a score list corresponding to Fig.2 of Keith Hall's 'K-best Spanning Tree Parsing' paper:: scores = [[[], [5], [1], [1]], [[], [], [11], [4]], [[], [10], [], [5]], [[], [8], [8], []]] When used in conjunction with a MaxEntClassifier, each score would correspond to the confidence of a particular edge being classified with the positive training examples. """ raise NotImplementedError() ################################################################# # NaiveBayesDependencyScorer ################################################################# class NaiveBayesDependencyScorer(DependencyScorerI): """ A dependency scorer built around a MaxEnt classifier. In this particular class that classifier is a ``NaiveBayesClassifier``. It uses head-word, head-tag, child-word, and child-tag features for classification. >>> from nltk.parse.dependencygraph import DependencyGraph, conll_data2 >>> graphs = [DependencyGraph(entry) for entry in conll_data2.split('\\n\\n') if entry] >>> npp = ProbabilisticNonprojectiveParser() >>> npp.train(graphs, NaiveBayesDependencyScorer()) >>> parses = npp.parse(['Cathy', 'zag', 'hen', 'zwaaien', '.'], ['N', 'V', 'Pron', 'Adj', 'N', 'Punc']) >>> len(list(parses)) 1 """ def __init__(self): pass # Do nothing without throwing error def train(self, graphs): """ Trains a ``NaiveBayesClassifier`` using the edges present in graphs list as positive examples, the edges not present as negative examples. Uses a feature vector of head-word, head-tag, child-word, and child-tag. :type graphs: list(DependencyGraph) :param graphs: A list of dependency graphs to train the scorer. """ from nltk.classify import NaiveBayesClassifier # Create training labeled training examples labeled_examples = [] for graph in graphs: for head_node in graph.nodes.values(): for child_index, child_node in graph.nodes.items(): if child_index in head_node["deps"]: label = "T" else: label = "F" labeled_examples.append( ( dict( a=head_node["word"], b=head_node["tag"], c=child_node["word"], d=child_node["tag"], ), label, ) ) self.classifier = NaiveBayesClassifier.train(labeled_examples) def score(self, graph): """ Converts the graph into a feature-based representation of each edge, and then assigns a score to each based on the confidence of the classifier in assigning it to the positive label. Scores are returned in a multidimensional list. :type graph: DependencyGraph :param graph: A dependency graph to score. :rtype: 3 dimensional list :return: Edge scores for the graph parameter. """ # Convert graph to feature representation edges = [] for head_node in graph.nodes.values(): for child_node in graph.nodes.values(): edges.append( dict( a=head_node["word"], b=head_node["tag"], c=child_node["word"], d=child_node["tag"], ) ) # Score edges edge_scores = [] row = [] count = 0 for pdist in self.classifier.prob_classify_many(edges): logger.debug("%.4f %.4f", pdist.prob("T"), pdist.prob("F")) # smoothing in case the probability = 0 row.append([math.log(pdist.prob("T") + 0.00000000001)]) count += 1 if count == len(graph.nodes): edge_scores.append(row) row = [] count = 0 return edge_scores ################################################################# # A Scorer for Demo Purposes ################################################################# # A short class necessary to show parsing example from paper class DemoScorer(DependencyScorerI): def train(self, graphs): print("Training...") def score(self, graph): # scores for Keith Hall 'K-best Spanning Tree Parsing' paper return [ [[], [5], [1], [1]], [[], [], [11], [4]], [[], [10], [], [5]], [[], [8], [8], []], ] ################################################################# # Non-Projective Probabilistic Parsing ################################################################# class ProbabilisticNonprojectiveParser: """A probabilistic non-projective dependency parser. Nonprojective dependencies allows for "crossing branches" in the parse tree which is necessary for representing particular linguistic phenomena, or even typical parses in some languages. This parser follows the MST parsing algorithm, outlined in McDonald(2005), which likens the search for the best non-projective parse to finding the maximum spanning tree in a weighted directed graph. >>> class Scorer(DependencyScorerI): ... def train(self, graphs): ... pass ... ... def score(self, graph): ... return [ ... [[], [5], [1], [1]], ... [[], [], [11], [4]], ... [[], [10], [], [5]], ... [[], [8], [8], []], ... ] >>> npp = ProbabilisticNonprojectiveParser() >>> npp.train([], Scorer()) >>> parses = npp.parse(['v1', 'v2', 'v3'], [None, None, None]) >>> len(list(parses)) 1 Rule based example >>> from nltk.grammar import DependencyGrammar >>> grammar = DependencyGrammar.fromstring(''' ... 'taught' -> 'play' | 'man' ... 'man' -> 'the' | 'in' ... 'in' -> 'corner' ... 'corner' -> 'the' ... 'play' -> 'golf' | 'dachshund' | 'to' ... 'dachshund' -> 'his' ... ''') >>> ndp = NonprojectiveDependencyParser(grammar) >>> parses = ndp.parse(['the', 'man', 'in', 'the', 'corner', 'taught', 'his', 'dachshund', 'to', 'play', 'golf']) >>> len(list(parses)) 4 """ def __init__(self): """ Creates a new non-projective parser. """ logging.debug("initializing prob. nonprojective...") def train(self, graphs, dependency_scorer): """ Trains a ``DependencyScorerI`` from a set of ``DependencyGraph`` objects, and establishes this as the parser's scorer. This is used to initialize the scores on a ``DependencyGraph`` during the parsing procedure. :type graphs: list(DependencyGraph) :param graphs: A list of dependency graphs to train the scorer. :type dependency_scorer: DependencyScorerI :param dependency_scorer: A scorer which implements the ``DependencyScorerI`` interface. """ self._scorer = dependency_scorer self._scorer.train(graphs) def initialize_edge_scores(self, graph): """ Assigns a score to every edge in the ``DependencyGraph`` graph. These scores are generated via the parser's scorer which was assigned during the training process. :type graph: DependencyGraph :param graph: A dependency graph to assign scores to. """ self.scores = self._scorer.score(graph) def collapse_nodes(self, new_node, cycle_path, g_graph, b_graph, c_graph): """ Takes a list of nodes that have been identified to belong to a cycle, and collapses them into on larger node. The arcs of all nodes in the graph must be updated to account for this. :type new_node: Node. :param new_node: A Node (Dictionary) to collapse the cycle nodes into. :type cycle_path: A list of integers. :param cycle_path: A list of node addresses, each of which is in the cycle. :type g_graph, b_graph, c_graph: DependencyGraph :param g_graph, b_graph, c_graph: Graphs which need to be updated. """ logger.debug("Collapsing nodes...") # Collapse all cycle nodes into v_n+1 in G_Graph for cycle_node_index in cycle_path: g_graph.remove_by_address(cycle_node_index) g_graph.add_node(new_node) g_graph.redirect_arcs(cycle_path, new_node["address"]) def update_edge_scores(self, new_node, cycle_path): """ Updates the edge scores to reflect a collapse operation into new_node. :type new_node: A Node. :param new_node: The node which cycle nodes are collapsed into. :type cycle_path: A list of integers. :param cycle_path: A list of node addresses that belong to the cycle. """ logger.debug("cycle %s", cycle_path) cycle_path = self.compute_original_indexes(cycle_path) logger.debug("old cycle %s", cycle_path) logger.debug("Prior to update: %s", self.scores) for i, row in enumerate(self.scores): for j, column in enumerate(self.scores[i]): logger.debug(self.scores[i][j]) if j in cycle_path and i not in cycle_path and self.scores[i][j]: subtract_val = self.compute_max_subtract_score(j, cycle_path) logger.debug("%s - %s", self.scores[i][j], subtract_val) new_vals = [] for cur_val in self.scores[i][j]: new_vals.append(cur_val - subtract_val) self.scores[i][j] = new_vals for i, row in enumerate(self.scores): for j, cell in enumerate(self.scores[i]): if i in cycle_path and j in cycle_path: self.scores[i][j] = [] logger.debug("After update: %s", self.scores) def compute_original_indexes(self, new_indexes): """ As nodes are collapsed into others, they are replaced by the new node in the graph, but it's still necessary to keep track of what these original nodes were. This takes a list of node addresses and replaces any collapsed node addresses with their original addresses. :type new_indexes: A list of integers. :param new_indexes: A list of node addresses to check for subsumed nodes. """ swapped = True while swapped: originals = [] swapped = False for new_index in new_indexes: if new_index in self.inner_nodes: for old_val in self.inner_nodes[new_index]: if old_val not in originals: originals.append(old_val) swapped = True else: originals.append(new_index) new_indexes = originals return new_indexes def compute_max_subtract_score(self, column_index, cycle_indexes): """ When updating scores the score of the highest-weighted incoming arc is subtracted upon collapse. This returns the correct amount to subtract from that edge. :type column_index: integer. :param column_index: A index representing the column of incoming arcs to a particular node being updated :type cycle_indexes: A list of integers. :param cycle_indexes: Only arcs from cycle nodes are considered. This is a list of such nodes addresses. """ max_score = -100000 for row_index in cycle_indexes: for subtract_val in self.scores[row_index][column_index]: if subtract_val > max_score: max_score = subtract_val return max_score def best_incoming_arc(self, node_index): """ Returns the source of the best incoming arc to the node with address: node_index :type node_index: integer. :param node_index: The address of the 'destination' node, the node that is arced to. """ originals = self.compute_original_indexes([node_index]) logger.debug("originals: %s", originals) max_arc = None max_score = None for row_index in range(len(self.scores)): for col_index in range(len(self.scores[row_index])): if col_index in originals and ( max_score is None or self.scores[row_index][col_index] > max_score ): max_score = self.scores[row_index][col_index] max_arc = row_index logger.debug("%s, %s", row_index, col_index) logger.debug(max_score) for key in self.inner_nodes: replaced_nodes = self.inner_nodes[key] if max_arc in replaced_nodes: return key return max_arc def original_best_arc(self, node_index): originals = self.compute_original_indexes([node_index]) max_arc = None max_score = None max_orig = None for row_index in range(len(self.scores)): for col_index in range(len(self.scores[row_index])): if col_index in originals and ( max_score is None or self.scores[row_index][col_index] > max_score ): max_score = self.scores[row_index][col_index] max_arc = row_index max_orig = col_index return [max_arc, max_orig] def parse(self, tokens, tags): """ Parses a list of tokens in accordance to the MST parsing algorithm for non-projective dependency parses. Assumes that the tokens to be parsed have already been tagged and those tags are provided. Various scoring methods can be used by implementing the ``DependencyScorerI`` interface and passing it to the training algorithm. :type tokens: list(str) :param tokens: A list of words or punctuation to be parsed. :type tags: list(str) :param tags: A list of tags corresponding by index to the words in the tokens list. :return: An iterator of non-projective parses. :rtype: iter(DependencyGraph) """ self.inner_nodes = {} # Initialize g_graph g_graph = DependencyGraph() for index, token in enumerate(tokens): g_graph.nodes[index + 1].update( {"word": token, "tag": tags[index], "rel": "NTOP", "address": index + 1} ) # Fully connect non-root nodes in g_graph g_graph.connect_graph() original_graph = DependencyGraph() for index, token in enumerate(tokens): original_graph.nodes[index + 1].update( {"word": token, "tag": tags[index], "rel": "NTOP", "address": index + 1} ) b_graph = DependencyGraph() c_graph = DependencyGraph() for index, token in enumerate(tokens): c_graph.nodes[index + 1].update( {"word": token, "tag": tags[index], "rel": "NTOP", "address": index + 1} ) # Assign initial scores to g_graph edges self.initialize_edge_scores(g_graph) logger.debug(self.scores) # Initialize a list of unvisited vertices (by node address) unvisited_vertices = [vertex["address"] for vertex in c_graph.nodes.values()] # Iterate over unvisited vertices nr_vertices = len(tokens) betas = {} while unvisited_vertices: # Mark current node as visited current_vertex = unvisited_vertices.pop(0) logger.debug("current_vertex: %s", current_vertex) # Get corresponding node n_i to vertex v_i current_node = g_graph.get_by_address(current_vertex) logger.debug("current_node: %s", current_node) # Get best in-edge node b for current node best_in_edge = self.best_incoming_arc(current_vertex) betas[current_vertex] = self.original_best_arc(current_vertex) logger.debug("best in arc: %s --> %s", best_in_edge, current_vertex) # b_graph = Union(b_graph, b) for new_vertex in [current_vertex, best_in_edge]: b_graph.nodes[new_vertex].update( {"word": "TEMP", "rel": "NTOP", "address": new_vertex} ) b_graph.add_arc(best_in_edge, current_vertex) # Beta(current node) = b - stored for parse recovery # If b_graph contains a cycle, collapse it cycle_path = b_graph.contains_cycle() if cycle_path: # Create a new node v_n+1 with address = len(nodes) + 1 new_node = {"word": "NONE", "rel": "NTOP", "address": nr_vertices + 1} # c_graph = Union(c_graph, v_n+1) c_graph.add_node(new_node) # Collapse all nodes in cycle C into v_n+1 self.update_edge_scores(new_node, cycle_path) self.collapse_nodes(new_node, cycle_path, g_graph, b_graph, c_graph) for cycle_index in cycle_path: c_graph.add_arc(new_node["address"], cycle_index) # self.replaced_by[cycle_index] = new_node['address'] self.inner_nodes[new_node["address"]] = cycle_path # Add v_n+1 to list of unvisited vertices unvisited_vertices.insert(0, nr_vertices + 1) # increment # of nodes counter nr_vertices += 1 # Remove cycle nodes from b_graph; B = B - cycle c for cycle_node_address in cycle_path: b_graph.remove_by_address(cycle_node_address) logger.debug("g_graph: %s", g_graph) logger.debug("b_graph: %s", b_graph) logger.debug("c_graph: %s", c_graph) logger.debug("Betas: %s", betas) logger.debug("replaced nodes %s", self.inner_nodes) # Recover parse tree logger.debug("Final scores: %s", self.scores) logger.debug("Recovering parse...") for i in range(len(tokens) + 1, nr_vertices + 1): betas[betas[i][1]] = betas[i] logger.debug("Betas: %s", betas) for node in original_graph.nodes.values(): # TODO: It's dangerous to assume that deps it a dictionary # because it's a default dictionary. Ideally, here we should not # be concerned how dependencies are stored inside of a dependency # graph. node["deps"] = {} for i in range(1, len(tokens) + 1): original_graph.add_arc(betas[i][0], betas[i][1]) logger.debug("Done.") yield original_graph ################################################################# # Rule-based Non-Projective Parser ################################################################# class NonprojectiveDependencyParser: """ A non-projective, rule-based, dependency parser. This parser will return the set of all possible non-projective parses based on the word-to-word relations defined in the parser's dependency grammar, and will allow the branches of the parse tree to cross in order to capture a variety of linguistic phenomena that a projective parser will not. """ def __init__(self, dependency_grammar): """ Creates a new ``NonprojectiveDependencyParser``. :param dependency_grammar: a grammar of word-to-word relations. :type dependency_grammar: DependencyGrammar """ self._grammar = dependency_grammar def parse(self, tokens): """ Parses the input tokens with respect to the parser's grammar. Parsing is accomplished by representing the search-space of possible parses as a fully-connected directed graph. Arcs that would lead to ungrammatical parses are removed and a lattice is constructed of length n, where n is the number of input tokens, to represent all possible grammatical traversals. All possible paths through the lattice are then enumerated to produce the set of non-projective parses. param tokens: A list of tokens to parse. type tokens: list(str) return: An iterator of non-projective parses. rtype: iter(DependencyGraph) """ # Create graph representation of tokens self._graph = DependencyGraph() for index, token in enumerate(tokens): self._graph.nodes[index] = { "word": token, "deps": [], "rel": "NTOP", "address": index, } for head_node in self._graph.nodes.values(): deps = [] for dep_node in self._graph.nodes.values(): if ( self._grammar.contains(head_node["word"], dep_node["word"]) and head_node["word"] != dep_node["word"] ): deps.append(dep_node["address"]) head_node["deps"] = deps # Create lattice of possible heads roots = [] possible_heads = [] for i, word in enumerate(tokens): heads = [] for j, head in enumerate(tokens): if (i != j) and self._grammar.contains(head, word): heads.append(j) if len(heads) == 0: roots.append(i) possible_heads.append(heads) # Set roots to attempt if len(roots) < 2: if len(roots) == 0: for i in range(len(tokens)): roots.append(i) # Traverse lattice analyses = [] for _ in roots: stack = [] analysis = [[] for i in range(len(possible_heads))] i = 0 forward = True while i >= 0: if forward: if len(possible_heads[i]) == 1: analysis[i] = possible_heads[i][0] elif len(possible_heads[i]) == 0: analysis[i] = -1 else: head = possible_heads[i].pop() analysis[i] = head stack.append([i, head]) if not forward: index_on_stack = False for stack_item in stack: if stack_item[0] == i: index_on_stack = True orig_length = len(possible_heads[i]) if index_on_stack and orig_length == 0: for j in range(len(stack) - 1, -1, -1): stack_item = stack[j] if stack_item[0] == i: possible_heads[i].append(stack.pop(j)[1]) elif index_on_stack and orig_length > 0: head = possible_heads[i].pop() analysis[i] = head stack.append([i, head]) forward = True if i + 1 == len(possible_heads): analyses.append(analysis[:]) forward = False if forward: i += 1 else: i -= 1 # Filter parses # ensure 1 root, every thing has 1 head for analysis in analyses: if analysis.count(-1) > 1: # there are several root elements! continue graph = DependencyGraph() graph.root = graph.nodes[analysis.index(-1) + 1] for address, (token, head_index) in enumerate( zip(tokens, analysis), start=1 ): head_address = head_index + 1 node = graph.nodes[address] node.update({"word": token, "address": address}) if head_address == 0: rel = "ROOT" else: rel = "" graph.nodes[head_index + 1]["deps"][rel].append(address) # TODO: check for cycles yield graph ################################################################# # Demos ################################################################# def demo(): # hall_demo() nonprojective_conll_parse_demo() rule_based_demo() def hall_demo(): npp = ProbabilisticNonprojectiveParser() npp.train([], DemoScorer()) for parse_graph in npp.parse(["v1", "v2", "v3"], [None, None, None]): print(parse_graph) def nonprojective_conll_parse_demo(): from nltk.parse.dependencygraph import conll_data2 graphs = [DependencyGraph(entry) for entry in conll_data2.split("\n\n") if entry] npp = ProbabilisticNonprojectiveParser() npp.train(graphs, NaiveBayesDependencyScorer()) for parse_graph in npp.parse( ["Cathy", "zag", "hen", "zwaaien", "."], ["N", "V", "Pron", "Adj", "N", "Punc"] ): print(parse_graph) def rule_based_demo(): from nltk.grammar import DependencyGrammar grammar = DependencyGrammar.fromstring( """ 'taught' -> 'play' | 'man' 'man' -> 'the' | 'in' 'in' -> 'corner' 'corner' -> 'the' 'play' -> 'golf' | 'dachshund' | 'to' 'dachshund' -> 'his' """ ) print(grammar) ndp = NonprojectiveDependencyParser(grammar) graphs = ndp.parse( [ "the", "man", "in", "the", "corner", "taught", "his", "dachshund", "to", "play", "golf", ] ) print("Graphs:") for graph in graphs: print(graph) if __name__ == "__main__": demo()