# Natural Language Toolkit: CONLL Corpus Reader # # Copyright (C) 2001-2012 NLTK Project # Author: Steven Bird # Edward Loper # URL: # For license information, see LICENSE.TXT """ Read CoNLL-style chunk fileids. """ import os import codecs import textwrap from nltk.tree import Tree from nltk.util import LazyMap, LazyConcatenation from util import * from api import * class ConllCorpusReader(CorpusReader): """ A corpus reader for CoNLL-style files. These files consist of a series of sentences, separated by blank lines. Each sentence is encoded using a table (or "grid") of values, where each line corresponds to a single word, and each column corresponds to an annotation type. The set of columns used by CoNLL-style files can vary from corpus to corpus; the ``ConllCorpusReader`` constructor therefore takes an argument, ``columntypes``, which is used to specify the columns that are used by a given corpus. @todo: Add support for reading from corpora where different parallel files contain different columns. @todo: Possibly add caching of the grid corpus view? This would allow the same grid view to be used by different data access methods (eg words() and parsed_sents() could both share the same grid corpus view object). @todo: Better support for -DOCSTART-. Currently, we just ignore it, but it could be used to define methods that retrieve a document at a time (eg parsed_documents()). """ #///////////////////////////////////////////////////////////////// # Column Types #///////////////////////////////////////////////////////////////// WORDS = 'words' #: column type for words POS = 'pos' #: column type for part-of-speech tags TREE = 'tree' #: column type for parse trees CHUNK = 'chunk' #: column type for chunk structures NE = 'ne' #: column type for named entities SRL = 'srl' #: column type for semantic role labels IGNORE = 'ignore' #: column type for column that should be ignored #: A list of all column types supported by the conll corpus reader. COLUMN_TYPES = (WORDS, POS, TREE, CHUNK, NE, SRL, IGNORE) #///////////////////////////////////////////////////////////////// # Constructor #///////////////////////////////////////////////////////////////// def __init__(self, root, fileids, columntypes, chunk_types=None, top_node='S', pos_in_tree=False, srl_includes_roleset=True, encoding=None, tree_class=Tree, tag_mapping_function=None): for columntype in columntypes: if columntype not in self.COLUMN_TYPES: raise ValueError('Bad column type %r' % columntype) if isinstance(chunk_types, basestring): chunk_types = [chunk_types] self._chunk_types = chunk_types self._colmap = dict((c,i) for (i,c) in enumerate(columntypes)) self._pos_in_tree = pos_in_tree self._top_node = top_node # for chunks self._srl_includes_roleset = srl_includes_roleset self._tree_class = tree_class CorpusReader.__init__(self, root, fileids, encoding) self._tag_mapping_function = tag_mapping_function #///////////////////////////////////////////////////////////////// # Data Access Methods #///////////////////////////////////////////////////////////////// def raw(self, fileids=None): if fileids is None: fileids = self._fileids elif isinstance(fileids, basestring): fileids = [fileids] return concat([self.open(f).read() for f in fileids]) def words(self, fileids=None): self._require(self.WORDS) return LazyConcatenation(LazyMap(self._get_words, self._grids(fileids))) def sents(self, fileids=None): self._require(self.WORDS) return LazyMap(self._get_words, self._grids(fileids)) def tagged_words(self, fileids=None, simplify_tags=False): self._require(self.WORDS, self.POS) def get_tagged_words(grid): return self._get_tagged_words(grid, simplify_tags) return LazyConcatenation(LazyMap(get_tagged_words, self._grids(fileids))) def tagged_sents(self, fileids=None, simplify_tags=False): self._require(self.WORDS, self.POS) def get_tagged_words(grid): return self._get_tagged_words(grid, simplify_tags) return LazyMap(get_tagged_words, self._grids(fileids)) def chunked_words(self, fileids=None, chunk_types=None, simplify_tags=False): self._require(self.WORDS, self.POS, self.CHUNK) if chunk_types is None: chunk_types = self._chunk_types def get_chunked_words(grid): # capture chunk_types as local var return self._get_chunked_words(grid, chunk_types, simplify_tags) return LazyConcatenation(LazyMap(get_chunked_words, self._grids(fileids))) def chunked_sents(self, fileids=None, chunk_types=None, simplify_tags=False): self._require(self.WORDS, self.POS, self.CHUNK) if chunk_types is None: chunk_types = self._chunk_types def get_chunked_words(grid): # capture chunk_types as local var return self._get_chunked_words(grid, chunk_types, simplify_tags) return LazyMap(get_chunked_words, self._grids(fileids)) def parsed_sents(self, fileids=None, pos_in_tree=None, simplify_tags=False): self._require(self.WORDS, self.POS, self.TREE) if pos_in_tree is None: pos_in_tree = self._pos_in_tree def get_parsed_sent(grid): # capture pos_in_tree as local var return self._get_parsed_sent(grid, pos_in_tree, simplify_tags) return LazyMap(get_parsed_sent, self._grids(fileids)) def srl_spans(self, fileids=None): self._require(self.SRL) return LazyMap(self._get_srl_spans, self._grids(fileids)) def srl_instances(self, fileids=None, pos_in_tree=None, flatten=True): self._require(self.WORDS, self.POS, self.TREE, self.SRL) if pos_in_tree is None: pos_in_tree = self._pos_in_tree def get_srl_instances(grid): # capture pos_in_tree as local var return self._get_srl_instances(grid, pos_in_tree) result = LazyMap(get_srl_instances, self._grids(fileids)) if flatten: result = LazyConcatenation(result) return result def iob_words(self, fileids=None, simplify_tags=False): """ :return: a list of word/tag/IOB tuples :rtype: list(tuple) :param fileids: the list of fileids that make up this corpus :type fileids: None or str or list """ self._require(self.WORDS, self.POS, self.CHUNK) def get_iob_words(grid): return self._get_iob_words(grid, simplify_tags) return LazyConcatenation(LazyMap(get_iob_words, self._grids(fileids))) def iob_sents(self, fileids=None, simplify_tags=False): """ :return: a list of lists of word/tag/IOB tuples :rtype: list(list) :param fileids: the list of fileids that make up this corpus :type fileids: None or str or list """ self._require(self.WORDS, self.POS, self.CHUNK) def get_iob_words(grid): return self._get_iob_words(grid, simplify_tags) return LazyMap(get_iob_words, self._grids(fileids)) #///////////////////////////////////////////////////////////////// # Grid Reading #///////////////////////////////////////////////////////////////// def _grids(self, fileids=None): # n.b.: we could cache the object returned here (keyed on # fileids), which would let us reuse the same corpus view for # different things (eg srl and parse trees). return concat([StreamBackedCorpusView(fileid, self._read_grid_block, encoding=enc) for (fileid, enc) in self.abspaths(fileids, True)]) def _read_grid_block(self, stream): grids = [] for block in read_blankline_block(stream): block = block.strip() if not block: continue grid = [line.split() for line in block.split('\n')] # If there's a docstart row, then discard. ([xx] eventually it # would be good to actually use it) if grid[0][self._colmap.get('words', 0)] == '-DOCSTART-': del grid[0] # Check that the grid is consistent. for row in grid: if len(row) != len(grid[0]): raise ValueError('Inconsistent number of columns:\n%s' % block) grids.append(grid) return grids #///////////////////////////////////////////////////////////////// # Transforms #///////////////////////////////////////////////////////////////// # given a grid, transform it into some representation (e.g., # a list of words or a parse tree). def _get_words(self, grid): return self._get_column(grid, self._colmap['words']) def _get_tagged_words(self, grid, simplify_tags=False): pos_tags = self._get_column(grid, self._colmap['pos']) if simplify_tags: pos_tags = [self._tag_mapping_function(t) for t in pos_tags] return zip(self._get_column(grid, self._colmap['words']), pos_tags) def _get_iob_words(self, grid, simplify_tags=False): pos_tags = self._get_column(grid, self._colmap['pos']) if simplify_tags: pos_tags = [self._tag_mapping_function(t) for t in pos_tags] return zip(self._get_column(grid, self._colmap['words']), pos_tags, self._get_column(grid, self._colmap['chunk'])) def _get_chunked_words(self, grid, chunk_types, simplify_tags=False): # n.b.: this method is very similar to conllstr2tree. words = self._get_column(grid, self._colmap['words']) pos_tags = self._get_column(grid, self._colmap['pos']) if simplify_tags: pos_tags = [self._tag_mapping_function(t) for t in pos_tags] chunk_tags = self._get_column(grid, self._colmap['chunk']) stack = [Tree(self._top_node, [])] for (word, pos_tag, chunk_tag) in zip(words, pos_tags, chunk_tags): if chunk_tag == 'O': state, chunk_type = 'O', '' else: (state, chunk_type) = chunk_tag.split('-') # If it's a chunk we don't care about, treat it as O. if chunk_types is not None and chunk_type not in chunk_types: state = 'O' # Treat a mismatching I like a B. if state == 'I' and chunk_type != stack[-1].node: state = 'B' # For B or I: close any open chunks if state in 'BO' and len(stack) == 2: stack.pop() # For B: start a new chunk. if state == 'B': new_chunk = Tree(chunk_type, []) stack[-1].append(new_chunk) stack.append(new_chunk) # Add the word token. stack[-1].append((word, pos_tag)) return stack[0] def _get_parsed_sent(self, grid, pos_in_tree, simplify_tags=False): words = self._get_column(grid, self._colmap['words']) pos_tags = self._get_column(grid, self._colmap['pos']) if simplify_tags: pos_tags = [self._tag_mapping_function(t) for t in pos_tags] parse_tags = self._get_column(grid, self._colmap['tree']) treestr = '' for (word, pos_tag, parse_tag) in zip(words, pos_tags, parse_tags): if word == '(': word = '-LRB-' if word == ')': word = '-RRB-' if pos_tag == '(': pos_tag = '-LRB-' if pos_tag == ')': pos_tag = '-RRB-' (left, right) = parse_tag.split('*') right = right.count(')')*')' # only keep ')'. treestr += '%s (%s %s) %s' % (left, pos_tag, word, right) try: tree = self._tree_class.parse(treestr) except (ValueError, IndexError): tree = self._tree_class.parse('(%s %s)' % (self._top_node, treestr)) if not pos_in_tree: for subtree in tree.subtrees(): for i, child in enumerate(subtree): if (isinstance(child, nltk.Tree) and len(child)==1 and isinstance(child[0], basestring)): subtree[i] = (child[0], child.node) return tree def _get_srl_spans(self, grid): """ list of list of (start, end), tag) tuples """ if self._srl_includes_roleset: predicates = self._get_column(grid, self._colmap['srl']+1) start_col = self._colmap['srl']+2 else: predicates = self._get_column(grid, self._colmap['srl']) start_col = self._colmap['srl']+1 # Count how many predicates there are. This tells us how many # columns to expect for SRL data. num_preds = len([p for p in predicates if p != '-']) spanlists = [] for i in range(num_preds): col = self._get_column(grid, start_col+i) spanlist = [] stack = [] for wordnum, srl_tag in enumerate(col): (left, right) = srl_tag.split('*') for tag in left.split('('): if tag: stack.append((tag, wordnum)) for i in range(right.count(')')): (tag, start) = stack.pop() spanlist.append( ((start, wordnum+1), tag) ) spanlists.append(spanlist) return spanlists def _get_srl_instances(self, grid, pos_in_tree): tree = self._get_parsed_sent(grid, pos_in_tree) spanlists = self._get_srl_spans(grid) if self._srl_includes_roleset: predicates = self._get_column(grid, self._colmap['srl']+1) rolesets = self._get_column(grid, self._colmap['srl']) else: predicates = self._get_column(grid, self._colmap['srl']) rolesets = [None] * len(predicates) instances = ConllSRLInstanceList(tree) for wordnum, predicate in enumerate(predicates): if predicate == '-': continue # Decide which spanlist to use. Don't assume that they're # sorted in the same order as the predicates (even though # they usually are). for spanlist in spanlists: for (start, end), tag in spanlist: if wordnum in range(start,end) and tag in ('V', 'C-V'): break else: continue break else: raise ValueError('No srl column found for %r' % predicate) instances.append(ConllSRLInstance(tree, wordnum, predicate, rolesets[wordnum], spanlist)) return instances #///////////////////////////////////////////////////////////////// # Helper Methods #///////////////////////////////////////////////////////////////// def _require(self, *columntypes): for columntype in columntypes: if columntype not in self._colmap: raise ValueError('This corpus does not contain a %s ' 'column.' % columntype) @staticmethod def _get_column(grid, column_index): return [grid[i][column_index] for i in range(len(grid))] class ConllSRLInstance(object): """ An SRL instance from a CoNLL corpus, which identifies and providing labels for the arguments of a single verb. """ # [xx] add inst.core_arguments, inst.argm_arguments? def __init__(self, tree, verb_head, verb_stem, roleset, tagged_spans): self.verb = [] """A list of the word indices of the words that compose the verb whose arguments are identified by this instance. This will contain multiple word indices when multi-word verbs are used (e.g. 'turn on').""" self.verb_head = verb_head """The word index of the head word of the verb whose arguments are identified by this instance. E.g., for a sentence that uses the verb 'turn on,' ``verb_head`` will be the word index of the word 'turn'.""" self.verb_stem = verb_stem self.roleset = roleset self.arguments = [] """A list of ``(argspan, argid)`` tuples, specifying the location and type for each of the arguments identified by this instance. ``argspan`` is a tuple ``start, end``, indicating that the argument consists of the ``words[start:end]``.""" self.tagged_spans = tagged_spans """A list of ``(span, id)`` tuples, specifying the location and type for each of the arguments, as well as the verb pieces, that make up this instance.""" self.tree = tree """The parse tree for the sentence containing this instance.""" self.words = tree.leaves() """A list of the words in the sentence containing this instance.""" # Fill in the self.verb and self.arguments values. for (start, end), tag in tagged_spans: if tag in ('V', 'C-V'): self.verb += range(start, end) else: self.arguments.append( ((start, end), tag) ) def __repr__(self): plural = len(self.arguments)!=1 and 's' or '' return '' % ( (self.verb_stem, len(self.arguments), plural)) def pprint(self): verbstr = ' '.join(self.words[i][0] for i in self.verb) hdr = 'SRL for %r (stem=%r):\n' % (verbstr, self.verb_stem) s = '' for i, word in enumerate(self.words): if isinstance(word, tuple): word = word[0] for (start, end), argid in self.arguments: if i == start: s += '[%s ' % argid if i == end: s += '] ' if i in self.verb: word = '<<%s>>' % word s += word + ' ' return hdr + textwrap.fill(s.replace(' ]', ']'), initial_indent=' ', subsequent_indent=' ') class ConllSRLInstanceList(list): """ Set of instances for a single sentence """ def __init__(self, tree, instances=()): self.tree = tree list.__init__(self, instances) def __str__(self): return self.pprint() def pprint(self, include_tree=False): # Sanity check: trees should be the same for inst in self: if inst.tree != self.tree: raise ValueError('Tree mismatch!') # If desired, add trees: if include_tree: words = self.tree.leaves() pos = [None] * len(words) synt = ['*'] * len(words) self._tree2conll(self.tree, 0, words, pos, synt) s = '' for i in range(len(words)): # optional tree columns if include_tree: s += '%-20s ' % words[i] s += '%-8s ' % pos[i] s += '%15s*%-8s ' % tuple(synt[i].split('*')) # verb head column for inst in self: if i == inst.verb_head: s += '%-20s ' % inst.verb_stem break else: s += '%-20s ' % '-' # Remaining columns: self for inst in self: argstr = '*' for (start, end), argid in inst.tagged_spans: if i==start: argstr = '(%s%s' % (argid, argstr) if i==(end-1): argstr += ')' s += '%-12s ' % argstr s += '\n' return s def _tree2conll(self, tree, wordnum, words, pos, synt): assert isinstance(tree, Tree) if len(tree) == 1 and isinstance(tree[0], basestring): pos[wordnum] = tree.node assert words[wordnum] == tree[0] return wordnum+1 elif len(tree) == 1 and isinstance(tree[0], tuple): assert len(tree[0]) == 2 pos[wordnum], pos[wordnum] = tree[0] return wordnum+1 else: synt[wordnum] = '(%s%s' % (tree.node, synt[wordnum]) for child in tree: wordnum = self._tree2conll(child, wordnum, words, pos, synt) synt[wordnum-1] += ')' return wordnum class ConllChunkCorpusReader(ConllCorpusReader): """ A ConllCorpusReader whose data file contains three columns: words, pos, and chunk. """ def __init__(self, root, fileids, chunk_types, encoding=None, tag_mapping_function=None): ConllCorpusReader.__init__( self, root, fileids, ('words', 'pos', 'chunk'), chunk_types=chunk_types, encoding=encoding, tag_mapping_function=tag_mapping_function)