# Natural Language Toolkit: Relation Extraction # # Copyright (C) 2001-2012 NLTK Project # Author: Ewan Klein # URL: # For license information, see LICENSE.TXT """ Code for extracting relational triples from the ieer and conll2002 corpora. Relations are stored internally as dictionaries ('reldicts'). The two serialization outputs are "rtuple" and "clause". - An rtuple is a tuple of the form ``(subj, filler, obj)``, where ``subj`` and ``obj`` are pairs of Named Entity mentions, and ``filler`` is the string of words occurring between ``sub`` and ``obj`` (with no intervening NEs). Strings are printed via ``repr()`` to circumvent locale variations in rendering utf-8 encoded strings. - A clause is an atom of the form ``relsym(subjsym, objsym)``, where the relation, subject and object have been canonicalized to single strings. """ # todo: get a more general solution to canonicalized symbols for clauses -- maybe use xmlcharrefs? from collections import defaultdict from string import join import re import htmlentitydefs from itertools import ifilter # Dictionary that associates corpora with NE classes NE_CLASSES = { 'ieer': ['LOCATION', 'ORGANIZATION', 'PERSON', 'DURATION', 'DATE', 'CARDINAL', 'PERCENT', 'MONEY', 'MEASURE'], 'conll2002': ['LOC', 'PER', 'ORG'], 'ace': ['LOCATION', 'ORGANIZATION', 'PERSON', 'DURATION', 'DATE', 'CARDINAL', 'PERCENT', 'MONEY', 'MEASURE', 'FACILITY', 'GPE'], } # Allow abbreviated class labels short2long = dict(LOC = 'LOCATION', ORG = 'ORGANIZATION', PER = 'PERSON') long2short = dict(LOCATION ='LOC', ORGANIZATION = 'ORG', PERSON = 'PER') def _expand(type): """ Expand an NE class name. :type type: str :rtype: str """ try: return short2long[type] except KeyError: return type def class_abbrev(type): """ Abbreviate an NE class name. :type type: str :rtype: str """ try: return long2short[type] except KeyError: return type def _join(lst, sep=' ', untag=False): """ Join a list into a string, turning tags tuples into tag strings or just words. :param untag: if ``True``, omit the tag from tagged input strings. :type lst: list :rtype: str """ try: return join(lst, sep=sep) except TypeError: if untag: return join([tup[0] for tup in lst], sep=sep) from nltk.tag import tuple2str return join([tuple2str(tup) for tup in lst], sep=sep) def descape_entity(m, defs=htmlentitydefs.entitydefs): """ Translate one entity to its ISO Latin value. Inspired by example from effbot.org """ #s = 'mcglashan_&_sarrail' #l = ['mcglashan', '&', 'sarrail'] #pattern = re.compile("&(\w+?);") #new = list2sym(l) #s = pattern.sub(descape_entity, s) #print s, new try: return defs[m.group(1)] except KeyError: return m.group(0) # use as is def list2sym(lst): """ Convert a list of strings into a canonical symbol. :type lst: list :return: a Unicode string without whitespace :rtype: unicode """ sym = _join(lst, '_', untag=True) sym = sym.lower() ENT = re.compile("&(\w+?);") sym = ENT.sub(descape_entity, sym) sym = sym.replace('.', '') return sym def mk_pairs(tree): """ Group a chunk structure into a list of pairs of the form (list(str), ``Tree``) In order to facilitate the construction of (``Tree``, string, ``Tree``) triples, this identifies pairs whose first member is a list (possibly empty) of terminal strings, and whose second member is a ``Tree`` of the form (NE_label, terminals). :param tree: a chunk tree :return: a list of pairs (list(str), ``Tree``) :rtype: list of tuple """ from nltk.tree import Tree pairs = [] pair = [[], None] for dtr in tree: if not isinstance(dtr, Tree): pair[0].append(dtr) else: # dtr is a Tree pair[1] = dtr pairs.append(pair) pair = [[], None] return pairs def mk_reldicts(pairs, window=5, trace=0): """ Converts the pairs generated by ``mk_pairs`` into a 'reldict': a dictionary which stores information about the subject and object NEs plus the filler between them. Additionally, a left and right context of length =< window are captured (within a given input sentence). :param pairs: a pair of list(str) and ``Tree``, as generated by :param window: a threshold for the number of items to include in the left and right context :type window: int :return: 'relation' dictionaries whose keys are 'lcon', 'subjclass', 'subjtext', 'subjsym', 'filler', objclass', objtext', 'objsym' and 'rcon' :rtype: list(defaultdict) """ result = [] while len(pairs) > 2: reldict = defaultdict(str) reldict['lcon'] = _join(pairs[0][0][-window:]) reldict['subjclass'] = pairs[0][1].node reldict['subjtext'] = _join(pairs[0][1].leaves()) reldict['subjsym'] = list2sym(pairs[0][1].leaves()) reldict['filler'] = _join(pairs[1][0]) reldict['objclass'] = pairs[1][1].node reldict['objtext'] = _join(pairs[1][1].leaves()) reldict['objsym'] = list2sym(pairs[1][1].leaves()) reldict['rcon'] = _join(pairs[2][0][:window]) if trace: print "(rel(%s, %s)" % (reldict['subjclass'], reldict['objclass']) result.append(reldict) pairs = pairs[1:] return result def extract_rels(subjclass, objclass, doc, corpus='ace', pattern=None, window=10): """ Filter the output of ``mk_reldicts`` according to specified NE classes and a filler pattern. The parameters ``subjclass`` and ``objclass`` can be used to restrict the Named Entities to particular types (any of 'LOCATION', 'ORGANIZATION', 'PERSON', 'DURATION', 'DATE', 'CARDINAL', 'PERCENT', 'MONEY', 'MEASURE'). :param subjclass: the class of the subject Named Entity. :type subjclass: str :param objclass: the class of the object Named Entity. :type objclass: str :param doc: input document :type doc: ieer document or a list of chunk trees :param corpus: name of the corpus to take as input; possible values are 'ieer' and 'conll2002' :type corpus: str :param pattern: a regular expression for filtering the fillers of retrieved triples. :type pattern: SRE_Pattern :param window: filters out fillers which exceed this threshold :type window: int :return: see ``mk_reldicts`` :rtype: list(defaultdict) """ if subjclass and subjclass not in NE_CLASSES[corpus]: if _expand(subjclass) in NE_CLASSES[corpus]: subjclass = _expand(subjclass) else: raise ValueError, "your value for the subject type has not been recognized: %s" % subjclass if objclass and objclass not in NE_CLASSES[corpus]: if _expand(objclass) in NE_CLASSES[corpus]: objclass = _expand(objclass) else: raise ValueError, "your value for the object type has not been recognized: %s" % objclass if corpus == 'ace' or corpus == 'conll2002': pairs = mk_pairs(doc) elif corpus == 'ieer': pairs = mk_pairs(doc.text) + mk_pairs(doc.headline) else: raise ValueError, "corpus type not recognized" reldicts = mk_reldicts(pairs) relfilter = lambda x: (x['subjclass'] == subjclass and len(x['filler'].split()) <= window and pattern.match(x['filler']) and x['objclass'] == objclass) return filter(relfilter, reldicts) def show_raw_rtuple(reldict, lcon=False, rcon=False): """ Pretty print the reldict as an rtuple. :param reldict: a relation dictionary :type reldict: defaultdict """ items = [class_abbrev(reldict['subjclass']), reldict['subjtext'], reldict['filler'], class_abbrev(reldict['objclass']), reldict['objtext']] format = '[%s: %r] %r [%s: %r]' if lcon: items = [reldict['lcon']] + items format = '...%r)' + format if rcon: items.append(reldict['rcon']) format = format + '(%r...' printargs = tuple(items) return format % printargs def show_clause(reldict, relsym): """ Print the relation in clausal form. :param reldict: a relation dictionary :type reldict: defaultdict :param relsym: a label for the relation :type relsym: str """ items = (relsym, reldict['subjsym'], reldict['objsym']) return "%s(%r, %r)" % items ####################################################### # Demos of relation extraction with regular expressions ####################################################### ############################################ # Example of in(ORG, LOC) ############################################ def in_demo(trace=0, sql=True): """ Select pairs of organizations and locations whose mentions occur with an intervening occurrence of the preposition "in". If the sql parameter is set to True, then the entity pairs are loaded into an in-memory database, and subsequently pulled out using an SQL "SELECT" query. """ from nltk.corpus import ieer if sql: try: import sqlite3 connection = sqlite3.connect(":memory:") connection.text_factory = sqlite3.OptimizedUnicode cur = connection.cursor() cur.execute("""create table Locations (OrgName text, LocationName text, DocID text)""") except ImportError: import warnings warnings.warn("Cannot import sqlite; sql flag will be ignored.") IN = re.compile(r'.*\bin\b(?!\b.+ing)') print print "IEER: in(ORG, LOC) -- just the clauses:" print "=" * 45 for file in ieer.fileids(): for doc in ieer.parsed_docs(file): if trace: print doc.docno print "=" * 15 for rel in extract_rels('ORG', 'LOC', doc, corpus='ieer', pattern=IN): print show_clause(rel, relsym='IN') if sql: try: rtuple = (rel['subjtext'], rel['objtext'], doc.docno) cur.execute("""insert into Locations values (?, ?, ?)""", rtuple) connection.commit() except NameError: pass if sql: try: cur.execute("""select OrgName from Locations where LocationName = 'Atlanta'""") print print "Extract data from SQL table: ORGs in Atlanta" print "-" * 15 for row in cur: print row except NameError: pass ############################################ # Example of has_role(PER, LOC) ############################################ def roles_demo(trace=0): from nltk.corpus import ieer roles = """ (.*( # assorted roles analyst| chair(wo)?man| commissioner| counsel| director| economist| editor| executive| foreman| governor| head| lawyer| leader| librarian).*)| manager| partner| president| producer| professor| researcher| spokes(wo)?man| writer| ,\sof\sthe?\s* # "X, of (the) Y" """ ROLES = re.compile(roles, re.VERBOSE) print print "IEER: has_role(PER, ORG) -- raw rtuples:" print "=" * 45 for file in ieer.fileids(): for doc in ieer.parsed_docs(file): lcon = rcon = False if trace: print doc.docno print "=" * 15 lcon = rcon = True for rel in extract_rels('PER', 'ORG', doc, corpus='ieer', pattern=ROLES): print show_raw_rtuple(rel, lcon=lcon, rcon=rcon) ############################################## ### Show what's in the IEER Headlines ############################################## def ieer_headlines(): from nltk.corpus import ieer from nltk.tree import Tree print "IEER: First 20 Headlines" print "=" * 45 trees = [doc.headline for file in ieer.fileids() for doc in ieer.parsed_docs(file)] for tree in trees[:20]: print print "%s:\n%s" % (doc.docno, tree) ############################################# ## Dutch CONLL2002: take_on_role(PER, ORG ############################################# def conllned(trace=1): """ Find the copula+'van' relation ('of') in the Dutch tagged training corpus from CoNLL 2002. """ from nltk.corpus import conll2002 vnv = """ ( is/V| # 3rd sing present and was/V| # past forms of the verb zijn ('be') werd/V| # and also present wordt/V # past of worden ('become) ) .* # followed by anything van/Prep # followed by van ('of') """ VAN = re.compile(vnv, re.VERBOSE) print print "Dutch CoNLL2002: van(PER, ORG) -- raw rtuples with context:" print "=" * 45 for doc in conll2002.chunked_sents('ned.train'): lcon = rcon = False if trace: lcon = rcon = True for rel in extract_rels('PER', 'ORG', doc, corpus='conll2002', pattern=VAN, window=10): print show_raw_rtuple(rel, lcon=True, rcon=True) ############################################# ## Spanish CONLL2002: (PER, ORG) ############################################# def conllesp(): from nltk.corpus import conll2002 de = """ .* ( de/SP| del/SP ) """ DE = re.compile(de, re.VERBOSE) print print "Spanish CoNLL2002: de(ORG, LOC) -- just the first 10 clauses:" print "=" * 45 rels = [rel for doc in conll2002.chunked_sents('esp.train') for rel in extract_rels('ORG', 'LOC', doc, corpus='conll2002', pattern = DE)] for r in rels[:10]: print show_clause(r, relsym='DE') print def ne_chunked(): IN = re.compile(r'.*\bin\b(?!\b.+ing)') rels = [] for sent in nltk.corpus.treebank.tagged_sents()[:100]: sent = nltk.ne_chunk(sent) print extract_rels('ORG', 'LOC', sent, corpus='ace', pattern = IN) if __name__ == '__main__': import nltk from nltk.sem import relextract in_demo(trace=0) roles_demo(trace=0) conllned() conllesp() ieer_headlines()