def make_tweet_entities_csv_for_turk(): twitter_site = short_text_websites.get_twitter_site() entities_to_evaluate = entity_dataset_mgr.get_valid_ne_candidates(twitter_site) if entities_to_evaluate is None: print "No ambiguous entities + candidates in cache. Run run_all_dataset_generators "+\ "script and choose to first fetch and store more entities from short texts." return judged_row_plus_headers = csv_util.query_csv_for_rows(__entities_results_csv_path__, False) judged_row_num = 0 already_judged = [] # list of (entity id, candidate link) for judge_row in judged_row_plus_headers: try: if judged_row_num==0: # row 0 is header entity_id_col = judge_row.index('Input.entity_id') candidate_link_col = judge_row.index('Input.candidate_link') else: judged_tuple = (judge_row[entity_id_col], judge_row[candidate_link_col]) if not judged_tuple in already_judged: already_judged.append(judged_tuple) judged_row_num = judged_row_num+1 except: continue # just ignore a problematic row # Determine what entity+candidate tasks we actually want to write to a spreadsheet # and send to mturk since we don't have resources for unlimited mturk tasks tasks = {} # NamedEntity object -> candidate judgment tasks we actually want performed user_entities = defaultdict(list) # username -> [NamedEntity obj] done_shorttexts = [] # list of shorttext id random.shuffle(entities_to_evaluate) # so we get a random subset of a user's entities for ne_obj in entities_to_evaluate: # "40 nouns usually enough to establish statistically significant # differences between WSD algorithms" (Santamaria et al., 2010) username = ne_obj.username if len(user_entities[username]) > 50: continue # have enough entities for this user # limiting our dataset to one named entity per short text shorttext_id = ne_obj.shorttext_id if shorttext_id in done_shorttexts: continue # no need to create tasks for candidates we already have annotator judgments for entity_id = ne_obj.get_entity_id() candidate_URLs = ne_obj.get_candidate_wikiURLs() valid_candidate_tasks = [] for candidate_URL in candidate_URLs: if ((entity_id, candidate_URL) in already_judged): continue valid_candidate_tasks.append(candidate_URL) if len(valid_candidate_tasks)==0: continue # already have annotator judgments for all of this entity's candidates if len(candidate_URLs)+len(valid_candidate_tasks) < 2: # this would be a non-ambiguous entity, and we should never reach this # point because such entities should have been filtered out by now raise tasks[entity_id] = valid_candidate_tasks user_entities[username].append(ne_obj) done_shorttexts.append(shorttext_id) # put valid entities + candidates in the spreadsheet until reach our limit of tasks task_max = 1400 rows = [] headers = ['entity_id', 'short_text', 'ambiguous_entity', 'candidate_link'] rows.append(headers) for username in user_entities: # add users until reach our limit on the number of tasks we can afford, # but break at this point in the loop rather than in the inner loop to # ensure that we do have at least 50 entities per user (even if this # means we go over our task limit a little in order to reach that amount) if len(rows) > task_max: break # bypass users who haven't written the minimum number of valid entities # required to establish statistical significance between the algorithms if len(user_entities[username]) < 50: continue # should be 50 NamedEntity objects per user, and we'll make tasks for their candidates for ne_obj in user_entities[username]: entity_id = ne_obj.get_entity_id() # make sure the entity presented to a Turker looks the same as # it appears in the short text (ie with the same capitalization) original_shorttext = ne_obj.shorttext_str.decode('latin-1') surface_form = ne_obj.surface_form if not surface_form in original_shorttext: surface_form = __match_appearance__(surface_form, original_shorttext) # shuffle candidates so that they don't appear # in wikiminer's/dbpedia's ranking order and bias the turker candidate_URLs = tasks[entity_id] random.shuffle(candidate_URLs) choices = candidate_URLs[:] # copy (list slicing) for choice in choices: # make a separate row for each candidate link # rather than putting all links in a single cell row = [entity_id, original_shorttext, surface_form, choice] rows.append(row) if len(rows)%50==0: # write the rows every once in a while in case we reach an error print "Updating spreadsheet..."+str(len(rows)) csv_util.write_to_spreadsheet(__entities_to_judge_csv_path__, rows) # dump to csv csv_util.write_to_spreadsheet(__entities_to_judge_csv_path__, rows)
def run_all_algorithms(RESLVE_alg, site, use_cache): ''' @param RESLVE_alg: A constructed reslve_algorithm object @param use_cache: False if still working on algorithms and boosting performance and therefore don't want to cache their rankings in a file yet; True if ready to cache algorithms' rankings ''' # Valid entities and their labels annotated by Mechanical Turk workers entities_to_evaluate = entity_dataset_mgr.get_valid_ne_candidates(site) entity_judgments = entity_dataset_mgr.get_entity_judgements(site) if (entities_to_evaluate is None or len(entities_to_evaluate)==0 or entity_judgments is None or len(entity_judgments))==0: print "No labeled ambiguous entities + candidates available. Run appropriate scripts first." return {} # entities that have been labeled by human judges entities_to_resolve = [ne_obj for ne_obj in entities_to_evaluate if ne_obj.get_entity_id() in entity_judgments] print str(len(entities_to_evaluate))+" and "+str(len(entity_judgments))+\ " judgments available, resulting in "+str(len(entities_to_resolve))+" entities to resolve" # Usernames that do not belong to the same individual on the site and # Wikipedia and that we'll use as a baseline for no background knowledge nonmatch_usernames = crosssite_username_dataset_mgr.get_confirmed_nonmatch_usernames(site) resolved_entities = [] for ne_obj in entities_to_resolve: print str(len(resolved_entities))+" out of "+\ str(len(entities_to_resolve))+" resolved.." entity_id = ne_obj.get_entity_id() evaluated_candidates = entity_judgments[entity_id] # construct a ResolvedEntity object to represent this # ambiguous entity and its various candidate rankings resolved_entity = ResolvedEntity(ne_obj, evaluated_candidates) resolved_entities.append(resolved_entity) reslve_algorithms = [RESLVE_alg] for reslve_alg in reslve_algorithms: print "Ranking candidates using RESLVE's "+str(reslve_alg.alg_type)+" algorithm..." candidate_titles = ne_obj.get_candidate_titles() # perform the RESLVE ranking.. reslve_ranking_user_match = reslve_alg.rank_candidates(candidate_titles, ne_obj.username) # perform the same algorithm's ranking again but this time use # a non-match user's interest model as background information, # which according to our hypothesis should provide less relevant # semantic background knowledge and thus have lower performance random.shuffle(nonmatch_usernames) random_nonmatch_username = nonmatch_usernames[0] reslve_ranking_user_nonmatch = reslve_alg.rank_candidates(candidate_titles, random_nonmatch_username) resolved_entity.add_reslve_ranking(reslve_alg.alg_id, reslve_ranking_user_match, reslve_ranking_user_nonmatch) # cache intermittently in case we need to exit.. __save_resolved_entities__(resolved_entities, site, use_cache) __save_resolved_entities__(resolved_entities, site, use_cache) # Cache resolved entities return resolved_entities