#!/usr/bin/python3 import threading import time import math import sys import pdb import requests import urllib.parse from common import * import config_utils as cf import json from collections import OrderedDict import argparse import numpy as np MASK = ":__entity__" RESULT_MASK = "NER_FINAL_RESULTS:" DEFAULT_CONFIG = "./ensemble_config.json" DEFAULT_TEST_BATCH_FILE="bootstrap_test_set.txt" NER_OUTPUT_FILE="ner_output.txt" DEFAULT_THRESHOLD = 1 #1 standard deviation from nean - for cross over prediction actions_arr = [] class AggregateNER: def __init__(self,config_file): global actions_arr base_path = cf.read_config(config_file)["BASE_PATH"] if ("BASE_PATH" in cf.read_config(config_file)) else "./" self.error_fp = open(base_path + "failed_queries_log.txt","a") self.rfp = open(base_path + "query_response_log.txt","a") self.query_log_fp = open(base_path + "query_logs.txt","a") self.inferred_entities_log_fp = open(base_path + "inferred_entities_log.txt","a") self.threshold = DEFAULT_THRESHOLD #TBD read this from confg. cf.read_config()["CROSS_OVER_THRESHOLD_SIGMA"] self.servers = cf.read_config(config_file)["NER_SERVERS"] actions_arr = [ {"url":cf.read_config(config_file)["actions_arr"][0]["url"],"desc":cf.read_config(config_file)["actions_arr"][0]["desc"], "precedence":cf.read_config(config_file)["bio_precedence_arr"],"common":cf.read_config(config_file)["common_entities_arr"]}, {"url":cf.read_config(config_file)["actions_arr"][1]["url"],"desc":cf.read_config(config_file)["actions_arr"][1]["desc"],"precedence":cf.read_config(config_file)["phi_precedence_arr"],"common":cf.read_config(config_file)["common_entities_arr"]}, ] def add_term_punct(self,sent): if (len(sent) > 1): end_tokens = "!,.:;?" last_char = sent[-1] if (last_char not in end_tokens): #End all sentences with a period if not already present in sentence. sent = sent + ' . ' print("End punctuated sent:",sent) return sent def fetch_all(self,inp,model_results_arr): self.query_log_fp.write(inp+"\n") self.query_log_fp.flush() inp = self.add_term_punct(inp) results = model_results_arr #print(json.dumps(results,indent=4)) #this updates results with ensembled results results = self.ensemble_processing(inp,results) return_stat = "Failed" if len(results["ensembled_ner"]) == 0 else "Success" results["stats"] = { "Ensemble server count" : str(len(model_results_arr)), "return_status": return_stat} self.rfp.write( "\n" + json.dumps(results,indent=4)) self.rfp.flush() return results def get_conflict_resolved_entity(self,results,term_index,terms_count,servers_arr): pos_index = str(term_index + 1) s1_entity = extract_main_entity(results,0,pos_index) s2_entity = extract_main_entity(results,1,pos_index) span_count1 = get_span_info(results,0,term_index,terms_count) span_count2 = get_span_info(results,1,term_index,terms_count) if(span_count1 != span_count2): print("Both input spans dont match. This is the effect of normalized casing that is model specific. Picking min span length") span_count1 = span_count1 if span_count1 <= span_count2 else span_count2 if (s1_entity == s2_entity): server_index = 0 if (s1_entity in servers_arr[0]["precedence"]) else 1 if (s1_entity != "O"): print("Both servers agree on prediction for term:",results[0]["ner"][pos_index]["term"],":",s1_entity) return server_index,span_count1,-1 else: print("Servers do not agree on prediction for term:",results[0]["ner"][pos_index]["term"],":",s1_entity,s2_entity) if (s2_entity == "O"): print("Server 2 returned O. Picking server 1") return 0,span_count1,-1 if (s1_entity == "O"): print("Server 1 returned O. Picking server 2") return 1,span_count2,-1 #Both the servers dont agree on their predictions. First server is BIO server. Second is PHI #Examine both server predictions. #Case 1: If just one of them makes a single prediction, then just pick that - it indicates one model is confident while the other isnt. #Else. # If the top prediction of one of them is a cross prediction, then again drop that prediction and pick the server being cross predicted. # Else. Return both predictions, but with the higher confidence prediction first #Case 2: Both dont cross predict. Then just return both predictions with higher confidence prediction listed first #Cross prediction is checked only for predictions a server makes ABOVE prediction mean. picked_server_index,cross_prediction_count = self.pick_single_server_if_possible(results,term_index,servers_arr) return picked_server_index,span_count1,cross_prediction_count def pick_single_server_if_possible(self,results,term_index,servers_arr): ''' Return param : index of picked server ''' pos_index = str(term_index + 1) predictions_dict = {} orig_cs_predictions_dict = {} single_prediction_count = 0 single_prediction_server_index = -1 for server_index in range(len(results)): if (pos_index in results[server_index]["entity_distribution"]): predictions = self.get_predictions_above_threshold(results[server_index]["entity_distribution"][pos_index]) predictions_dict[server_index] = predictions #This is used below to only return top server prediction orig_cs_predictions = self.get_predictions_above_threshold(results[server_index]["orig_cs_prediction_details"][pos_index]) orig_cs_predictions_dict[server_index] = orig_cs_predictions #this is used below for cross prediction determination since it is just a CS prediction #single_prediction_count += 1 if (len(orig_cs_predictions) == 1) else 0 #if (len(orig_cs_predictions) == 1): # single_prediction_server_index = server_index if (single_prediction_count == 1): is_included = is_included_in_server_entities(orig_cs_predictions_dict[single_prediction_server_index],servers_arr[single_prediction_server_index],False) if(is_included == False) : print("This is an odd case of single server prediction, that is a cross over") ret_index = 0 if single_prediction_server_index == 1 else 1 return ret_index,-1 else: print("Returning the index of single prediction server") return single_prediction_server_index,-1 elif (single_prediction_count == 2): print("Both have single predictions") cross_predictions = {} cross_prediction_count = 0 for server_index in range(len(results)): if (pos_index in results[server_index]["entity_distribution"]): is_included = is_included_in_server_entities(orig_cs_predictions_dict[server_index],servers_arr[server_index],False) cross_predictions[server_index] = not is_included cross_prediction_count += 1 if not is_included else 0 if (cross_prediction_count == 2): #this is an odd case of both cross predicting with high confidence. Not sure if we will ever come here. print("*********** BOTH servers are cross predicting! ******") return self.pick_top_server_prediction(predictions_dict),2 elif (cross_prediction_count == 0): #Neither are cross predecting print("*********** BOTH servers have single predictions within their domain - returning both ******") return self.pick_top_server_prediction(predictions_dict),2 else: print("Returning just the server that is not cross predicting, dumping the cross prediction") ret_index = 1 if cross_predictions[0] == True else 0 #Given a server cross predicts, return the other server index return ret_index,-1 else: print("*** Both servers have multiple predictions above mean") #both have multiple predictions above mean cross_predictions = {} strict_cross_predictions = {} cross_prediction_count = 0 strict_cross_prediction_count = 0 for server_index in range(len(results)): if (pos_index in results[server_index]["entity_distribution"]): is_included = is_included_in_server_entities(orig_cs_predictions_dict[server_index],servers_arr[server_index],False) strict_is_included = strict_is_included_in_server_entities(orig_cs_predictions_dict[server_index],servers_arr[server_index],False) cross_predictions[server_index] = not is_included strict_cross_predictions[server_index] = not strict_is_included cross_prediction_count += 1 if not is_included else 0 strict_cross_prediction_count += 1 if not strict_is_included else 0 if (cross_prediction_count == 2): print("*********** BOTH servers are ALSO cross predicting and have multiple predictions above mean ******") return self.pick_top_server_prediction(predictions_dict),2 elif (cross_prediction_count == 0): print("*********** BOTH servers are ALSO predicting within their domain ******") #if just one of them is predicting in the common set, then just pick the server that is predicting in its primary set. #if (strict_cross_prediction_count == 1): # ret_index = 1 if (0 not in strict_cross_predictions or strict_cross_predictions[0] == True) else 0 #Given a server cross predicts, return the other server index # return ret_index,-1 #else: # return self.pick_top_server_prediction(predictions_dict),2 return self.pick_top_server_prediction(predictions_dict),2 else: print("Returning just the server that is not cross predicting, dumping the cross prediction. This is mainly to reduce the noise in prefix predictions that show up in CS context predictions") ret_index = 1 if (0 not in cross_predictions or cross_predictions[0] == True) else 0 #Given a server cross predicts, return the other server index return ret_index,-1 #print("*********** One of them is also cross predicting ******") #return self.pick_top_server_prediction(predictions_dict),2 def pick_top_server_prediction(self,predictions_dict): ''' ''' if (len(predictions_dict) != 2): return 0 assert(len(predictions_dict) == 2) return 0 if (predictions_dict[0][0]["conf"] >= predictions_dict[1][0]["conf"]) else 1 def get_predictions_above_threshold(self,predictions): dist = predictions["cs_distribution"] sum_predictions = 0 ret_arr = [] if(len(dist) != 0): mean_score = 1.0/len(dist) #input is a prob distriubution. so sum is 1 else: mean_score = 0 #sum_deviation = 0 #for node in dist: # sum_deviation += (mean_score - node["confidence"])*(mean_score - node["confidence"]) #variance = sum_deviation/len(dist) #std_dev = math.sqrt(variance) #threshold = mean_score + std_dev*self.threshold #default is 1 standard deviation from mean threshold = mean_score pick_count = 1 for node in dist: if (node["confidence"] >= threshold): ret_arr.append({"e":node["e"],"conf":node["confidence"]}) pick_count += 1 else: break #this is a reverse sorted list. So no need to check anymore if (len(dist) > 0): assert(len(ret_arr) > 0) return ret_arr def check_if_entity_in_arr(self,entity,arr): for node in arr: if (entity == node["e"].split('[')[0]): return True return False def gen_resolved_entity(self,results,server_index,pivot_index,run_index,cross_prediction_count,servers_arr): if (cross_prediction_count == 1 or cross_prediction_count == -1): #This is the case where we are emitting just one server prediction. In this case, if CS and consolidated dont match, emit both if (pivot_index in results[server_index]["orig_cs_prediction_details"]): if (len(results[server_index]["orig_cs_prediction_details"][pivot_index]['cs_distribution']) == 0): #just use the ci prediction in this case. This happens only for boundary cases of a single entity in a sentence and there is no context orig_cs_entity = results[server_index]["orig_ci_prediction_details"][pivot_index]['cs_distribution'][0] else: orig_cs_entity = results[server_index]["orig_cs_prediction_details"][pivot_index]['cs_distribution'][0] orig_ci_entity = results[server_index]["orig_ci_prediction_details"][pivot_index]['cs_distribution'][0] m1 = orig_cs_entity["e"].split('[')[0] m1_ci = orig_ci_entity["e"].split('[')[0] is_ci_included = True if (m1_ci in servers_arr[server_index]["precedence"]) else False consolidated_entity = results[server_index]["ner"][pivot_index] m2,dummy = prefix_strip(consolidated_entity["e"].split('[')[0]) if (m1 != m2): #if we come here consolidated is not same as cs prediction. So we emit both consolidated and cs ret_obj = results[server_index]["ner"][run_index].copy() dummy,prefix = prefix_strip(ret_obj["e"]) n1 = flip_category(orig_cs_entity) n1["e"] = prefix + n1["e"] n2 = flip_category(consolidated_entity) ret_obj["e"] = n2["e"] + "/" + n1["e"] return ret_obj else: #if we come here consolidated is same as cs prediction. So we try to either use ci or the second cs prediction if ci is out of domain if (m1 != m1_ci): #CS and CI are not same if (is_ci_included): #Emity both CS and CI ret_obj = results[server_index]["ner"][run_index].copy() dummy,prefix = prefix_strip(ret_obj["e"]) n1 = flip_category(orig_cs_entity) n1["e"] = prefix + n1["e"] n2 = flip_category(orig_ci_entity) n2["e"] = prefix + n2["e"] ret_obj["e"] = n1["e"] + "/" + n2["e"] return ret_obj else: #We come here for the case where CI is not in server list. So we pick the second cs as an option if meaningful if (len(results[server_index]["orig_cs_prediction_details"][pivot_index]['cs_distribution']) >= 2): ret_arr = self.get_predictions_above_threshold(results[server_index]["orig_cs_prediction_details"][pivot_index]) orig_cs_second_entity = results[server_index]["orig_cs_prediction_details"][pivot_index]['cs_distribution'][1] m2_cs = orig_cs_second_entity["e"].split('[')[0] is_cs_included = True if (m2_cs in servers_arr[server_index]["precedence"]) else False is_cs_included = True #Disabling cs included check. If prediction above threshold is cross prediction, then letting it through assert (m2_cs != m1) if (is_cs_included and self.check_if_entity_in_arr(m2_cs,ret_arr)): ret_obj = results[server_index]["ner"][run_index].copy() dummy,prefix = prefix_strip(ret_obj["e"]) n1 = flip_category(orig_cs_second_entity) n1["e"] = prefix + n1["e"] n2 = flip_category(orig_cs_entity) n2["e"] = prefix + n2["e"] ret_obj["e"] = n2["e"] + "/" + n1["e"] return ret_obj else: return flip_category(results[server_index]["ner"][run_index]) else: return flip_category(results[server_index]["ner"][run_index]) else: #here cs and ci are same. So use two cs predictions if meaningful if (len(results[server_index]["orig_cs_prediction_details"][pivot_index]['cs_distribution']) >= 2): ret_arr = self.get_predictions_above_threshold(results[server_index]["orig_cs_prediction_details"][pivot_index]) orig_cs_second_entity = results[server_index]["orig_cs_prediction_details"][pivot_index]['cs_distribution'][1] m2_cs = orig_cs_second_entity["e"].split('[')[0] is_cs_included = True if (m2_cs in servers_arr[server_index]["precedence"]) else False is_cs_included = True #Disabling cs included check. If prediction above threshold is cross prediction, then letting it through assert (m2_cs != m1) if (is_cs_included and self.check_if_entity_in_arr(m2_cs,ret_arr)): ret_obj = results[server_index]["ner"][run_index].copy() dummy,prefix = prefix_strip(ret_obj["e"]) n1 = flip_category(orig_cs_second_entity) n1["e"] = prefix + n1["e"] n2 = flip_category(orig_cs_entity) n2["e"] = prefix + n2["e"] ret_obj["e"] = n2["e"] + "/" + n1["e"] return ret_obj else: return flip_category(results[server_index]["ner"][run_index]) else: return flip_category(results[server_index]["ner"][run_index]) else: return flip_category(results[server_index]["ner"][run_index]) else: #Case where both servers dont match ret_obj = results[server_index]["ner"][run_index].copy() #ret_obj["e"] = results[0]["ner"][run_index]["e"] + "/" + results[1]["ner"][run_index]["e"] index2 = 1 if server_index == 0 else 0 #this is the index of the dominant server with hihgher prediction confidence n1 = flip_category(results[server_index]["ner"][run_index]) n2 = flip_category(results[index2]["ner"][run_index]) ret_obj["e"] = n1["e"] + "/" + n2["e"] return ret_obj def confirm_same_size_responses(self,sent,results): count = 0 for i in range(len(results)): if ("ner" in results[i]): ner = results[i]["ner"] else: print("Server",i," returned invalid response;",results[i]) self.error_fp.write("Server " + str(i) + " failed for query: " + sent + "\n") self.error_fp.flush() return 0 if(count == 0): assert(len(ner) > 0) count = len(ner) else: if (count != len(ner)): print("Warning. The return sizes of both servers do not match. This must be truncated sentence, where tokenization causes different length truncations. Using min length") count = count if count < len(ner) else len(ner) return count def get_ensembled_entities(self,sent,results,servers_arr): ensembled_ner = OrderedDict() orig_cs_predictions = OrderedDict() orig_ci_predictions = OrderedDict() ensembled_conf = OrderedDict() ambig_ensembled_conf = OrderedDict() ensembled_ci = OrderedDict() ensembled_cs = OrderedDict() ambig_ensembled_ci = OrderedDict() ambig_ensembled_cs = OrderedDict() print("Ensemble candidates") terms_count = self.confirm_same_size_responses(sent,results) if (terms_count == 0): return ensembled_ner,ensembled_conf,ensembled_ci,ensembled_cs,ambig_ensembled_conf,ambig_ensembled_ci,ambig_ensembled_cs,orig_cs_predictions,orig_ci_predictions assert(len(servers_arr) == len(results)) term_index = 0 while (term_index < terms_count): pos_index = str(term_index + 1) assert(len(servers_arr) == 2) #TBD. Currently assumes two servers in prototype to see if this approach works. To be extended to multiple servers server_index,span_count,cross_prediction_count = self.get_conflict_resolved_entity(results,term_index,terms_count,servers_arr) pivot_index = str(term_index + 1) for span_index in range(span_count): run_index = str(term_index + 1 + span_index) ensembled_ner[run_index] = self.gen_resolved_entity(results,server_index,pivot_index,run_index,cross_prediction_count,servers_arr) if (run_index in results[server_index]["entity_distribution"]): ensembled_conf[run_index] = results[server_index]["entity_distribution"][run_index] ensembled_conf[run_index]["e"] = strip_prefixes(ensembled_ner[run_index]["e"]) #this is to make sure the same tag can be taken from NER result or this structure. #When both server responses are required, just return the details of first server for now ensembled_ci[run_index] = results[server_index]["ci_prediction_details"][run_index] ensembled_cs[run_index] = results[server_index]["cs_prediction_details"][run_index] orig_cs_predictions[run_index] = results[server_index]["orig_cs_prediction_details"][run_index] orig_ci_predictions[run_index] = results[server_index]["orig_ci_prediction_details"][run_index] if (cross_prediction_count == 0 or cross_prediction_count == 2): #This is an ambiguous prediction. Send both server responses second_server = 1 if server_index == 0 else 1 if (run_index in results[second_server]["entity_distribution"]): #It may not be present if the B/I tags are out of sync from servers. ambig_ensembled_conf[run_index] = results[second_server]["entity_distribution"][run_index] ambig_ensembled_conf[run_index]["e"] = ensembled_ner[run_index]["e"] #this is to make sure the same tag can be taken from NER result or this structure. ambig_ensembled_ci[run_index] = results[second_server]["ci_prediction_details"][run_index] if (ensembled_ner[run_index]["e"] != "O"): self.inferred_entities_log_fp.write(results[0]["ner"][run_index]["term"] + " " + ensembled_ner[run_index]["e"] + "\n") term_index += span_count self.inferred_entities_log_fp.flush() return ensembled_ner,ensembled_conf,ensembled_ci,ensembled_cs,ambig_ensembled_conf,ambig_ensembled_ci,ambig_ensembled_cs,orig_cs_predictions,orig_ci_predictions def ensemble_processing(self,sent,results): global actions_arr ensembled_ner,ensembled_conf,ci_details,cs_details,ambig_ensembled_conf,ambig_ci_details,ambig_cs_details,orig_cs_predictions,orig_ci_predictions = self.get_ensembled_entities(sent,results,actions_arr) final_ner = OrderedDict() final_ner["ensembled_ner"] = ensembled_ner final_ner["ensembled_prediction_details"] = ensembled_conf final_ner["ci_prediction_details"] = ci_details final_ner["cs_prediction_details"] = cs_details final_ner["ambig_prediction_details_conf"] = ambig_ensembled_conf final_ner["ambig_prediction_details_ci"] = ambig_ci_details final_ner["ambig_prediction_details_cs"] = ambig_cs_details final_ner["orig_cs_prediction_details"] = orig_cs_predictions final_ner["orig_ci_prediction_details"] = orig_ci_predictions #final_ner["individual"] = results return final_ner class myThread (threading.Thread): def __init__(self, url,param,desc): threading.Thread.__init__(self) self.url = url self.param = param self.desc = desc self.results = {} def run(self): print ("Starting " + self.url + self.param) escaped_url = self.url + self.param.replace("#","-") #TBD. This is a nasty hack for client side handling of #. To be fixed. For some reason, even replacing with parse.quote or just with %23 does not help. The fragment after # is not sent to server. Works just fine in wget with %23 print("ESCAPED:",escaped_url) out = requests.get(escaped_url) try: self.results = json.loads(out.text,object_pairs_hook=OrderedDict) except: print("Empty response from server for input:",self.param) self.results = json.loads("{}",object_pairs_hook=OrderedDict) self.results["server"] = self.desc print ("Exiting " + self.url + self.param) # Create new threads def create_workers(inp_dict,inp): threads_arr = [] for i in range(len(inp_dict)): threads_arr.append(myThread(inp_dict[i]["url"],inp,inp_dict[i]["desc"])) return threads_arr def start_workers(threads_arr): for thread in threads_arr: thread.start() def wait_for_completion(threads_arr): for thread in threads_arr: thread.join() def get_results(threads_arr): results = [] for thread in threads_arr: results.append(thread.results) return results def prefix_strip(term): prefix = "" if (term.startswith("B_") or term.startswith("I_")): prefix = term[:2] term = term[2:] return term,prefix def strip_prefixes(term): split_entities = term.split('/') if (len(split_entities) == 2): term1,dummy = prefix_strip(split_entities[0]) term2,dummy = prefix_strip(split_entities[1]) return term1 + '/' + term2 else: assert(len(split_entities) == 1) term1,dummy = prefix_strip(split_entities[0]) return term1 #This hack is simply done for downstream API used for UI displays the entity instead of the class. Details has all additional info def flip_category(obj): new_obj = obj.copy() entity_type_arr = obj["e"].split("[") if (len(entity_type_arr) > 1): term = entity_type_arr[0] if (term.startswith("B_") or term.startswith("I_")): prefix = term[:2] new_obj["e"] = prefix + entity_type_arr[1].rstrip("]") + "[" + entity_type_arr[0][2:] + "]" else: new_obj["e"] = entity_type_arr[1].rstrip("]") + "[" + entity_type_arr[0] + "]" return new_obj def extract_main_entity(results,server_index,pos_index): main_entity = results[server_index]["ner"][pos_index]["e"].split('[')[0] main_entity,dummy = prefix_strip(main_entity) return main_entity def get_span_info(results,server_index,term_index,terms_count): pos_index = str(term_index + 1) entity = results[server_index]["ner"][pos_index]["e"] span_count = 1 if (entity.startswith("I_")): print("Skipping an I tag for server:",server_index,". This has to be done because of mismatched span because of model specific casing normalization that changes POS tagging. This happens only for sentencees user does not explicirly tag with ':__entity__'") return span_count assert(not entity.startswith("I_")) if (entity.startswith("B_")): term_index += 1 while(term_index < terms_count): pos_index = str(term_index + 1) entity = results[server_index]["ner"][pos_index]["e"] if (entity == "O"): break span_count += 1 term_index += 1 return span_count def is_included_in_server_entities(predictions,s_arr,check_first_only): for entity in predictions: entity = entity['e'].split('[')[0] if ((entity not in s_arr["precedence"]) and (entity not in s_arr["common"])): #do not treat the presence of an entity in common as a cross over return False if (check_first_only): return True #Just check the top prediction for inclusion in the new semantics return True def strict_is_included_in_server_entities(predictions,s_arr,check_first_only): for entity in predictions: entity = entity['e'].split('[')[0] if ((entity not in s_arr["precedence"])): #do not treat the presence of an entity in common as a cross over return False if (check_first_only): return True #Just check the top prediction for inclusion in the new semantics return True if __name__ == '__main__': parser = argparse.ArgumentParser(description='main NER for a single model ',formatter_class=argparse.ArgumentDefaultsHelpFormatter) parser.add_argument('-input', action="store", dest="input",default=DEFAULT_TEST_BATCH_FILE,help='Input file for batch run option') parser.add_argument('-config', action="store", dest="config", default=DEFAULT_CONFIG,help='config file path') parser.add_argument('-output', action="store", dest="output",default=NER_OUTPUT_FILE,help='Output file for batch run option') parser.add_argument('-option', action="store", dest="option",default="canned",help='Valid options are canned,batch,interactive. canned - test few canned sentences used in medium artice. batch - tag sentences in input file. Entities to be tagged are determing used POS tagging to find noun phrases.interactive - input one sentence at a time') results = parser.parse_args() config_file = results.config