From 5b86b63e0cfe7d7a6bdaafbe1f962bfc5d0dc009 Mon Sep 17 00:00:00 2001
From: chernenko <chernenko@cl.uni-heidelberg.de>
Date: Wed, 28 Mar 2018 14:52:44 +0200
Subject: [PATCH] Eine Neue Datei hochladen
---
experiments/variant_34.py | 305 ++++++++++++++++++++++++++++++++++++++
1 file changed, 305 insertions(+)
create mode 100644 experiments/variant_34.py
diff --git a/experiments/variant_34.py b/experiments/variant_34.py
new file mode 100644
index 0000000..691fa1a
--- /dev/null
+++ b/experiments/variant_34.py
@@ -0,0 +1,305 @@
+#!/bin/env python3.5
+#variant_34.py
+#
+#usage:
+#$ ./variant_34.py
+#
+#author: <Tatjana Chernenko, Utaemon Toyota>
+
+
+"""
+<VARIANT 34>
+
+------------------- DESCRIPTION -------------------
+
+The pipeline system performs the 34th variant of the system for WSI (word sense induction) task (the Task 11 at SemEval 2013).
+
+The system creates semantic related clusters from the given snippets (the text fragments we get back from the search engine) for each pre-defined ambigue topic.
+
+------------------- METHODS -------------------
+
+For the WSI purposes it uses the following methods:
+
+- For pre-rpocessing: tokenization + remove punctuation
+- Language model: sent2vec, plaintexts_bigramm - trained Model (Wikipedia 2017)
+- Compositional semantics: vector mixture model (BOW (bag-of-words) representation with summarization for each snippet)
+- Clustering: Cos similarity with max Similarity
+
+------------------- EVALUATION -------------------
+
+=========== Final average value of F1: =====================
+average F1 = 0.6013936257538159
+
+=========== Final average value of Rand Index: =============
+average Rand Index = 0.38823232323232326
+
+=========== Final average value of Adjusted Rand Index: ====
+average Adj Rand Index = -0.0038108019555718464
+
+=========== Final average value of Jaccard Index: ==========
+average Jaccard Index = 0.21276973487337159
+
+================ Statistics: ====================================
+============ average number of created clusters: 6.25
+============ average cluster size: 12.740079365079364
+
+"""
+import sys
+sys.path.append("/home/students/toyota")
+import sent2vec
+from collections import defaultdict, deque
+import re
+import math
+import numpy as np
+from sklearn.cluster import MeanShift
+import nltk
+from scipy import spatial
+
+# Read Data; get the number of topics and number of subtopics for each topic:
+def read_data(topics_file, subtopics_file, results_file):
+ with open(topics_file, "r") as f:
+ topics_data = f.readlines()
+ with open(subtopics_file, "r") as f:
+ subtopics_data = f.readlines()
+ with open(results_file, "r") as f:
+ results_data = f.readlines()
+ number_topics = 0
+ for line in topics_data:
+ if not line.startswith("ID"):
+ number_topics += 1
+ subtopics = {}
+ for line in subtopics_data:
+ if not line.startswith("ID"):
+ topic = line.split(" ")[0].split(".")[0]
+ if topic not in subtopics.keys():
+ subtopics[topic]=1
+ else:
+ subtopics[topic]+=1
+ return topics_data, subtopics_data, results_data, number_topics, subtopics
+
+# Create a vocabulary with topics as keys and lists with lists of snippets with IDs as values:
+def devide_text_into_topics(results_data, number_topics):
+ text = defaultdict(list)
+ for line in results_data:
+ if not line.startswith("ID"):
+ if line.split()[0].split(".")[0] not in text.keys():
+ text[line.split()[0].split(".")[0]] = []
+ help_structure = []
+ help_structure.append(line.split(" ")[0])
+ help_structure.append(line.split(" ")[2:])
+ text[line.split()[0].split(".")[0]].append(help_structure)
+ else:
+ help_structure = []
+ help_structure.append(line.split(" ")[0])
+ help_structure.append(line.split(" ")[2:])
+ text[line.split()[0].split(".")[0]].append(help_structure)
+ # Clean sentences from "\n":
+ for values in text.values():
+ for paragr in values:
+ for sent in paragr:
+ if sent == "\n":
+ paragr.remove(sent)
+ #print(text["45"]) # example of the output for the topic "45"
+ return text
+
+# Preprocess Data (tokenize every sentence in every topic):
+def preprocess_data(text):
+ # Tokenize:
+ for value in text.values():
+ for paragr in value:
+ for i in range(1,len(paragr)):
+ tokens = re.findall(r"\w+", str(paragr[i])) # remove punctuation
+ words = []
+ for word in tokens:
+ if word == "n":
+ words.append(" ")
+ else:
+ words.append(word.strip()) # delete first empty placeholder
+ paragr[i] = words
+
+ prepr_data = text
+ #print(prepr_data["45"]) # example of the output for the topic "45"
+ return prepr_data
+
+# For every word in a sentence make a vector representation with sense2vec; make a compositional vector for every sentence as sum of BOW vectors:
+def compos_sent2vec(prepr_data,len_vector):
+ model = sent2vec.Sent2vecModel()
+ model.load_model("/proj/toyota/plaintexts_bigramm.bin")
+ for value in prepr_data.values():
+ for paragr in value: #one snippet
+ par_list = [] # list with a snippet
+ vector_paragr = np.zeros(len_vector) #vector for one snippet for a sum
+ for sent in paragr[1:]: #sent in a snippet
+ #print("sent: ", sent)
+ vector_sent = [] #list for all sentences in a snippet
+ for word in sent: #word
+ try:
+ query_vector = model.embed_sentence(word)
+ vector_sent.append(query_vector) #add a word-vector to a list for sentences in a snippet - BOW for all words in a snippet - now for a sentence - for every sentence
+ except:
+ continue
+ summe = np.zeros(len_vector) # vector for a summ
+ for vector in vector_sent: # for one word in all sentences
+ summe+=vector # summ all words in a snippet - vector for a snippet
+ par_list.append(summe)#?#add a summ(vector for a snippet) to a list with a snippet
+ for sentence in par_list: # for all snippet-vectors
+ vector_paragr+=sentence # sum all snippets
+ paragr.append(vector_paragr) #add to a snippet a summ of all snippets
+ compos_vectors = prepr_data
+ #print(compos_vectors["45"]) # example of the output for the topic "45"
+ return compos_vectors
+
+# Create a vocabulary for subtopics with topics as keys and lists with subtopics with IDs as values:
+def devide_subtopics_into_topics(subtopics_data, number_topics):
+ subtopics_vectors = defaultdict(list)
+ for line in subtopics_data:
+ if not line.startswith("ID"):
+ if line.split()[0].split(".")[0] not in subtopics_vectors.keys():
+ subtopics_vectors[line.split()[0].split(".")[0]] = []
+ help_structure = []
+ help_structure.append(line.split(" ")[0])
+ help_structure.append(line.split(" ")[1])
+ subtopics_vectors[line.split()[0].split(".")[0]].append(help_structure)
+ else:
+ help_structure = []
+ help_structure.append(line.split(" ")[0])
+ help_structure.append(line.split(" ")[1])
+ subtopics_vectors[line.split()[0].split(".")[0]].append(help_structure)
+ # Clean sentences from "\n":
+ for values in subtopics_vectors.values():
+ for paragr in values:
+ for sent in paragr:
+ if sent == "\n":
+ paragr.remove(sent)
+ #print("Subtopics Vectors: ", subtopics_vectors)
+ return subtopics_vectors
+
+# Make vector representation of Subtopics with sense2vec, sum of BOW with sense2vec:
+def compose_sent2vec_subtopics(subtopics_vectors, len_vector):
+ model = sent2vec.Sent2vecModel()
+ model.load_model("/proj/toyota/plaintexts_bigramm.bin")
+ for value in subtopics_vectors.values():
+ for paragr in value: #one snippet
+ par_list = [] # list with a snippet
+ vector_paragr = np.zeros(len_vector) #vector for one snippet for a sum
+ for sent in paragr[1:]: #sent in a snippet
+ vector_sent = [] #list for all sentences in a snippet
+ for word in sent: #word
+ try:
+ query_vector = model.embed_sentence(word)
+ vector_sent.append(query_vector) #add a word-vector to a list for sentences in a snippet - BOW for all words in a snippet - now for a sentence - for every sentence
+ except:
+ continue
+ summe = np.zeros(len_vector) # vector for a summ
+ for vector in vector_sent: # for one word in all sentences
+ summe+=vector # summ all words in a snippet - vector for a snippet
+ par_list.append(summe)#?#add a summ(vector for a snippet) to a list with a snippet
+ for sentence in par_list: # for all snippet-vectors
+ vector_paragr+=sentence # sum all snippets
+ paragr.append(vector_paragr) #add to a snippet a summ of all snippets
+ compos_subtopics_vectors = subtopics_vectors
+ #print("Composed Subtopics Vectors: ", compos_subtopics_vectors)
+ return compos_subtopics_vectors
+
+
+# Create a vocabulary with cos similarities to subtopics for every snippet:
+def cos_sim_vocab(compos_subtopics_vectors, compos_vectors):
+ all_sim = {}
+ for value in compos_vectors.values(): #value = all sent with metainfo for one topic
+ #print("\n\nTOPIC: ", value[0][0].split(".")[0], "\n")
+ for snippet in value: #snippet for one topic ( ['47.100', ['The|DET', ...], array([-2.11360547e+00, ...]] )
+ similarities = {}
+ for all_subt in compos_subtopics_vectors[value[0][0].split(".")[0]]:
+ sim = 1 - spatial.distance.cosine(snippet[-1], all_subt[2])
+ a = []
+ a.append(all_subt[0])
+ a.append(sim)
+ if snippet[0] not in similarities.keys():
+ similarities[snippet[0]] = []
+ similarities[snippet[0]].append(a)
+ else:
+ similarities[snippet[0]].append(a)
+ #print("\n\nSimilarities: ", similarities)
+ if value[0][0].split(".")[0] not in all_sim.keys():
+ all_sim[value[0][0].split(".")[0]] = []
+ all_sim[value[0][0].split(".")[0]].append(similarities)
+ else:
+ all_sim[value[0][0].split(".")[0]].append(similarities)
+ #print("\nAll similarities: ", all_sim)
+ return all_sim
+
+#LIKE WSD: Cluster sentences (snippets) based on cos similarity to Subtopics without sim_factor (use max cos sim); create an output file:
+def cos_sim_clustering(cos_sim_vocab, output):
+ sim_factor = 0
+ f = open(output, "a")
+ f.write("subTopicID"+" "+"resultID\n")
+ lines = []
+ result = {}
+ for topic in cos_sim_vocab.keys(): # 46
+ #print("TOPIC: ", topic)
+ if topic not in result.keys():
+ result[topic] = []
+ for snippet in cos_sim_vocab[topic]:# (list with) one vocab with keys=snippet ids
+ #print("\nSNIPPET: ", snippet)
+ voc = {}
+ max_sim = np.float(0)
+ max_id = ""
+ for simil in snippet.values():
+ #print("\nsimil: ", simil)
+ for el in simil:
+ #print(type(el[1]), type(max_sim))
+ if el[1] > max_sim and el[1] > np.float(sim_factor):
+ max_sim = el[1]
+ max_id = el[0]
+ max_value = []
+ structure = []
+ if max_sim != np.float(0):
+ structure.append(max_id)
+ structure.append(max_sim)
+ max_value.append(structure)
+ #print("\nMax_value: ", max_value)
+ if max_value != []:
+ for el in snippet.keys():
+ #print("el ", el)
+ if el not in voc.keys():
+ voc[el] = []
+ voc[el].append(max_value)
+ else:
+ voc[el].append(max_value)
+ #print("\nvoc: ", voc, "for snippet: ", snippet)
+ if voc!={}:
+ result[topic].append(voc)
+ for value in result.values():
+ #print("\nvalue:\n", value)
+ for el in value:
+ for e in el.keys():
+ #print("subtid: ", el[e][0][0][0])
+ one_line = str(el[e][0][0][0]) + " " + (str(e)+"\n")
+ lines.append(one_line)
+
+ sort = sorted(lines)
+ for el in sort:
+ f.write(el)
+ f.close()
+ return result
+
+
+
+if __name__ == "__main__":
+ path = "/proj/toyota/Pool"
+ trial_path = "/semeval-2013_task11_trial"
+ trial_topics = "/topics.txt"
+ trial_subtopics = "/subTopics.txt"
+ trial_results = "/results.txt"
+
+ topics_data, subtopics_data, results_data, number_topics, subtopics = read_data(path+trial_path+trial_topics,path+trial_path+trial_subtopics,path+trial_path+trial_results)
+ text = devide_text_into_topics(results_data, number_topics)
+ prepr_data = preprocess_data(text)
+ compos_vectors = compos_sent2vec(prepr_data, 700)
+ subtopics_vectors = devide_subtopics_into_topics(subtopics_data, number_topics)
+ compos_subtopics_vectors = compose_sent2vec_subtopics(subtopics_vectors, 700)
+ cos_sim_vocab = cos_sim_vocab(compos_subtopics_vectors, compos_vectors)
+ clusters = cos_sim_clustering(cos_sim_vocab, path+"/output_34.txt")
+ print("Done.")
+
+
--
GitLab