def load(self,
folder=None,
filenameW2V=None,
filenameMatrix=None,
filenameSearchTree=None,
filenameMessageMap=None):
if folder is None:
folder = self.folder
if filenameW2V is None:
filenameW2V = self.filenameW2V
if filenameMatrix is None:
filenameMatrix = self.filenameMatrix
if filenameSearchTree is None:
filenameSearchTree = self.filenameSearchTree
if filenameMessageMap is None:
filenameMessageMap = self.filenameMessageMap
# Load w2v object
self.W2V = Word2Vec(filenameW2V, folder=folder, load_from_file=True)
print('loaded W2V')
# Load vectorizer
self.vectorizer = SentenceVectorizer(self.W2V)
print('loaded vectorizer')
# Load search tree
self.tree = SearchTree(sentences_matrix=None,
filename=filenameSearchTree,
folder=folder)
print('loaded tree')
# Load sentence matrix and prepare submatrices
fn = os.path.join(folder, filenameMatrix)
with open(fn, 'rb') as fid:
self.sentenceMatrix = pickle.load(fid)
self.sentenceMatrixNoNans, self.correspondenceVector = remove_NaNs_from(
self.sentenceMatrix)
# Extend sentence matrix with one dimention equal to level
print('loaded matrix')
# Load message map
fn = os.path.join(folder, filenameMessageMap)
with open(fn, 'rb') as fid:
self.FilteredToUnfilteredMessageMap = pickle.load(fid)
print('loaded message map')
def load(self, folder = None, filenameW2V = None, filenameMatrix = None,
filenameSearchTree = None, filenameMessageMap = None):
if folder is None:
folder = self.folder
if filenameW2V is None:
filenameW2V = self.filenameW2V
if filenameMatrix is None:
filenameMatrix = self.filenameMatrix
if filenameSearchTree is None:
filenameSearchTree = self.filenameSearchTree
if filenameMessageMap is None:
filenameMessageMap = self.filenameMessageMap
# Load w2v object
self.W2V = Word2Vec(filenameW2V, folder = folder,
load_from_file = True)
print('loaded W2V')
# Load vectorizer
self.vectorizer = SentenceVectorizer(self.W2V)
print('loaded vectorizer')
# Load search tree
self.treeExtended = SearchTree(sentences_matrix=None,
filename=filenameSearchTree,
folder=folder)
print('loaded tree')
# Load sentence matrix and prepare submatrices
fn = os.path.join(folder, filenameMatrix)
with open(fn, 'rb') as fid:
self.sentenceMatrix = pickle.load(fid)
self.sentenceMatrixNoNans, self.correspondenceVector = remove_NaNs_from(self.sentenceMatrix)
self.MaximumVectorLength = np.max(sp.linalg.norm(self.sentenceMatrixNoNans,ord=2,axis=1))
self.sentenceMatrixNormalized = self.sentenceMatrixNoNans/self.MaximumVectorLength
# Extend sentence matrix with one dimention equal to level
self.sentenceMatrixExtended=np.hstack((self.sentenceMatrixNormalized,
self.sentenceLevelNoNans[self.correspondenceVector]*self.levelFactor))
print('loaded matrix')
# Load message map
fn = os.path.join(folder, filenameMessageMap)
with open(fn, 'rb') as fid:
self.FilteredToUnfilteredMessageMap = pickle.load(fid)
print('loaded message map')
def _initChatBot(self,
folder,
filenameClient,
filenameAgent,
filenameW2V,
filenameSearchTree,
filenameMessageMap,
load_from_file,
retrain_on_agent_file=False):
# Load messages dataframe
self.MessagesDataFrame = mdb.load_data_frame('MessagesDF.pkl', folder)
fn = os.path.join(folder, filenameMessageMap)
with open(fn, 'rb') as fid:
self.FilteredToUnfilteredMessageMap = pickle.load(fid)
# Either load everything from files:
if load_from_file:
self.W2V = Word2Vec(filenameW2V,
folder=folder,
load_from_file=True)
self.tree = SearchTree(sentences_matrix=None,
filename=filenameSearchTree,
folder=folder)
# or generate it from scratch:
else:
# Train W2V model on client file
self.W2V = Word2Vec(filenameClient, folder=folder, workers=7)
# Retrain (optionally) on agent file
if retrain_on_agent_file:
self.W2V.retrain(filenameAgent, folder=folder)
# Make the sentence vectorizer object, get sentence matrix
# and generate search tree
self.tree = SearchTree(self.matrix)
self.vectorizer = SentenceVectorizer(self.W2V)
self.matrix = self.vectorizer.get_sentences_matrix(filenameClient,
folder=folder)
Created on Fri Aug 19 16:52:19 2016 @author: piromast """ # In[1]: from vectorisations.vectorization import Word2Vec, SentenceVectorizer, SearchTree import message_db.message_db as mdb folder = 'C:\\Users\\Client\\Dropbox\\S2DS - M&S\\Data' filenameClient = '04_clientMessagesFiltered.txt' filenameAgent = '04_agentMessagesFiltered.txt' W2V = Word2Vec(filenameClient, folder=folder, workers=7) #W2V.retrain(filenameAgent, folder = folder) vectorizer = SentenceVectorizer(W2V) matrix = vectorizer.get_sentences_matrix(filenameClient, folder=folder) tree = SearchTree(matrix) # In[2]: import pandas as pd import numpy as np import os filenameIndexTable = os.path.join(folder, 'client_agent_summary.csv') indexTable = pd.read_csv(filenameIndexTable) linePositionInFilteredFile = indexTable['linePosFiltered'].values linePosition = indexTable['linePos'].values # Construct map from filtered to unfiltered file line number
def _initChatBot(self, folder, filenameClient, filenameAgent, filenameMessageDF,
filenameMessageInfoTable, load_from_file, retrain_on_agent_file = True):
# Load messages dataframe
self.MessagesDataFrame = mdb.load_data_frame(filenameMessageDF, folder)
# Import chat info table
filenameIndexTable = os.path.join(folder,filenameMessageInfoTable)
indexTable = pd.read_csv(filenameIndexTable)
# Extract sentence level information and conversation number
# (the order of each sentence in a conversation)
self.sentenceLevel = indexTable['convPos'].values.reshape((-1,1))
self.conversationNumber = indexTable['convID'].values.reshape((-1,1))
linePositionInFilteredFile = indexTable['linePosFiltered'].values
linePosition = indexTable['linePos'].values
self.sentenceLevelNoNans = self.sentenceLevel[~np.isnan(linePositionInFilteredFile)]
# Either load everything from files:
if load_from_file:
self.load()
# or generate it from scratch:
else:
# Train W2V model on client file
self.W2V = Word2Vec(filenameClient, folder = folder, workers = 7)
# Retrain (optionally) on agent file
if retrain_on_agent_file:
self.W2V.retrain(filenameAgent, folder = folder, workers = 7)
# Make the sentence vectorizer object, get sentence matrix
# and generate search tree
self.vectorizer = SentenceVectorizer(self.W2V)
self.sentenceMatrix = self.vectorizer.get_sentences_matrix(filenameClient,
folder = folder)
# Construct map from filtered to unfiltered file line number
self.FilteredToUnfilteredMessageMap = [np.int(linePosition[i])
for i in linePosition.tolist() if not np.isnan(linePositionInFilteredFile[i])]
# Normalize sentenceMatrix
self.sentenceMatrixNoNans, self.correspondenceVector = remove_NaNs_from(self.sentenceMatrix)
self.MaximumVectorLength = np.max(sp.linalg.norm(self.sentenceMatrixNoNans,ord=2,axis=1))
self.sentenceMatrixNormalized = self.sentenceMatrixNoNans/self.MaximumVectorLength
# Extend sentence matrix with one dimention equal to level
self.sentenceMatrixExtended=np.hstack((self.sentenceMatrixNormalized,
self.sentenceLevelNoNans[self.correspondenceVector]*self.levelFactor))
# Calculate compound filtered to unfiltered message map
self.FilteredToUnfilteredMessageMap = [self.FilteredToUnfilteredMessageMap[i] for i in self.correspondenceVector]
self.treeExtended = SearchTree(self.sentenceMatrixExtended)
# Instantiate input cleaner object
self.cleaner = InputCleaner(folder)
class ChattyV2(object):
def __init__(self, folder = 'C:\\Users\\Client\\Dropbox\\S2DS - M&S\\Data',
filenameClient = '05_clientMessagesFilteredfastText.txt',
filenameAgent = '05_agentMessagesFilteredfastText.txt',
filenameMessageDF = 'MessagesDF.pkl',
filenameW2V = 'W2V_Chatty2_0.mdl',
filenameSearchTree = 'SearchTree_Chatty2_0.pkl',
filenameMessageMap = 'Filtered2UnfilteredMessageMap_Chatty2_0.pkl',
filenameMessageInfoTable = 'client_agent_summary3.csv',
filenameMatrix = 'SentenceMatrix_Chatty2_0.pkl',
load_from_file = True, levelFactor = 0.35):
# Init properties
self.levelFactor = levelFactor
self.conversation_level = 0
self.name = 'Chatty 2.0'
self.folder = folder
self.filenameClient = filenameClient
self.filenameAgent = filenameAgent
self.filenameMessageDF = filenameMessageDF
self.filenameW2V = filenameW2V
self.filenameSearchTree = filenameSearchTree
self.filenameMessageMap = filenameMessageMap
self.filenameMessageInfoTable = filenameMessageInfoTable
self.filenameMatrix = filenameMatrix
# Init bot
self._initChatBot(folder, filenameClient, filenameAgent,
filenameMessageDF, filenameMessageInfoTable, load_from_file)
def _initChatBot(self, folder, filenameClient, filenameAgent, filenameMessageDF,
filenameMessageInfoTable, load_from_file, retrain_on_agent_file = True):
# Load messages dataframe
self.MessagesDataFrame = mdb.load_data_frame(filenameMessageDF, folder)
# Import chat info table
filenameIndexTable = os.path.join(folder,filenameMessageInfoTable)
indexTable = pd.read_csv(filenameIndexTable)
# Extract sentence level information and conversation number
# (the order of each sentence in a conversation)
self.sentenceLevel = indexTable['convPos'].values.reshape((-1,1))
self.conversationNumber = indexTable['convID'].values.reshape((-1,1))
linePositionInFilteredFile = indexTable['linePosFiltered'].values
linePosition = indexTable['linePos'].values
self.sentenceLevelNoNans = self.sentenceLevel[~np.isnan(linePositionInFilteredFile)]
# Either load everything from files:
if load_from_file:
self.load()
# or generate it from scratch:
else:
# Train W2V model on client file
self.W2V = Word2Vec(filenameClient, folder = folder, workers = 7)
# Retrain (optionally) on agent file
if retrain_on_agent_file:
self.W2V.retrain(filenameAgent, folder = folder, workers = 7)
# Make the sentence vectorizer object, get sentence matrix
# and generate search tree
self.vectorizer = SentenceVectorizer(self.W2V)
self.sentenceMatrix = self.vectorizer.get_sentences_matrix(filenameClient,
folder = folder)
# Construct map from filtered to unfiltered file line number
self.FilteredToUnfilteredMessageMap = [np.int(linePosition[i])
for i in linePosition.tolist() if not np.isnan(linePositionInFilteredFile[i])]
# Normalize sentenceMatrix
self.sentenceMatrixNoNans, self.correspondenceVector = remove_NaNs_from(self.sentenceMatrix)
self.MaximumVectorLength = np.max(sp.linalg.norm(self.sentenceMatrixNoNans,ord=2,axis=1))
self.sentenceMatrixNormalized = self.sentenceMatrixNoNans/self.MaximumVectorLength
# Extend sentence matrix with one dimention equal to level
self.sentenceMatrixExtended=np.hstack((self.sentenceMatrixNormalized,
self.sentenceLevelNoNans[self.correspondenceVector]*self.levelFactor))
# Calculate compound filtered to unfiltered message map
self.FilteredToUnfilteredMessageMap = [self.FilteredToUnfilteredMessageMap[i] for i in self.correspondenceVector]
self.treeExtended = SearchTree(self.sentenceMatrixExtended)
# Instantiate input cleaner object
self.cleaner = InputCleaner(folder)
def getReplyFromChatBot(self, input_message):
try:
# Clean message
input_message = self.cleaner.clean_input(input_message)
# Vectorize message
vect = self._getSentenceVector(input_message, self.conversation_level)
self.conversation_level += 1
# Find nearest message in client corpus
ind, _ = self.treeExtended.findNearestVector(vect)
print(ind)
# Unmap to account for filter-removed messages
ind = self.FilteredToUnfilteredMessageMap[ind]
# Get closest client and agent message
df = self.MessagesDataFrame
ClientMessage = mdb.get_client_message_from_dataframe(df, ind)
AgentMessage = mdb.get_agent_message_from_dataframe(df, ind)
return (AgentMessage, ClientMessage)
except Exception as inst:
print(type(inst)) # the exception instance
print(inst.args) # arguments stored in .args
print(inst) # __str__ allows args to be printed directly
def resetChat(self):
self.conversation_level = 0
def save(self, folder = None, filenameW2V = None, filenameMatrix = None,
filenameSearchTree = None, filenameMessageMap = None):
if folder is None:
folder = self.folder
if filenameW2V is None:
filenameW2V = self.filenameW2V
if filenameMatrix is None:
filenameMatrix = self.filenameMatrix
if filenameSearchTree is None:
filenameSearchTree = self.filenameSearchTree
if filenameMessageMap is None:
filenameMessageMap = self.filenameMessageMap
self.W2V.save(filenameW2V, folder)
print('saved W2V')
self.treeExtended.save_tree(filenameSearchTree, folder)
print('saved tree')
with open(os.path.join(folder,filenameMessageMap),'wb') as f:
pickle.dump(self.FilteredToUnfilteredMessageMap, f, protocol=2)
print('saved message map')
with open(os.path.join(folder,filenameMatrix),'wb') as f:
pickle.dump(self.sentenceMatrix, f, protocol=2)
print('saved matrix')
def load(self, folder = None, filenameW2V = None, filenameMatrix = None,
filenameSearchTree = None, filenameMessageMap = None):
if folder is None:
folder = self.folder
if filenameW2V is None:
filenameW2V = self.filenameW2V
if filenameMatrix is None:
filenameMatrix = self.filenameMatrix
if filenameSearchTree is None:
filenameSearchTree = self.filenameSearchTree
if filenameMessageMap is None:
filenameMessageMap = self.filenameMessageMap
# Load w2v object
self.W2V = Word2Vec(filenameW2V, folder = folder,
load_from_file = True)
print('loaded W2V')
# Load vectorizer
self.vectorizer = SentenceVectorizer(self.W2V)
print('loaded vectorizer')
# Load search tree
self.treeExtended = SearchTree(sentences_matrix=None,
filename=filenameSearchTree,
folder=folder)
print('loaded tree')
# Load sentence matrix and prepare submatrices
fn = os.path.join(folder, filenameMatrix)
with open(fn, 'rb') as fid:
self.sentenceMatrix = pickle.load(fid)
self.sentenceMatrixNoNans, self.correspondenceVector = remove_NaNs_from(self.sentenceMatrix)
self.MaximumVectorLength = np.max(sp.linalg.norm(self.sentenceMatrixNoNans,ord=2,axis=1))
self.sentenceMatrixNormalized = self.sentenceMatrixNoNans/self.MaximumVectorLength
# Extend sentence matrix with one dimention equal to level
self.sentenceMatrixExtended=np.hstack((self.sentenceMatrixNormalized,
self.sentenceLevelNoNans[self.correspondenceVector]*self.levelFactor))
print('loaded matrix')
# Load message map
fn = os.path.join(folder, filenameMessageMap)
with open(fn, 'rb') as fid:
self.FilteredToUnfilteredMessageMap = pickle.load(fid)
print('loaded message map')
def changeLevelFactor(self, levelFactor):
self.levelFactor = levelFactor
# Extend sentence matrix with one dimention equal to level
self.sentenceMatrixExtended=np.hstack((self.sentenceMatrixNormalized,
self.sentenceLevelNoNans[self.correspondenceVector]*self.levelFactor))
self.treeExtended = SearchTree(self.sentenceMatrixExtended)
print('Chenged level to'+str(self.levelFactor))
def _getSentenceVector(self, text, level):
vect = self.vectorizer.get_sentence_vector(text)/self.MaximumVectorLength
vect = np.hstack((vect,level*self.levelFactor))
return vect
def __get_score(self, train_index, test_index, size, window , min_count, workers):
"""
Private method doing the hard work described in test method
"""
for i, inputFile in enumerate(self.inputFileNames):
print('Training iteration:',str(i))
sentenceGeneratorTrainSetInput = SentencesGeneratorInList(inputFile, train_index)
if i == 0: # First train is done creating instance of Word2Vec
W2VInput = gensim.models.Word2Vec(sentenceGeneratorTrainSetInput, size, window, min_count, workers)
W2VInput = Word2Vec(W2VInput)
else: # Subsequent training is done calling train method
W2VInput.model.train(sentenceGeneratorTrainSetInput, size, window, min_count, workers)
# Generate input and output sentence generators
sentenceGeneratorTestSetInput = SentencesGeneratorInList(self.inputFileNames[0], test_index)
sentenceGeneratorTestSetOutput = SentencesGeneratorInList(self.outputFileName, test_index)
# Train input and output word2vec models
W2VOutput = gensim.models.Word2Vec(sentenceGeneratorTestSetOutput,
size, window, min_count, workers)
W2VOutput = Word2Vec(W2VOutput)
vectorizerInput = SentenceVectorizer(W2VInput)
vectorizerOutput = SentenceVectorizer(W2VOutput)
# calculate matrix of sentence vectors of whole text file
matrixInput = vectorizerInput.get_sentences_matrix(sentenceGenerator =
sentenceGeneratorTestSetInput)
matrixInput,_ = remove_NaNs_from(matrixInput)
matrixOutput = vectorizerOutput.get_sentences_matrix(sentenceGenerator =
sentenceGeneratorTestSetOutput)
matrixOutput,_ = remove_NaNs_from(matrixOutput)
# Generate search tree (to find nearest vector)
treeInput = SearchTree(matrixInput)
#treeOutput = SearchTree(matrixOutput)
n = 0 # Sentence number counter
# Initialize different distance measures
# (input and output eucledian distance and cos similarity)
total_input_distance = 0
total_output_distance = 0
total_cosine_similarity = 0
for vIn, vOut in zip(matrixInput, matrixOutput):
#print('Print:',str(i))
n += 1
# Find nearest vector to input in sentence corpus
indexInput, distanceInput = treeInput.findNearestVector(vIn)
# Accumulate input vector distance to nearest neighbout
total_input_distance += distanceInput
# Calculating output vector corresponding
# to nearest input sentence
vOutPredicted = matrixOutput[indexInput,]
# Caluclating eucledian distance
# between observed and predicted output
distanceOutput = spd.euclidean(vOut, vOutPredicted)
total_output_distance += distanceOutput
# Calculating cosine similarity between observation and prediction
cosineSimilarity = spd.cosine(vOut, vOutPredicted)
total_cosine_similarity += cosineSimilarity
return (total_cosine_similarity,
total_input_distance, total_output_distance)
@author: piromast """ # In[1]: from vectorisations.vectorization import Word2Vec, SentenceVectorizer, SearchTree, remove_NaNs_from import message_db.message_db as mdb import os from preprocessing.filtering import filter_line, build_names, NLTKPreprocessor folder = 'C:\\Users\\Client\\Dropbox\\S2DS - M&S\\Data' #filenameAgent = '04_agentMessagesFilteredTrigramLemmatized.txt' filenameClient = '04_clientMessagesFilteredOld.txt' filenameAgent = '04_agentMessagesFilteredOld.txt' W2V = Word2Vec(filenameClient, folder = folder) vectorizer = SentenceVectorizer(W2V) sentenceMatrix = vectorizer.get_sentences_matrix(filenameClient, folder = folder) tree = SearchTree(sentenceMatrix) # In[2]: import pandas as pd import numpy as np import scipy as sp filenameIndexTable = os.path.join(folder,'client_agent_summaryOld.csv') indexTable = pd.read_csv(filenameIndexTable) sentenceLevel = indexTable['convPos'].values.reshape((-1,1)) conversationNumber = indexTable['convID'].values.reshape((-1,1)) linePositionInFilteredFile = indexTable['linePosFiltered'].values
class ChattyV1(object):
def __init__(
self,
folder='C:\\Users\\Client\\Dropbox\\S2DS - M&S\\Data',
filenameClient='04_clientMessagesFiltered.txt',
filenameAgent='04_agentMessagesFiltered.txt',
filenameMessageDF='MessagesDF.pkl',
filenameW2V='W2V_Chatty1_0.mdl',
filenameSearchTree='SearchTree_Chatty1_0.pkl',
filenameMessageMap='Filtered2UnfilteredMessageMap_Chatty1_0.pkl',
load_from_file=True):
# Init bot
self._initChatBot(folder, filenameClient, filenameAgent, filenameW2V,
filenameSearchTree, filenameMessageMap,
load_from_file)
def _initChatBot(self,
folder,
filenameClient,
filenameAgent,
filenameW2V,
filenameSearchTree,
filenameMessageMap,
load_from_file,
retrain_on_agent_file=False):
# Load messages dataframe
self.MessagesDataFrame = mdb.load_data_frame('MessagesDF.pkl', folder)
fn = os.path.join(folder, filenameMessageMap)
with open(fn, 'rb') as fid:
self.FilteredToUnfilteredMessageMap = pickle.load(fid)
# Either load everything from files:
if load_from_file:
self.W2V = Word2Vec(filenameW2V,
folder=folder,
load_from_file=True)
self.tree = SearchTree(sentences_matrix=None,
filename=filenameSearchTree,
folder=folder)
# or generate it from scratch:
else:
# Train W2V model on client file
self.W2V = Word2Vec(filenameClient, folder=folder, workers=7)
# Retrain (optionally) on agent file
if retrain_on_agent_file:
self.W2V.retrain(filenameAgent, folder=folder)
# Make the sentence vectorizer object, get sentence matrix
# and generate search tree
self.tree = SearchTree(self.matrix)
self.vectorizer = SentenceVectorizer(self.W2V)
self.matrix = self.vectorizer.get_sentences_matrix(filenameClient,
folder=folder)
def getReplyFromChatBot(self, input_message):
try:
# Vectorize message
vect = self.vectorizer.get_sentence_vector(input_message)
print(vect)
# Find nearest message in client corpus
ind, _ = self.tree.findNearestVector(vect)
print(ind)
# Unmap to account for filter-removed messages
ind = self.FilteredToUnfilteredMessageMap[ind]
# Get closest client and agent message
df = self.MessagesDataFrame
ClientMessage = mdb.get_client_message_from_dataframe(df, ind)
AgentMessage = mdb.get_agent_message_from_dataframe(df, ind)
return (AgentMessage, ClientMessage)
except Exception as inst:
print type(inst) # the exception instance
print inst.args # arguments stored in .args
print inst # __str__ allows args to be printed directly
# Add app folder to PYTHONPATH
from vectorisations.vectorization import Word2Vec, SentenceVectorizer, SearchTree
import os.path
filename = os.path.join("/Users", "piromast", "Dropbox", "S2DS - M&S", "Data",
"04_clientMessagesFiltered.txt")
W2V = Word2Vec(filename) # Train word2vec with file in filename
vectorizer = SentenceVectorizer(W2V) # make sentence vectorizer object
matrix = vectorizer.get_sentences_matrix(
filename) # calculate matrix of sentence vectors of whole text file
tree = SearchTree(matrix) # Generate search tree (to find nearest vector)
# Lookup example
v = vectorizer.get_sentence_vector(
"I lost my order") # Convert sentence to vector
index, distance = tree.findNearestVector(v)