def __init__(self):
self.parser = Parser()
self.bids = Crawler().finished_set
if os.path.exists(os.path.join(Matrix.dirname, 'dictionary.pkl')):
self.dictionary = None
else:
self.dictionary = self.load_dictionary()
if os.path.exists(os.path.join(Matrix.dirname, 'higher_tier_dictionary.pkl')):
self.higher_tier_dictionary = None
else:
self.load_higher_tier_dictionary()
self.term_to_row = self.load_term_to_row()
self.bid_to_col = self.load_bid_to_col()
self.N = len(self.bid_to_col) # Total number of documents
self.M = len(self.term_to_row) # Total number of terms
self.col_to_bid = self.load_col_to_bid()
self.row_to_term = self.load_row_to_term() # Used in similar search
self.term_bid_matrix = self.load_term_bid_matrix()
self.title_term_bid_matrix = self.load_title_term_bid_matrix()
# Csr sparse matrix sed for slice term rows
self.term_bid_matrix_csr = self.term_bid_matrix.tocsr()
self.title_term_bid_matrix_csr = self.title_term_bid_matrix.tocsr()
self.stop_words = self.load_stop_words()
# Tiers, tier[0] is the lowest tier
self.mats_csc = (self.term_bid_matrix,
self.title_term_bid_matrix)
self.mats_csr = (self.term_bid_matrix_csr,
self.title_term_bid_matrix_csr)
def __init__(self, documents=[], transforms=[TFIDF, LSA]):
self.collection_of_document_term_vectors = []
self.parser = Parser()
if len(documents) > 0:
self._build(documents, transforms)
class Matrix(object):
k1, k2, b = 1.5, 1.5, 0.75 # Parameters
modified = {'dictionary': False, 'higher_tier_dictionary': False} # Member vars modified mark
dirname = os.path.dirname(os.path.realpath(__file__))
def __init__(self):
self.parser = Parser()
self.bids = Crawler().finished_set
if os.path.exists(os.path.join(Matrix.dirname, 'dictionary.pkl')):
self.dictionary = None
else:
self.dictionary = self.load_dictionary()
if os.path.exists(os.path.join(Matrix.dirname, 'higher_tier_dictionary.pkl')):
self.higher_tier_dictionary = None
else:
self.load_higher_tier_dictionary()
self.term_to_row = self.load_term_to_row()
self.bid_to_col = self.load_bid_to_col()
self.N = len(self.bid_to_col) # Total number of documents
self.M = len(self.term_to_row) # Total number of terms
self.col_to_bid = self.load_col_to_bid()
self.row_to_term = self.load_row_to_term() # Used in similar search
self.term_bid_matrix = self.load_term_bid_matrix()
self.title_term_bid_matrix = self.load_title_term_bid_matrix()
# Csr sparse matrix sed for slice term rows
self.term_bid_matrix_csr = self.term_bid_matrix.tocsr()
self.title_term_bid_matrix_csr = self.title_term_bid_matrix.tocsr()
self.stop_words = self.load_stop_words()
# Tiers, tier[0] is the lowest tier
self.mats_csc = (self.term_bid_matrix,
self.title_term_bid_matrix)
self.mats_csr = (self.term_bid_matrix_csr,
self.title_term_bid_matrix_csr)
def __del__(self):
self.save_dictionary()
self.save_higher_tier_dictionary()
self.save_term_to_row()
self.save_bid_to_col()
self.save_term_bid_matrix()
self.save_title_term_bid_matrix()
self.save_row_to_term()
self.save_col_to_bid()
self.save_stop_words()
def saving(func):
@wraps(func)
def wrapper(*args, **kwargs):
stuff = func.__name__[5:]
if Matrix.modified[stuff]:
filename = os.path.join(Matrix.dirname, stuff + '.pkl')
print('Saving {} ...'.format(stuff))
with open(filename, 'rb') as f:
pickle.dump(getattr(self, stuff), f)
print('{} saved.'.format(stuff))
return wrapper
def loading(func):
@wraps(func)
def wrapper(*args, **kwargs):
stuff = func.__name__[5:]
filename = os.path.join(Matrix.dirname, stuff + '.pkl')
if os.path.exists(filename):
print('Loading {} ...'.format(stuff))
with open(filename, 'rb') as f:
Matrix.modified[stuff] = False
r = pickle.load(f)
print('{} loaded.'.format(stuff))
else:
print('Building {} ...'.format(stuff))
r = func(*args, **kwargs)
Matrix.modified[stuff] = True
print('{} built.'.format(stuff))
return r
return wrapper
@loading
def load_dictionary(self):
dictionary = defaultdict(Counter) # {term -> {bid -> frequency}}
for bid in self.bids:
path = os.path.join('text', bid + '.txt')
try:
with open(path, 'r') as f:
text = f.read()
except FileNotFoundError as E:
with open('parser.log', 'a') as f:
f.write('{}\n'.format(E))
for token in self.parser.parse(text):
dictionary[token][bid] += 1
return dictionary
@loading
def load_higher_tier_dictionary(self):
higher_tier_dictionary = defaultdict(Counter)
conn = sqlite3.connect('books.db')
c = conn.cursor()
c.execute('SELECT bid, title, author, tags FROM books')
while True:
record = c.fetchone()
if record:
bid, title, author, tags = record
text = '{} {} {}'.format(title, author, tags)
for token in self.parser.parse(text):
higher_tier_dictionary[token][str(bid)] += 1
else:
break
conn.close()
return higher_tier_dictionary
@loading
def load_term_to_row(self):
return {term: row for row, term in enumerate(self.dictionary)}
@loading
def load_bid_to_col(self):
return {bid: col for col, bid in enumerate(self.bids)}
@loading
def load_term_bid_matrix(self):
return self._dict_to_matrix(self.dictionary)
@loading
def load_title_term_bid_matrix(self):
return self._dict_to_matrix(self.higher_tier_dictionary)
@loading
def load_row_to_term(self):
return {row: term for term, row in self.term_to_row.items()}
@loading
def load_col_to_bid(self):
return {col: bid for bid, col in self.bid_to_col.items()}
@loading
def load_stop_words(self, N=25000):
"""Return a boolean vector in which True entries stand for lower occurence words,
while False entries stand for higher occurence words.
N is the threshold values that how many stop words will be selected"""
def df(term):
return len(self.dictionary[term])
def nlargest_df(N):
return heapq.nlargest(N, ((df(term), term) for term in self.dictionary))
v = np.ones((self.M, 1), dtype=bool) # High occurence words (stop words) vector
for df, term in nlargest_df(N):
v[self.term_to_row[term], 0] = False
return v
def _dict_to_matrix(self, dictionary):
"""Get term/bid matrix."""
m = sparse.dok_matrix((self.M, self.N), dtype=int)
for term, bids in dictionary.items():
row = self.term_to_row[term]
for bid, freq in bids.items():
col = self.bid_to_col[bid]
m[row, col] = freq
return m.tocsc()
@saving
def save_term_bid_matrix(self):
pass
@saving
def save_title_term_bid_matrix(self):
pass
@saving
def save_term_to_row(self):
pass
@saving
def save_bid_to_col(self):
pass
@saving
def save_dictionary(self):
pass
@saving
def save_higher_tier_dictionary(self):
pass
@saving
def save_row_to_term(self):
pass
@saving
def save_col_to_bid(self):
pass
@saving
def save_stop_words(self):
pass
def tf(self, freq):
return 1 + np.log10(freq)
def idf(self, term, tier=0):
matrix = self.mats_csr[tier]
def df(term):
row_num = self.term_to_row[term]
row = matrix.getrow(row_num)
return row.nnz
return np.log10(self.N / df(term))
def make_query_vector(self, query_terms, tier):
"""Generator a query tf-idf vector according to terms counter.
Parameter query_terms is a Counter data structure."""
q = sparse.dok_matrix((1, self.M)) # 1 * #term_num# vector
# Build query tf-idf vector
for term, freq in query_terms.items():
q[0,self.term_to_row[term]] = self.tf(freq) * self.idf(term, tier)
q = normalize(q.tocsr(), axis=1, copy=False)
return q
def boolean_search(self, query_terms, tier):
"""Search a query according to its ocurrence in the documents and return matched results.
When tier is 'l', search the lower tier, and if tier is 'h' search the higher tier.
Return the column indices of matched books."""
matrix = self.mats_csr[tier]
docs = np.zeros(self.N, dtype=bool) # Get union of docs containing query
for term in query_terms:
row = self.term_to_row.get(term)
if row is None:
continue
else:
np.logical_or(docs, matrix.getrow(row).toarray()[0], docs)
return docs.nonzero()[0]
def make_matched_matrix(self, cols, tier):
matrix = self.mats_csc[tier]
# Pick matched columns
m = matrix[:,cols]
m.data = self.tf(m.data)
m = normalize(m.T, axis=1, copy=False).T # No copy to make faster
return m
def tiered_search(self, query, K=50):
query_terms = self.parser.parse_query(query)
# Higher tier
try:
hres = heapq.nlargest(K, self.cos_sim_search(query_terms, tier=1))
except: # Terms don't occur in titles, continue to search the lower tier
hres = []
lres = []
if len(hres) < K: # Lower tier
try:
hres_bids = [r[1] for r in hres]
# Docs that hasn't been retrieve
lres = heapq.nlargest(K - len(hres), # Find at most K docs
filter(lambda pair: pair[0] not in hres_bids,
self.cos_sim_search(query_terms, tier=0)))
except: # Term don't occur in all docs, return nothing
return ()
return (int(r[1]) for r in itertools.chain(hres, lres)) # Return K highest ranked bids
def cos_sim_search(self, query_terms, tier=0):
"""Return an unsorted iterator of (similarity, bid) tuples given the query terms.
If qv is not provided, it will be computed.
Default tier is 0"""
qv = self.make_query_vector(query_terms, tier)
cols = self.boolean_search(query_terms, tier)
m = self.make_matched_matrix(cols, tier)
cos_sims = qv.dot(m).toarray()[0]
bids = (self.col_to_bid[col] for col in cols)
return zip(cos_sims, bids)
def find_most_similar(self, bids, K_sim=14):
"""Return the bid of the most similar book to parameter bid except the given bid."""
termv = sparse.csc_matrix((self.M, 1), dtype=int)
for bid in bids:
col_num = self.bid_to_col.get(str(bid))
if col_num is not None:
termv = termv + self.term_bid_matrix.getcol(col_num)
if termv.nnz == 0:
return ()
termva = termv.toarray() # Generate a vector for terms
stop_words_removed = np.logical_and(termva, self.stop_words)
nonzero = stop_words_removed.nonzero()[0] # Nonzero indices
rest_term_rows = self.term_bid_matrix_csr[nonzero]
docs = np.zeros(self.N, dtype=bool)
for row in rest_term_rows:
np.logical_or(docs, row.toarray()[0], docs)
cols = docs.nonzero()[0]
matched_matrix = self.term_bid_matrix[:,cols]
termv.data = self.tf(termv.data) * np.array([self.idf(self.row_to_term[row])
for row in termv.indices])
termv = normalize(termv.T, axis=1, copy=False)
matched_matrix.data = self.tf(matched_matrix.data)
matched_matrix = normalize(matched_matrix.T, axis=1, copy=False).T
cos_sims = termv.dot(matched_matrix).toarray()[0]
found_bids = (self.col_to_bid[col] for col in cols)
return islice((int(r[1])
for r in heapq.nlargest(K_sim, zip(cos_sims, found_bids))
if int(r[1]) not in bids),
9)
def find_most_similar_tags(self, bid, K_sim=10):
conn = sqlite3.connect('books.db')
c = conn.cursor()
c.execute('SELECT tags FROM books WHERE bid = ?', (int(bid),))
query_terms = self.parser.parse_query(c.fetchone()[0])
return (int(r[1]) for r in heapq.nlargest(K_sim, self.cos_sim_search(query_terms)))
def bm25_search(self, query):
pass
class VectorSpace:
collection_of_document_term_vectors = []
vector_index_to_keyword_mapping = []
parser = None
def __init__(self, documents=[], transforms=[TFIDF, LSA]):
self.collection_of_document_term_vectors = []
self.parser = Parser()
if len(documents) > 0:
self._build(documents, transforms)
def _build(self, documents, transforms):
"""
Create the vector space for the passed document strings
"""
self.vector_index_to_keyword_mapping = self._get_vector_keyword_index(documents)
pprint(self.vector_index_to_keyword_mapping)
matrix = [self._make_vector(document) for document in documents]
matrix = reduce(lambda matrix, transform: transform(matrix).transform(), transforms, matrix)
self.collection_of_document_term_vectors = matrix
def _get_vector_keyword_index(self, document_list):
"""
Zwraca słownik zawierający pary "słowo" : pozycja w liście rdzeni
"""
vocabulary_list = self.parser.tokenise_and_remove_stop_words(document_list)
unique_vocabulary_list = self._remove_duplicates(vocabulary_list)
vector_index = {}
offset = 0
# Associate a position with the keywords
# which maps to the dimension on the vector used to represent this word
for word in unique_vocabulary_list:
vector_index[word] = offset
offset += 1
return vector_index
def related(self, document_id):
""" find documents that are related to the document indexed by passed Id within the document Vectors"""
ratings = [self._cosine(self.collection_of_document_term_vectors[document_id], document_vector) for
document_vector in self.collection_of_document_term_vectors]
ratings.sort(reverse=True)
return ratings
def search(self, searchList):
""" search for documents that match based on a list of terms """
queryVector = self._build_query_vector(searchList)
ratings = [self._cosine(queryVector, documentVector) for documentVector in self.collection_of_document_term_vectors]
ratings.sort(reverse=True)
return ratings
def _make_vector(self, word_string):
""" @pre: unique(vectorIndex) """
vector = [0] * len(self.vector_index_to_keyword_mapping)
word_list = self.parser.tokenise_and_remove_stop_words(word_string.split(" "))
# Term Count Model
for word in word_list:
vector[self.vector_index_to_keyword_mapping[word]] += 1;
return vector
def _build_query_vector(self, term_list):
"""
convert query string into a term vector
"""
query = self._make_vector(" ".join(term_list))
return query
def _remove_duplicates(self, list):
"""
Usuwanie duplikatów słów
"""
return set((item for item in list))
def _cosine(self, vector1, vector2):
""" related documents j and q are in the concept space by comparing the vectors :
cosine = ( V1 * V2 ) / ||V1|| x ||V2|| """
dot_val = dot(vector1, vector2)
norms = norm(vector1) * norm(vector2)
if norms == 0.0:
return 0.0
else:
return float(dot_val / norms)
