def analogy(self, x, y, z):
"""
y is to ??? what z is to x
:param x:
:param y:
:param z:
:return:
"""
x = dpu.encode_cell(x)
y = dpu.encode_cell(y)
z = dpu.encode_cell(z)
indexes, metrics = self.M.analogy(pos=[x, y], neg=[z], n=10)
res = self.M.generate_response(indexes, metrics).tolist()
return res
def evaluate_table_attributes(api, args, table_df, entity_attribute, table_name, target_attribute, ranking_size=10, debug=True):
"""
Given a table dataframe (pandas), an entity attribute and a target attribute, makes questions and records the
position of the found answers
:param api: relational_emb api object
:param args: arguments passed to the program
:param table_df: dataframe holding the table of interest (pandas dataframe)
:param entity_attribute: attribute in table_df from where to draw entities
:param target_attribute: attribute in table_df for which we want to predict the answer
:param ranking_size: the size of the ranking
:return:
"""
should_sample = args.sample
evaluation_results = defaultdict(int)
num_questions = 0
key_error = 0
qs = 0
# Iterate rows of table to draw entity and target_attribute
for index, el in table_df.iterrows():
if should_sample:
if random.randint(1, 10) > 1:
continue
qs += 1
if (qs % 100) == 0:
print("#q: " + str(qs))
entity = dpu.encode_cell(el[entity_attribute])
ground_truth = dpu.encode_cell(el[target_attribute])
try:
ranking_result = api.concept_qa(entity, table_name, target_attribute, n=ranking_size)
# Record in which position does the right answer appear, if it does
for index, entry in enumerate(ranking_result):
answer, score = entry
found = (answer == ground_truth)
if found:
evaluation_results[index] += 1
break
num_questions += 1 # One more question
except KeyError:
key_error += 1
# We only recorded the first position where an answer appears, accumulate results to get easy-to-interpret perc
total_hits = 0
for index in range(ranking_size):
evaluation_results[index] += total_hits
total_hits = evaluation_results[index]
return evaluation_results, num_questions, key_error
def window_column(paths, output_file, debug=False):
try:
os.remove(output_file)
except FileNotFoundError:
print("Creating new file for writing data")
total = len(paths)
current = 0
for path in paths:
if debug:
print(str(current) + "/" + str(total))
current += 1
df = pd.read_csv(path, encoding='latin1')
# Check for valid relations only
if not dpu.valid_relation(df):
continue
columns = df.columns
f = csv.writer(open(output_file, 'a'),
delimiter=',',
quotechar='\"',
quoting=csv.QUOTE_MINIMAL)
# Columns
for c in columns:
col_data = df[c]
row = [
dpu.encode_cell(cell_value) for cell_value in col_data
if dpu.valid_cell(cell_value)
]
if len(row) > 0:
f.writerow(row)
# TODO: why is it necessary to indicate end of relation?
f.writerow(["~R!RR*~"])
def column_avg_unique_composition(df, we_model):
column_we = dict()
columns = df.columns
missing_words = 0
for c in columns:
col_wes = []
value = df[c].unique()
for el in value:
# Check validity of cell
if not dpu.valid_cell(el):
continue
el = dpu.encode_cell(el)
if " " in el:
els = el.split(" ")
vector = we_model.get_vector(els[0])
missing_words_mini = 0
for ee in range(1, len(els)):
try:
vector += we_model.get_vector(els[1])
except KeyError:
missing_words += 1
missing_words_mini += 1
vector /= (len(els) - missing_words_mini)
else:
try:
vector = we_model.get_vector(el)
except KeyError:
missing_words += 1
continue
col_wes.append(vector)
col_wes = np.asarray(col_wes)
col_we = np.mean(col_wes, axis=0)
column_we[c] = col_we
return column_we, missing_words
def row_avg_composition(df, we_model):
missing_words = 0
row_we_dict = dict()
columns = df.columns
for i, row in df.iterrows():
row_wes = []
for c in columns:
# Check validity of cell
if not dpu.valid_cell(row[c]):
continue
el = dpu.encode_cell(row[c])
if " " in el:
els = el.split(" ")
vector = we_model.get_vector(els[0])
missing_words_mini = 0
for ee in range(1, len(els)):
try:
vector += we_model.get_vector(els[1])
except KeyError:
missing_words += 1
missing_words_mini += 1
vector /= (len(els) - missing_words_mini)
else:
try:
vector = we_model.get_vector(el)
except KeyError:
missing_words += 1
continue
row_wes.append(vector)
row_wes = np.asarray(row_wes)
row_we = np.mean(row_wes, axis=0)
row_we_dict[i] = row_we
return row_we_dict, missing_words
def vector_for_entity(self, cell=None, attribute=None, table=None):
vec = None
if cell:
cell = dpu.encode_cell(cell)
vec = self.M.get_vector(cell)
elif table:
table = dpu.encode_cell(table)
if attribute:
attribute = dpu.encode_cell(attribute)
vec = self.RE[table]["columns"][attribute]
else:
vec = self.RE[table]["vector"]
elif attribute:
attribute = dpu.encode_cell(attribute)
print("Not supported yet!")
return
return vec
def entity_to_attribute(self, entities, n=2, simf=SIMF.COSINE):
res = []
for entity in entities:
entity = dpu.encode_cell(entity)
vec_e = self.M.get_vector(entity)
topk = self.topk_columns(vec_e, k=n, simf=simf)
res.append((entity, topk))
return res
def _read_columns_from_dataframe(df, columns):
for c in columns:
data_values = df[c]
for cell_value in data_values:
# We check the cell value is valid before continuing
if not dpu.valid_cell(cell_value):
continue
cell_value = dpu.encode_cell(cell_value)
yield cell_value
def topk_similar_vectors(self, input_string, k=10, simf=SIMF.COSINE):
el = dpu.encode_cell(input_string)
indexes = []
metrics = []
if simf == SIMF.COSINE:
indexes, metrics = self.M.cosine(el, n=k)
elif simf == SIMF.EUCLIDEAN:
indexes, metrics = self.M.euclidean(el, n=k)
res = self.M.generate_response(indexes, metrics).tolist()
return res
def _read_rows_from_dataframe(df, columns):
for index, el in df.iterrows():
for c in columns:
cell_value = el[c]
# We check the cell value is valid before continuing
if not dpu.valid_cell(cell_value):
continue
# If valid, we clean and format it and return it
cell_value = dpu.encode_cell(cell_value)
yield cell_value
def concept_qa(self, entity, relation, attribute, n=20, simf=SIMF.COSINE):
entity = dpu.encode_cell(entity)
if " " in entity:
# We have spaces/words now!!
entity_words = entity.split(" ")
indexes = []
metrics = []
if simf == SIMF.COSINE:
if " " in entity:
indexes, metrics = self.M.cosine_array(entity_words, n=n)
# print(indexes)
else:
indexes, metrics = self.M.cosine(entity, n=n)
elif simf == SIMF.EUCLIDEAN:
indexes, metrics = self.M.euclidean(entity, n=n)
#SPACES UNIMPLEMENTED TODO:
res = self.M.generate_response(indexes, metrics).tolist()
res = [(e, self.re_range_score(score)) for e, score in res]
vec_attribute = self.RE[relation]["columns"][attribute]
if type(vec_attribute) is not np.ndarray:
# print(attribute)
return []
# vec_attribute = self.RE[relation+"."+attribute]
candidate_attribute_sim = []
for e, score in res:
vec_e = self.M.get_vector(e) # no need to normalize e --- it's already normalized
similarity_to_attr = 0
if simf == SIMF.COSINE:
similarity_to_attr = np.dot(vec_e, vec_attribute)
similarity_to_attr = self.re_range_score(similarity_to_attr)
# distance_to_attr = cosine(vec_e, vec_attribute)
elif simf == SIMF.EUCLIDEAN:
similarity_to_attr = 1 - euclidean(vec_e, vec_attribute)
# avg distance between original entity to each ranking entity and each ranking entity and target attr
similarity = (similarity_to_attr + score) / 2
candidate_attribute_sim.append((e, similarity))
candidate_attribute_sim = sorted(candidate_attribute_sim, key=lambda x: x[1], reverse=True)
return candidate_attribute_sim
def _concept_qa_no_avg_rerank(self, entity, relation, attribute, n=20, simf=SIMF.COSINE):
entity = dpu.encode_cell(entity)
indexes = []
metrics = []
if simf == SIMF.COSINE:
indexes, metrics = self.M.cosine(entity, n=n)
elif simf == SIMF.EUCLIDEAN:
indexes, metrics = self.M.euclidean(entity, n=n)
res = self.M.generate_response(indexes, metrics).tolist()
vec_attribute = self.RE[relation]["columns"][attribute]
# vec_attribute = self.RE[relation+"."+attribute]
candidate_attribute_sim = []
for e, score in res:
vec_e = self.M.get_vector(e) # no need to normalize e --- it's already normalized
similarity = 0
if simf == SIMF.COSINE:
similarity = np.dot(vec_e, vec_attribute)
similarity = self.re_range_score(similarity)
elif simf == SIMF.EUCLIDEAN:
similarity = 1 - euclidean(vec_e, vec_attribute)
candidate_attribute_sim.append((e, similarity))
candidate_attribute_sim = sorted(candidate_attribute_sim, key=lambda x: x[1], reverse=True)
return candidate_attribute_sim
df = pd.read_csv(csv_filepath, encoding='latin1')
columns = list(df.columns.values)
columnsize = len(columns)
fh.write(f"D:Columns: {columns} \n")
fh.flush()
for index, el in df.iterrows():
if random.randint(1, 10) > 1:
continue
for i in range(3):
c = random.randint(0, columnsize - 1)
target_column = random.randint(0, columnsize - 1)
#SHOULD I CHECK IF
if c == target_column:
continue
# try:
value = dpu.encode_cell(el[c])
if len(
value
) < 4 or "/" in value: #We're only going 1 direction with the testing data AND NO DATES
continue
expected = dpu.encode_cell(el[target_column])
# print(value,expected)
try:
res = api.concept_qa(value,
csv_file,
columns[target_column],
n=RELEVANTS[-1])
y = 0
ind = 0
def similarity_between(self, entity1, entity2, simf=SIMF.COSINE):
x = dpu.encode_cell(entity1)
y = dpu.encode_cell(entity2)
vec_x = self.M.get_vector(x)
vec_y = self.M.get_vector(y)
return self.similarity_between_vectors(vec_x, vec_y, simf=simf)