# --------------------------------------------------------------------------------------------------------------------
print(u'Crunching numbers, sit tight...')
# errors = codecs.open(u"errors.tsv", u"w", encoding=u"utf-8")
# Uncomment the above line for error diagnostics, also the section below.
conn = sqlite3.connect(u'../data/geonames.db')
file_name = u"data/eval_" + test_data + u".txt"
final_errors = []
print("Processing file..." + file_name)
for prediction, (y, name, context) in zip(
model.predict_generator(
generate_arrays_from_file(file_name, word_to_index, train=False),
steps=int(check_output([u"wc", file_name]).split()[0]) /
BATCH_SIZE,
verbose=True), generate_strings_from_file(file_name)):
prediction = index_to_coord(REVERSE_MAP_2x2[np.argmax(prediction)], 2)
candidates = get_coordinates(conn.cursor(), name)
if len(candidates) == 0:
print(u"Don't have an entry for", name, u"in GeoNames")
raise Exception(u"Check your database, buddy :-)")
# candidates = [candidates[0]] # Uncomment for population heuristic.
# THE ABOVE IS THE POPULATION ONLY BASELINE IMPLEMENTATION
#print("Prediction..."+" ".join(y)+"..."+name+".."+str(context))
print("Prediction..." + " " + str(y) + "...." + name + ".." + context)
best_candidate = []
max_pop = candidates[0][2]
bias = 0.905 # the Bias parameter in the paper
for candidate in candidates:
def geoparse(text, result_string, model, word_to_index):
"""
This function allows one to geoparse text i.e. extract toponyms (place names) and disambiguate to coordinates.
:param text: to be parsed
:return: currently only prints results to the screen, feel free to modify to your task
"""
nlp = spacy.load(
u'en'
) # or spacy.load(u'en') depending on your Spacy Download (simple or full)
conn = sqlite3.connect(
u'/opt/gsda/EUPEG/Geoparsers/camCoder/data/geonames.db').cursor(
) # this DB can be downloaded using the GitHub link
padding = nlp(u"0")[0] # Do I need to explain? :-)
doc = nlp(text) # NER with Spacy NER
for entity in doc.ents:
if entity.label_ in [u"GPE", u"FACILITY", u"LOC", u"FAC", u"LOCATION"]:
name = entity.text if not entity.text.startswith(
'the') else entity.text[4:].strip()
start = entity.start_char if not entity.text.startswith(
'the') else entity.start_char + 4
end = entity.end_char
near_inp = pad_list(CONTEXT_LENGTH / 2, [x for x in doc[max(0, entity.start - CONTEXT_LENGTH / 2):entity.start]], True, padding) + \
pad_list(CONTEXT_LENGTH / 2, [x for x in doc[entity.end: entity.end + CONTEXT_LENGTH / 2]], False, padding)
far_inp = pad_list(CONTEXT_LENGTH / 2, [x for x in doc[max(0, entity.start - CONTEXT_LENGTH):max(0, entity.start - CONTEXT_LENGTH / 2)]], True, padding) + \
pad_list(CONTEXT_LENGTH / 2, [x for x in doc[entity.end + CONTEXT_LENGTH / 2: entity.end + CONTEXT_LENGTH]], False, padding)
map_vector = text2mapvec(doc=near_inp + far_inp,
mapping=ENCODING_MAP_1x1,
outliers=OUTLIERS_MAP_1x1,
polygon_size=1,
db=conn,
exclude=name)
context_words, entities_strings = [], []
target_string = pad_list(TARGET_LENGTH,
[x.text.lower() for x in entity], True,
u'0')
target_string = [
word_to_index[x]
if x in word_to_index else word_to_index[UNKNOWN]
for x in target_string
]
for words in [near_inp, far_inp]:
for word in words:
if word.text.lower() in word_to_index:
vec = word_to_index[word.text.lower()]
else:
vec = word_to_index[UNKNOWN]
if word.ent_type_ in [
u"GPE", u"FACILITY", u"LOC", u"FAC", u"LOCATION"
]:
entities_strings.append(vec)
context_words.append(word_to_index[u'0'])
elif word.is_alpha and not word.is_stop:
context_words.append(vec)
entities_strings.append(word_to_index[u'0'])
else:
context_words.append(word_to_index[u'0'])
entities_strings.append(word_to_index[u'0'])
prediction = model.predict([
np.array([context_words]),
np.array([context_words]),
np.array([entities_strings]),
np.array([entities_strings]),
np.array([map_vector]),
np.array([target_string])
])
prediction = index_to_coord(
REVERSE_MAP_2x2[np.argmax(prediction[0])], 2)
candidates = get_coordinates(conn, name)
if len(candidates) == 0:
# print(u"Don't have an entry for", name, u"in GeoNames")
continue
max_pop = candidates[0][2]
best_candidate = []
bias = 0.905 # Tweak the parameter depending on the domain you're working with.
# Less than 0.9 suitable for ambiguous text, more than 0.9 suitable for less ambiguous locations, see paper
for candidate in candidates:
err = great_circle(
prediction, (float(candidate[0]), float(candidate[1]))).km
best_candidate.append(
(err -
(err * max(1, candidate[2]) / max(1, max_pop)) * bias,
(float(candidate[0]), float(candidate[1]))))
best_candidate = sorted(best_candidate, key=lambda (a, b): a)[0]
# England,, England,, 51.5,, -0.11,, 669,, 676 || - use evaluation script to test correctness
# print name, start, end
# print u"Coordinates:", best_candidate[1]
name = name.encode('utf-8')
one_toponym = "{0},,{1},,{2},,{3},,{4},,{5}||".format(
name, name, best_candidate[1][0], best_candidate[1][1], start,
end)
result_string = result_string + one_toponym
return result_string
# ------------------------------------- END OF TRAINING, BEGINNING OF TESTING -----------------------------------
X = []
final_errors = []
clf = joblib.load(u'../data/bayes.pkl')
test_file = u"data/eval_" + data + u".txt" # which data to test on?
for (x, y) in generate_arrays_from_file_map2vec(test_file, looping=False):
X.extend(x[0]) # Load test instances
print(u"Testing with:", len(X), u"examples.")
conn = sqlite3.connect(u'../data/geonames.db')
for x, (y, name, context) in zip(clf.predict(X), generate_strings_from_file(test_file)):
p = index_to_coord(REVERSE_MAP_2x2[x], 2)
candidates = get_coordinates(conn.cursor(), name)
if len(candidates) == 0:
print(u"Don't have an entry for", name, u"in GeoNames")
raise Exception(u"Check your database, buddy :-)")
# candidates = [candidates[0]] # Uncomment for population heuristic.
# THE ABOVE IS THE POPULATION ONLY BASELINE IMPLEMENTATION
best_candidate = []
max_pop = candidates[0][2]
bias = 0.9 # bias parameter, see
for candidate in candidates:
err = great_circle(p, (float(candidate[0]), float(candidate[1]))).km
best_candidate.append((err - (err * max(1, candidate[2]) / max(1, max_pop)) * bias, (float(candidate[0]), float(candidate[1]))))