def main():
gold_tl_fname = sys.argv[1]
sys_tl_fname = sys.argv[2]
with open(gold_tl_fname, errors="ignore") as f:
gold_tl = Timeline.from_file(f)
with open(sys_tl_fname, errors="ignore") as f:
sys_tl = Timeline.from_file(f)
all_dates = sorted(set(gold_tl.get_dates()).union(set(sys_tl.get_dates())))
for date in all_dates:
print(date)
if date in gold_tl.dates_to_summaries:
gold_sum = "\n".join(gold_tl.dates_to_summaries[date])
else:
gold_sum = "-----"
if date in sys_tl.dates_to_summaries:
sys_sum = "\n".join(sys_tl.dates_to_summaries[date])
else:
sys_sum = "-----"
print_vertical_split(gold_sum, sys_sum)
print()
def generate_question_html(sys_tl_dir, tl_questions):
for (tl_topic, tl_name), questions in tl_questions.items():
sys_tl_path = os.path.join(sys_tl_dir, tl_topic, tl_name)
with open(sys_tl_path, errors="ignore") as f_tl:
tl = Timeline.from_file(f_tl)
yield tl_topic, tl_name, generate_template(tl, questions)
def read_gold_tl_dir(gold_dir):
timelines = {}
for tl_fname in os.listdir(gold_dir):
tl_path = os.path.join(gold_dir, tl_fname)
with open(tl_path, errors="ignore") as f:
timelines[tl_fname] = Timeline.from_file(f)
return timelines
def load_all_gold_timelines():
all_timelines = {}
gold_tl_dir = "gold-timelines"
for topic_gold_tl_dir in iter_dirs(gold_tl_dir):
for gold_tl_fname in iter_files(topic_gold_tl_dir, ".txt"):
with open(gold_tl_fname, errors="ignore") as f:
tl = Timeline.from_file(f)
all_timelines[os.path.basename(topic_gold_tl_dir),
os.path.basename(gold_tl_fname)] = tl
return all_timelines
def compute_reliability(date_scores):
tl_base_path = Path("./gold-timelines")
all_annotators = sorted(date_scores)
all_topics = set()
for scores in date_scores.values():
for topic, tl_name in scores:
all_topics.add((topic, tl_name))
for topic, tl_name in all_topics:
with open(tl_base_path / topic / (tl_name + ".txt"),
errors="ignore") as f:
tl = Timeline.from_file(f)
score_matrix = np.zeros((len(date_scores), len(tl.get_dates())))
all_dates = sorted(tl.get_dates())
for annotator_idx, annotator in enumerate(all_annotators):
annotator_tl_scores = date_scores[annotator][(topic, tl_name)]
sorted_dates = sorted(
all_dates,
key=lambda date: annotator_tl_scores.get(date, 0),
reverse=True)
new_annotator_tl_scores = {}
curr_idx = 0
prev_score = None
for date in sorted_dates:
score = annotator_tl_scores.get(date, 0)
if prev_score is None or prev_score != score:
curr_idx += 1
prev_score = score
new_annotator_tl_scores[date] = curr_idx
for date_idx, date in enumerate(all_dates):
score_matrix[annotator_idx,
date_idx] = new_annotator_tl_scores.get(
date, 0)
print(topic, tl_name,
k.alpha(score_matrix, level_of_measurement="ordinal"))
for annotator_1_idx, annotator_2_idx in it.combinations(
range(len(all_annotators)), 2):
annotator_1 = ANNOTATORS[all_annotators[annotator_1_idx]]
annotator_2 = ANNOTATORS[all_annotators[annotator_2_idx]]
annotator_rows = score_matrix[[annotator_1_idx, annotator_2_idx]]
print(annotator_1, annotator_2,
k.alpha(annotator_rows, level_of_measurement="interval"))
def main():
with open(sys.argv[2], "w") as f_out:
for topic_dirname in os.listdir(sys.argv[1]):
for tl_fname in os.listdir(os.path.join(sys.argv[1],
topic_dirname)):
with open(os.path.join(sys.argv[1], topic_dirname, tl_fname),
errors="ignore") as f_tl:
tl = Timeline.from_file(f_tl)
f_out.write("== {} {}\n".format(topic_dirname, tl_fname))
for date in sorted(tl.get_dates()):
f_out.write(str(date))
f_out.write("\t\t\n")
def main():
all_timelines = []
gold_tl_dir = "gold-timelines"
for topic_gold_tl_dir in iter_dirs(gold_tl_dir):
for gold_tl_fname in iter_files(topic_gold_tl_dir, ".txt"):
with open(gold_tl_fname, errors="ignore") as f:
tl = Timeline.from_file(f)
print(topic_gold_tl_dir, (gold_tl_fname))
if (os.path.split(topic_gold_tl_dir)[-1],
os.path.split(gold_tl_fname)[-1]) in [
("tl17-bpoil", "bbc.txt"),
("crisis-syria", "bbc.txt"),
("tl17-mj", "bbc.txt"),
("crisis-libya", "xinhua.txt")
]:
all_timelines.append(tl)
print(len(tl))
print("Original TL-Count", len(all_timelines))
all_timelines = list(filter(lambda i: len(i) <= 50, all_timelines))
print("New TL-Count", len(all_timelines))
tl_tuple_counts = [len(tl.get_dates()) * 2 for tl in all_timelines]
print(sum(tl_tuple_counts))
price_per_tuple = STUDENT_PRICE_PER_TUPLE
num_annotations = 3
tl_words = [
tokenize(" ".join(summary_sents)) for tl in all_timelines
for summary_sents in tl.dates_to_summaries.values()
]
tl_reading_times = [len(w) / 200 for w in tl_words]
print("Reading time", sum(tl_reading_times))
print(
"Reading cost",
sum(tl_reading_times) * STUDENT_PRICE_PER_READING_MINUTE *
num_annotations)
tl_tuple_price = sum(tl_tuple_counts) * price_per_tuple * num_annotations
print("Number of Tuples", sum(tl_tuple_counts))
print("Annotation Time", sum(tl_tuple_counts) / 1.5)
print("Annotation Cost", tl_tuple_price)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("timeline")
parser.add_argument("replay_file")
parser.add_argument("--replay", action="store_true", default=False)
args = parser.parse_args()
with open(args.timeline, errors="ignore") as f:
tl = Timeline.from_file(f)
comparisons = play_comparison_game(tl)
all_comparisons = []
if not args.replay:
with open(args.replay_file, "w") as f_out:
for date_1, date_2, result in comparisons:
f_out.write("{}\t{}\t{}\n".format(date_1, date_2, result))
all_comparisons.append((date_1, date_2, result))
else:
with open(args.replay_file) as f:
for line in f:
date_1, date_2, result = line.split()
date_1 = datetime.date(*map(int, date_1.split("-")))
date_2 = datetime.date(*map(int, date_2.split("-")))
result = int(result)
all_comparisons.append((date_1, date_2, result))
scores = {key: 0 for key in tl.get_dates()}
for date_1, date_2, result in all_comparisons:
if result == 0:
scores[date_1] += 0.5
scores[date_2] += 0.5
elif result == -1:
scores[date_1] += 1
elif result == 1:
scores[date_2] += 1
for date, score in sorted(scores.items(), key=lambda x: x[1],
reverse=True):
print(date, "\n", "\n".join(tl.dates_to_summaries[date]))
print()
def retrieve_hit_ranking(hit_task_name):
client = MTurkClient(hit_task_name)
hit_ids = client.list_reviewable_hit_ids_and_annotations()
date_scores = defaultdict(Counter)
for hit_id, info in hit_ids:
results = client.load_answers_for_hit(hit_id)
task_name, hit_info = info.split(":")
tl_topic, tl_name = hit_info.split("__")
for assignment in results:
for worker_id, assignment_answer in assignment:
for key, answer in assignment_answer.items():
_, row_idx, answer_type = key.split("-")
date_scores[tl_topic, tl_name][str_to_tl_date(
answer)] += 1 if answer_type == "best" else -1
if "K" not in hit_task_name:
print(answer, answer_type)
return date_scores
tl_base_path = Path("./gold-timelines")
timelines_and_scores = {}
for (topic, tl_name), results in date_scores.items():
with open(tl_base_path / topic / (tl_name + ".txt"),
errors="ignore") as f:
tl = Timeline.from_file(f)
timelines_and_scores[topic, tl_name] = (results, tl)
#summaries = sorted(tl.dates_to_summaries.items(), key=lambda it: results[it[0]], reverse=True)
#if (info_path / hit_id).is_file():
# pass
return timelines_and_scores
def compute_stats_for_tl(tlfname):
num_splitted_sents = 0
root_counts = []
with open(tlfname, errors="ignore") as f:
timeline = Timeline.from_file(f)
num_sents = 0
for date in timeline:
for sent in timeline[date]:
num_sents += 1
sents = list(nlp(sent).sents)
if len(sents) > 1:
num_splitted_sents += 1
for sent in sents:
num_roots = 0
for tok in sent:
if tok.head == tok:
num_roots += 1
root_counts.append(num_roots)
return num_splitted_sents, root_counts, num_sents
def cross_eval_main():
parser = argparse.ArgumentParser()
parser.add_argument("corpus_def")
parser.add_argument("config")
parser.add_argument("param_file")
args = parser.parse_args()
corpora_and_timelines = []
with open(args.corpus_def) as f:
corpus_defs = json.load(f)
for corpus_def in corpus_defs["corpora"]:
timeline_dir = corpus_def["tl_dir"]
corpus_pickle = corpus_def["corpus_pkl"]
corpus = load_corpus(corpus_pickle)
timelines = []
for tl_fname in iter_files(timeline_dir, ".txt"):
with open(tl_fname, encoding="latin-1") as f:
timeline = Timeline.from_file(f)
timelines.append((os.path.basename(tl_fname), timeline))
corpora_and_timelines.append((corpus, timelines))
with open(args.config) as f:
config = json.load(f)
tl_gen = GloballyClusteredSentenceCompressionTimelineGenerator(config)
parameters = tl_gen.run_scoring_cv_train_mode(corpora_and_timelines)
with open(args.param_file, "wb") as f_out:
pickle.dump(parameters, f_out)
def evaluate_tl_main():
parser = argparse.ArgumentParser()
parser.add_argument("-f",
dest="filter_corpus",
default=False,
action="store_true")
parser.add_argument("-c", dest="constraint", default="sent")
parser.add_argument("-t", dest="timelines", nargs="+")
parser.add_argument("-m",
dest="num_multi_selection_runs",
type=int,
default=None)
parser.add_argument("--queryfile")
parser.add_argument("corpus_pickle")
parser.add_argument("config")
args = parser.parse_args()
if args.constraint == "sent":
use_token_count = False
elif args.constraint == "tok":
use_token_count = True
else:
raise ValueError("Unknown constraint {}".format(args.constraint))
corpus = load_corpus(args.corpus_pickle,
filter_blacklist=args.filter_corpus)
timelines = []
for tl_fname in args.timelines:
with open(tl_fname, errors="ignore") as f:
timeline = Timeline.from_file(f)
timelines.append((os.path.basename(tl_fname), timeline))
#tl_gen = APClusteringTimelineGenerator(True)
with open(args.config) as f:
config = json.load(f)
tl_gen = GloballyClusteredSentenceCompressionTimelineGenerator(config)
corpus_basename = os.path.basename(corpus.name).split(".")[0]
print(corpus_basename)
config_basename = os.path.basename(args.config)
results_basename = config_basename
if args.queryfile:
results_basename += "+queryfilter"
out_timelines_dir = os.path.join("system_timelines",
results_basename + "+" + args.constraint,
corpus_basename)
results_dir = os.path.join("evaluation_results",
results_basename + "+" + args.constraint)
if not os.path.isdir(out_timelines_dir):
os.makedirs(out_timelines_dir)
if not os.path.isdir(results_dir):
os.makedirs(results_dir)
query_words = None
if args.queryfile is not None:
with open(args.queryfile) as f:
query_words = [l.strip() for l in f]
debug_identifier = results_basename + "+" + corpus_basename
if use_token_count:
config["scoring"]["use_length"] = True
if args.num_multi_selection_runs is None:
sys_timelines = tl_gen.generate_timelines(
corpus, [
determine_tl_parameters(tl, use_token_count=use_token_count)
for _, tl in timelines
],
reference_timelines=list(map(lambda x: x[1], timelines)),
query_words=query_words,
debug_identifier=debug_identifier)
write_results_file(os.path.join(results_dir, corpus_basename + ".txt"),
out_timelines_dir, timelines, sys_timelines)
else:
with open("multirun-results+{}.txt".format(config_basename),
"a") as f_out:
print(timelines)
evaluator = rouge.TimelineRougeEvaluator(
measures=["rouge_1", "rouge_2"])
all_run_timelines = tl_gen.generate_timelines(
corpus, [
determine_tl_parameters(tl,
use_token_count=use_token_count)
for _, tl in timelines
],
reference_timelines=list(map(lambda x: x[1], timelines)),
query_words=query_words,
debug_identifier=debug_identifier,
num_selection_runs=args.num_multi_selection_runs)
for sys_timelines in all_run_timelines:
for (timeline_name, gold_timeline), sys_timeline in zip(
timelines, sys_timelines):
reference_timeline = GroundTruth([gold_timeline])
eval_results = evaluator.evaluate_concat(
"TL", sys_timeline, reference_timeline)
eval_results_agree = evaluator.evaluate_agreement(
"TL", sys_timeline, reference_timeline)
eval_results_align = evaluator.evaluate_align_date_content_costs_many_to_one(
"TL", sys_timeline, reference_timeline)
f_out.write(" ".join(
map(str, [
eval_results["rouge_1"]["f_score"],
eval_results["rouge_2"]["f_score"],
eval_results_agree["rouge_1"]["f_score"],
eval_results_agree["rouge_2"]["f_score"],
eval_results_align["rouge_1"]["f_score"],
eval_results_align["rouge_2"]["f_score"]
])))
f_out.write("\n")
f_out.write("--------\n")
f_out.write("========\n")
def main():
parser = argparse.ArgumentParser()
parser.add_argument("timeline")
parser.add_argument("event_info")
parser.add_argument("--tuple-file")
args = parser.parse_args()
tl_path = PurePath(args.timeline)
tl_topic = tl_path.parts[-2]
tl_name = tl_path.parts[-1].split(".")[0]
with open(args.event_info) as f:
all_event_info = json.load(f)
event_info = all_event_info[tl_topic]
with open(args.timeline, errors="ignore") as f:
tl = Timeline.from_file(f)
if args.tuple_file is None:
result_tuples = generate_date_tuples(tl)
else:
with open(args.tuple_file) as f:
result_tuples = f.read().strip().split("\n")[1:]
tuples = [
tuple(
sorted((str_to_tl_date(date),
tl.dates_to_summaries[str_to_tl_date(date)])
for date in tuple_line.split("\t")))
for tuple_line in result_tuples
]
batches = []
curr_idx = 0
while curr_idx < len(tuples):
batches.append(tuples[curr_idx:curr_idx + 20])
curr_idx += 20
print("Uploading", len(tuples), "tuples")
client = MTurkClient("Timeline Importance Annotation VIII (K)")
hit_info_path = Path("./hitinfo")
hit_info_path.mkdir(parents=True, exist_ok=True)
for batch in batches:
hit_ids = client.upload_html([
("importance:" + tl_topic + "__" + tl_name,
generate_template(tl, batch, event_info))
])
hit_id = hit_ids[0]
hit_info_file_path = hit_info_path / hit_id
with open(hit_info_file_path, "w") as f:
f.write(tl_topic)
f.write("\t")
f.write(tl_name)
f.write("\n")
for candidate_tuple in batch:
f.write("\t".join(map(lambda it: str(it[0]), candidate_tuple)))
f.write("\n")
def main():
with open(sys.argv[1]) as f, open("readability-samples-fixed-withcopy.csv",
"w") as f_out:
reader = csv.reader(f)
writer = csv.writer(f_out)
sentences_per_system = defaultdict(list)
corpus_basepath = pathlib.Path("./corpora")
header = next(reader)
writer.writerow(header + ["Copied?"])
lines = sorted(enumerate(reader), key=lambda x: x[1][1], reverse=True)
prev_corpus = None
new_lines = []
for l_idx, line in lines:
did_copy = False
if line[0] != "gold":
corpus_path = corpus_basepath / (line[1] + ".pkl")
print(corpus_path)
corpus = CachedCorpusReader.load_corpus(str(corpus_path))
if prev_corpus != corpus:
corpus_sentences = set(
map(
lambda s: tuple(
s.as_token_attr_sequence("form_lowercase")),
corpus.sentences))
#corpus_sentences = set(map(lambda s: "".join(s.as_token_attr_sequence("form_lowercase")).translate(str.maketrans('', '', string.punctuation)).lower(), corpus.sentences))
#sents = line[-1].split(". ")
#did_copy = False
#for sent in sents:
# if len("".join(sent.split()).translate(str.maketrans('', '', string.punctuation)).lower()) == 0:
# continue
# if "".join(sent.split()).translate(str.maketrans('', '', string.punctuation)).lower() in corpus_sentences:
# did_copy = True
tl_name = line[2]
if "nn" not in line[0] and line[1] == "tl17-mj" and line[
2] == "bbc.txt":
tl_name = "bbc.co.uk.txt"
tl_path = os.path.join("filtered", "system_timelines", line[0],
line[1], tl_name)
print(tl_path)
with open(tl_path, errors="ignore") as f:
tl = Timeline.from_file(f)
summary = tl.dates_to_summaries[datetime.date(
*map(int, line[3].split("-")))]
for summary_line in summary:
summary_line = summary_line.strip()
if tuple(summary_line.lower().split()
) in corpus_sentences:
print("COPY!")
did_copy = True
new_line = list(line)
new_line.append("y" if did_copy else "n")
new_lines.append((l_idx, new_line))
for idx, new_line in sorted(new_lines):
writer.writerow(new_line)
from tilse.data.timelines import Timeline
import sys
if __name__ == "__main__":
timelines = []
for tl_fname in sys.argv[1:]:
with open(tl_fname, encoding="latin-1") as f:
#print(tl_fname)
timelines.append(Timeline.from_file(f))
timeline_dates = [sorted(tl.get_dates()) for tl in timelines]
rows = []
for date_idx in range(5):
date_row = []
sum_row = []
for tl_idx in range(len(timelines)):
tl_date = timeline_dates[tl_idx][date_idx]
tl_sum = timelines[tl_idx][tl_date]
date_row.append("\\textbf{" + str(tl_date) + "}")
sum_row.append(str(" \\newline ".join(tl_sum)))
rows.append(date_row)
rows.append(sum_row)
for row in rows:
print(" & ".join(row), "\\\\\\hline")
args = parser.parse_args()
relevant_systems = set(args.relevant_systems)
all_relevant_timelines = defaultdict(lambda: defaultdict(dict))
for directory in iter_dirs(args.system_tl_dir):
system_name = os.path.basename(directory)
for tl_dir in iter_dirs(directory):
for tlfilename in iter_files(tl_dir, ".txt"):
#print(system_name, relevant_systems)
if system_name in relevant_systems:
with open(tlfilename) as tlfile:
all_relevant_timelines[system_name][os.path.basename(
tl_dir)][os.path.basename(
tlfilename)] = Timeline.from_file(tlfile)
#for directory in iter_dirs(args.human_tl_dir):
# source_name = os.path.basename(directory)
# for tlfilename in iter_files(directory, ".txt"):
# with open(tlfilename, errors='ignore') as tlfile:
# all_relevant_timelines["human"][source_name][os.path.basename(tlfilename)] = Timeline.from_file(tlfile)
vectorized_timelines = vectorize_timelines(all_relevant_timelines)
num_samples_per_tl = 5
all_samples = []
for system, timelines in vectorized_timelines.items():
system_samples = set()
for topic_name, tl_name, timeline in timelines:
def main():
parser = argparse.ArgumentParser()
parser.add_argument("sys_1_results_dir")
parser.add_argument("sys_2_results_dir")
args = parser.parse_args()
results_1 = read_results_dir(args.sys_1_results_dir)
results_2 = read_results_dir(args.sys_2_results_dir)
score_diffs = []
available_sents = []
compression_rates = []
spreads = []
for corpus_name in results_1:
if corpus_name not in results_2:
continue
corpus = load_corpus("corpora/" + corpus_name.rsplit(".")[0] + ".pkl")
for tl_name, result_1 in results_1[corpus_name].items():
result_2 = results_2[corpus_name][tl_name]
with open("gold-timelines/" + corpus_name.split(".")[0] + "/" +
tl_name,
errors="ignore") as f:
print("gold-timelines/" + corpus_name.split(".")[0] + "/" +
tl_name)
gold_tl = Timeline.from_file(f)
total_tl_length = sum(map(len,
gold_tl.dates_to_summaries.values()))
total_corpus_length = len(corpus.sentences)
score_diffs.append(result_1.rouge_2_align.f1 -
result_2.rouge_2_align.f1)
available_sents.append(
len(
corpus.docs_between_dates(min(gold_tl.get_dates()),
max(gold_tl.get_dates()))))
compression_rates.append(1.0 -
(total_tl_length / total_corpus_length))
spreads.append(compute_spread(gold_tl))
print("Sents", scipy.stats.spearmanr(available_sents, score_diffs))
print("Compression", scipy.stats.spearmanr(compression_rates, score_diffs))
print("Spread", scipy.stats.spearmanr(spreads, score_diffs))
plt.axhline(color="b")
plt.scatter(
available_sents,
score_diffs,
c=["r" if score_diff <= 0.0 else "b" for score_diff in score_diffs])
plt.figure()
plt.axhline(color="b")
plt.scatter(
compression_rates,
score_diffs,
c=["r" if score_diff <= 0.0 else "b" for score_diff in score_diffs])
plt.figure()
plt.axhline(color="b")
plt.scatter(
spreads,
score_diffs,
c=["r" if score_diff <= 0.0 else "b" for score_diff in score_diffs])
plt.show()
def main():
base_config_path = sys.argv[1]
corpus_fname = sys.argv[2]
gold_tl_fname = sys.argv[3]
with open(gold_tl_fname, errors="ignore") as f:
gold_tl = Timeline.from_file(f)
with open(base_config_path) as f:
config = json.load(f)
corpus = reader.load_corpus(corpus_fname)
generator = GloballyClusteredSentenceCompressionTimelineGenerator(config)
corpus_promise, cluster_promise, dated_cluster_promise, cluster_candidates = generator.get_promises(corpus)
dated_clusters = dated_cluster_promise.get()
all_svo_tuples_per_date = defaultdict(Counter)
for cluster, date in dated_clusters:
for sentence in cluster:
for pred, subj, obj in sentence.dependency_tree.extract_svo_tuples():
if pred is not None:
pred = pred.lemma.lower()
if obj is not None:
obj = obj.lemma.lower()
if subj is not None:
subj = subj.lemma.lower()
all_svo_tuples_per_date[date][(pred, subj, obj)] += 1
triples_by_frequency = []
for date, counter in sorted(all_svo_tuples_per_date.items()):
for triple, count in counter.items():
triples_by_frequency.append((count, date, triple))
params = determine_tl_parameters(gold_tl)
selected_triples = dict()
for (cnt, date, triple) in sorted(triples_by_frequency, reverse=True, key=lambda x: x[0]):
if date < params.first_date or date > params.last_date:
continue
if date not in selected_triples:
if len(selected_triples) < params.max_date_count:
selected_triples[date] = []
else:
continue
if len(selected_triples[date]) > params.max_date_sent_count:
continue
if triple[0] in ("happen", "say"):
continue
selected_triples[date].append(str(triple))
print(Timeline(selected_triples))
def read_tl(fname):
with open(fname, errors="ignore") as f:
return Timeline.from_file(f)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("-e",
dest="evaluation_results_dir",
default="evaluation_results")
parser.add_argument("-s",
dest="system_timelines_dir",
default="system_timelines")
parser.add_argument("-g",
dest="gold_timelines_dir",
default="gold-timelines")
parser.add_argument("-c", dest="cutoff_constraint", default="none")
parser.add_argument("system_name")
parser.add_argument("system_tl_files", nargs="+")
args = parser.parse_args()
per_corpus_timelines = defaultdict(dict)
for tl_fname in args.system_tl_files:
corpus_name, tl_name = parse_tl_name(tl_fname)
with open(tl_fname) as f:
timeline = Timeline.from_file(f)
per_corpus_timelines[corpus_name][tl_name] = timeline
system_name = args.system_name + "+" + args.cutoff_constraint
system_sys_tl_dir = os.path.join(args.system_timelines_dir, system_name)
system_evaluation_dir = os.path.join(args.evaluation_results_dir,
system_name)
ensure_is_dir(system_sys_tl_dir)
ensure_is_dir(system_evaluation_dir)
for corpus_name, tls in per_corpus_timelines.items():
corpus_sys_tl_dir = os.path.join(system_sys_tl_dir, corpus_name)
ensure_is_dir(corpus_sys_tl_dir)
gold_tl_dir = os.path.join(args.gold_timelines_dir,
corpus_name[:-len(".pkl")])
gold_tls = read_gold_tl_dir(gold_tl_dir)
sys_tls = [it[1] for it in sorted(tls.items())]
gold_tls = [it for it in sorted(gold_tls.items())]
if args.cutoff_constraint != "none":
sys_tls = [
timeline_by_applying_constraints(
sys_tl,
determine_tl_parameters(gold_tl),
constraint_type=args.cutoff_constraint)
for sys_tl, (_, gold_tl) in zip(sys_tls, gold_tls)
]
write_results_file(
os.path.join(system_evaluation_dir,
corpus_name[:-len(".pkl")] + ".txt"),
corpus_sys_tl_dir, gold_tls, sys_tls)
def main_twotasks():
annotations_j = retrieve_hit_ranking(
"Timeline Importance Annotation V (J)")
annotations_k = retrieve_hit_ranking(
"Timeline Importance Annotation VI (K)")
assert list(annotations_j) == list(annotations_k)
tl_base_path = Path("./gold-timelines")
for topic, tl_name in annotations_j:
with open(tl_base_path / topic / (tl_name + ".txt"),
errors="ignore") as f:
tl = Timeline.from_file(f)
all_dates = tl.get_dates()
scores_j = annotations_j[topic, tl_name]
scores_k = annotations_k[topic, tl_name]
for date in all_dates:
if date not in scores_j:
scores_j[date] = 0
if date not in scores_k:
scores_k[date] = 0
obs_j = np.zeros(len(tl.get_dates()))
obs_k = np.zeros(len(tl.get_dates()))
top_5_dates_j = [
i[0] for i in sorted(scores_j.items(), key=lambda i: i[1])[:10]
]
top_5_dates_k = [
i[0] for i in sorted(scores_k.items(), key=lambda i: i[1])[:10]
]
print(top_5_dates_j)
print(top_5_dates_k)
print(
len(set(top_5_dates_j).intersection(set(top_5_dates_k))) /
len(top_5_dates_j))
ranking_j = map(
lambda x: x[0],
sorted(scores_j.items(), key=lambda x: x[1], reverse=True))
ranking_k = map(
lambda x: x[0],
sorted(scores_k.items(), key=lambda x: x[1], reverse=True))
ranks_j = dict(map(lambda x: tuple(reversed(x)), enumerate(ranking_j)))
ranks_k = dict(map(lambda x: tuple(reversed(x)), enumerate(ranking_k)))
print(ranks_j)
for idx, date in enumerate(all_dates):
score_j = scores_j[date]
score_k = scores_k[date]
obs_j[idx] = score_j
obs_k[idx] = score_k
print(ranks_j[date] + 1, ranks_k[date] + 1)
print(score_j, score_k)
print(date)
print(tl.dates_to_summaries[date])
print(
k.alpha(np.stack([obs_j, obs_k]), level_of_measurement="interval"))
print(scipy.stats.kendalltau(obs_j, obs_k))
db_file.write(
"\t".join([
os.path.basename(topic_gold_tl_dir),
os.path.basename(gold_tl_fname),
"detail",
str(date)
])
)
db_file.write("\n")
annotator_file.write("# Ask about a detail of what happened on that day.\n\n")
if __name__ == "__main__":
gold_tl_dir = sys.argv[1]
out_file_prefix = sys.argv[2]
db_file = open(out_file_prefix + ".db", "w")
annotator_file = open(out_file_prefix + ".txt", "w")
for topic_gold_tl_dir in iter_dirs(gold_tl_dir):
for gold_tl_fname in iter_files(topic_gold_tl_dir, ".txt"):
with open(gold_tl_fname, errors="ignore") as f:
tl = Timeline.from_file(f)
#write_entity_questions(topic_gold_tl_dir, gold_tl_fname, tl, db_file, annotator_file)
#write_date_questions(topic_gold_tl_dir, gold_tl_fname, tl, db_file, annotator_file)
write_entity_detail_question(topic_gold_tl_dir, gold_tl_fname, tl, db_file, annotator_file)
db_file.close()
annotator_file.close()
def main():
event_annotations = defaultdict(dict)
current_tl_key = None
system_name = None
if len(sys.argv) == 3:
system_name = sys.argv[2]
with open(sys.argv[1]) as f:
for lidx, line in enumerate(f):
if line.startswith("=="):
_, corpus, tl_name = line.split()
current_tl_key = (corpus, tl_name)
continue
elems = line.split()
if len(elems) == 1:
continue
date, response = elems
year, month, day = map(int, date.split("-"))
event_annotations[current_tl_key][datetime.date(year, month,
day)] = response
event_present_in_first_3_ratios = {}
event_present_in_last_3_ratios = {}
event_present_ratios = {}
for tl_key, annotations in event_annotations.items():
dates_to_consider = None
if system_name:
with open(os.path.join("gold-timelines", tl_key[0], tl_key[1]),
errors="ignore") as f:
gold_tl = Timeline.from_file(f)
with open(os.path.join("system_timelines", system_name, tl_key[0],
tl_key[1]),
errors="ignore") as f:
sys_tl = Timeline.from_file(f)
dates_to_consider = set(gold_tl.get_dates()).intersection(
set(sys_tl.get_dates()))
num_correct_first_3 = 0
num_correct_last_3 = 0
num_correct = 0
num_total = 0
annotations = annotations.items()
for idx, (date, annotation) in enumerate(sorted(annotations)):
if dates_to_consider and date not in dates_to_consider:
continue
if annotation != "-":
#print("==>", annotation)
num_correct += 1
if idx < 3:
num_correct_first_3 += 1
elif idx >= (len(annotations) - 3):
num_correct_last_3 += 1
num_total += 1
if num_total == 0:
event_present_ratios[tl_key] = 0
else:
event_present_ratios[tl_key] = num_correct / num_total
event_present_in_first_3_ratios[tl_key] = num_correct_first_3 / 3
event_present_in_last_3_ratios[tl_key] = num_correct_last_3 / 3
#print(event_present_ratios)
print(
"Last",
sum(event_present_in_last_3_ratios.values()) /
len(event_present_in_last_3_ratios))
print(
"First",
sum(event_present_in_first_3_ratios.values()) /
len(event_present_in_first_3_ratios))
print("Total",
sum(event_present_ratios.values()) / len(event_present_ratios))