Matej Klemen
commited on
Commit
•
b3d2e59
1
Parent(s):
c8f8673
Add paragraph-level aggregation
Browse files
README.md
CHANGED
@@ -41,7 +41,12 @@ document available at https://www.clarin.si/repository/xmlui/bitstream/handle/11
|
|
41 |
|
42 |
By default the dataset is provided at **sentence-level** (125867 instances): each instance contains a source (the original) and a target (the corrected) sentence. Note that either the source or the target sentence in an instance may be missing - this usually happens when a source sentence is marked as redundant or when a new sentence is added by the teacher. Additionally, a source or a target sentence may appear in multiple instances - for example, this happens when one sentence gets divided into multiple sentences.
|
43 |
|
44 |
-
There is also an option to aggregate the instances at the **document-level**
|
|
|
|
|
|
|
|
|
|
|
45 |
|
46 |
### Supported Tasks and Leaderboards
|
47 |
|
|
|
41 |
|
42 |
By default the dataset is provided at **sentence-level** (125867 instances): each instance contains a source (the original) and a target (the corrected) sentence. Note that either the source or the target sentence in an instance may be missing - this usually happens when a source sentence is marked as redundant or when a new sentence is added by the teacher. Additionally, a source or a target sentence may appear in multiple instances - for example, this happens when one sentence gets divided into multiple sentences.
|
43 |
|
44 |
+
There is also an option to aggregate the instances at the **document-level** or **paragraph-level**
|
45 |
+
by explicitly providing the correct config:
|
46 |
+
```
|
47 |
+
datasets.load_dataset("cjvt/solar3", "paragraph_level")`
|
48 |
+
datasets.load_dataset("cjvt/solar3", "document_level")`
|
49 |
+
```
|
50 |
|
51 |
### Supported Tasks and Leaderboards
|
52 |
|
solar3.py
CHANGED
@@ -126,6 +126,7 @@ def read_data(data_path):
|
|
126 |
data[id_sent] = {
|
127 |
"id_doc": id_text,
|
128 |
"doc_title": text_title,
|
|
|
129 |
"id_token": ids, "form": forms, "lemma": lemmas, "ana": msds_jos, "msd": msds_ud, "ne_tag": nes, "space_after": spaces_after,
|
130 |
"is_manually_validated": is_manually_validated
|
131 |
}
|
@@ -142,6 +143,8 @@ class Solar3(datasets.GeneratorBasedBuilder):
|
|
142 |
BUILDER_CONFIGS = [
|
143 |
datasets.BuilderConfig(name="sentence_level", version=VERSION,
|
144 |
description="Annotations at sentence-level."),
|
|
|
|
|
145 |
datasets.BuilderConfig(name="document_level", version=VERSION,
|
146 |
description="Annotations at document-level."),
|
147 |
]
|
@@ -225,10 +228,14 @@ class Solar3(datasets.GeneratorBasedBuilder):
|
|
225 |
|
226 |
if len(involved_src_sents) > 0:
|
227 |
src_sent_data = deepcopy(source_data[involved_src_sents[0]])
|
|
|
|
|
228 |
|
229 |
for src_sent_id in involved_src_sents[1:]:
|
230 |
curr_sent_data = source_data[src_sent_id]
|
|
|
231 |
src_sent_data["id_token"].extend(curr_sent_data["id_token"])
|
|
|
232 |
src_sent_data["form"].extend(curr_sent_data["form"])
|
233 |
src_sent_data["lemma"].extend(curr_sent_data["lemma"])
|
234 |
src_sent_data["ana"].extend(curr_sent_data["ana"])
|
@@ -244,10 +251,14 @@ class Solar3(datasets.GeneratorBasedBuilder):
|
|
244 |
|
245 |
if len(involved_tgt_sents) > 0:
|
246 |
tgt_sent_data = deepcopy(target_data[involved_tgt_sents[0]])
|
|
|
|
|
247 |
|
248 |
for tgt_sent_id in involved_tgt_sents[1:]:
|
249 |
curr_sent_data = target_data[tgt_sent_id]
|
|
|
250 |
tgt_sent_data["id_token"].extend(curr_sent_data["id_token"])
|
|
|
251 |
tgt_sent_data["form"].extend(curr_sent_data["form"])
|
252 |
tgt_sent_data["lemma"].extend(curr_sent_data["lemma"])
|
253 |
tgt_sent_data["ana"].extend(curr_sent_data["ana"])
|
@@ -288,6 +299,7 @@ class Solar3(datasets.GeneratorBasedBuilder):
|
|
288 |
"id_doc": id_doc[:-1], # doc ID without the "s" or "t" info
|
289 |
"doc_title": doc_title,
|
290 |
"is_manually_validated": is_manually_validated,
|
|
|
291 |
"id_src_tokens": src_sent_data.get("id_token", []),
|
292 |
"src_tokens": src_sent_data.get("form", []),
|
293 |
"src_ling_annotations": {
|
@@ -297,6 +309,7 @@ class Solar3(datasets.GeneratorBasedBuilder):
|
|
297 |
"ne_tag": src_sent_data.get("ne_tag", []),
|
298 |
"space_after": src_sent_data.get("space_after", [])
|
299 |
},
|
|
|
300 |
"id_tgt_tokens": tgt_sent_data.get("id_token", []),
|
301 |
"tgt_tokens": tgt_sent_data.get("form", []),
|
302 |
"tgt_ling_annotations": {
|
@@ -309,6 +322,94 @@ class Solar3(datasets.GeneratorBasedBuilder):
|
|
309 |
"corrections": corr_data
|
310 |
}
|
311 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
312 |
@staticmethod
|
313 |
def aggregate_docs(sent_level_data):
|
314 |
# NOTE: assuming here that `sent_level_data` is pre-sorted by id_doc, which is done in the raw data
|
@@ -397,9 +498,12 @@ class Solar3(datasets.GeneratorBasedBuilder):
|
|
397 |
sent_level_data = list(Solar3.generate_sentences(source_path, target_path, links_path))
|
398 |
|
399 |
if self.config.name == "sentence_level":
|
400 |
-
# Remove IDs that are only useful for aggregating the document-level data
|
401 |
for i, instance in sent_level_data:
|
402 |
-
yield i, {_k: _v for _k, _v in instance.items() if _k not in {"id_src_tokens", "id_tgt_tokens"
|
|
|
|
|
|
|
403 |
else:
|
404 |
yield from list(Solar3.aggregate_docs(sent_level_data))
|
405 |
|
|
|
126 |
data[id_sent] = {
|
127 |
"id_doc": id_text,
|
128 |
"doc_title": text_title,
|
129 |
+
"idx_par": idx_par,
|
130 |
"id_token": ids, "form": forms, "lemma": lemmas, "ana": msds_jos, "msd": msds_ud, "ne_tag": nes, "space_after": spaces_after,
|
131 |
"is_manually_validated": is_manually_validated
|
132 |
}
|
|
|
143 |
BUILDER_CONFIGS = [
|
144 |
datasets.BuilderConfig(name="sentence_level", version=VERSION,
|
145 |
description="Annotations at sentence-level."),
|
146 |
+
datasets.BuilderConfig(name="paragraph_level", version=VERSION,
|
147 |
+
description="Annotations at paragraph-level."),
|
148 |
datasets.BuilderConfig(name="document_level", version=VERSION,
|
149 |
description="Annotations at document-level."),
|
150 |
]
|
|
|
228 |
|
229 |
if len(involved_src_sents) > 0:
|
230 |
src_sent_data = deepcopy(source_data[involved_src_sents[0]])
|
231 |
+
if not isinstance(src_sent_data["idx_par"], list):
|
232 |
+
src_sent_data["idx_par"] = [src_sent_data["idx_par"]]
|
233 |
|
234 |
for src_sent_id in involved_src_sents[1:]:
|
235 |
curr_sent_data = source_data[src_sent_id]
|
236 |
+
|
237 |
src_sent_data["id_token"].extend(curr_sent_data["id_token"])
|
238 |
+
src_sent_data["idx_par"].append(curr_sent_data["idx_par"])
|
239 |
src_sent_data["form"].extend(curr_sent_data["form"])
|
240 |
src_sent_data["lemma"].extend(curr_sent_data["lemma"])
|
241 |
src_sent_data["ana"].extend(curr_sent_data["ana"])
|
|
|
251 |
|
252 |
if len(involved_tgt_sents) > 0:
|
253 |
tgt_sent_data = deepcopy(target_data[involved_tgt_sents[0]])
|
254 |
+
if not isinstance(tgt_sent_data["idx_par"], list):
|
255 |
+
tgt_sent_data["idx_par"] = [tgt_sent_data["idx_par"]]
|
256 |
|
257 |
for tgt_sent_id in involved_tgt_sents[1:]:
|
258 |
curr_sent_data = target_data[tgt_sent_id]
|
259 |
+
|
260 |
tgt_sent_data["id_token"].extend(curr_sent_data["id_token"])
|
261 |
+
tgt_sent_data["idx_par"].append(curr_sent_data["idx_par"])
|
262 |
tgt_sent_data["form"].extend(curr_sent_data["form"])
|
263 |
tgt_sent_data["lemma"].extend(curr_sent_data["lemma"])
|
264 |
tgt_sent_data["ana"].extend(curr_sent_data["ana"])
|
|
|
299 |
"id_doc": id_doc[:-1], # doc ID without the "s" or "t" info
|
300 |
"doc_title": doc_title,
|
301 |
"is_manually_validated": is_manually_validated,
|
302 |
+
"idx_src_par": src_sent_data.get("idx_par", []),
|
303 |
"id_src_tokens": src_sent_data.get("id_token", []),
|
304 |
"src_tokens": src_sent_data.get("form", []),
|
305 |
"src_ling_annotations": {
|
|
|
309 |
"ne_tag": src_sent_data.get("ne_tag", []),
|
310 |
"space_after": src_sent_data.get("space_after", [])
|
311 |
},
|
312 |
+
"idx_tgt_par": tgt_sent_data.get("idx_par", []),
|
313 |
"id_tgt_tokens": tgt_sent_data.get("id_token", []),
|
314 |
"tgt_tokens": tgt_sent_data.get("form", []),
|
315 |
"tgt_ling_annotations": {
|
|
|
322 |
"corrections": corr_data
|
323 |
}
|
324 |
|
325 |
+
@staticmethod
|
326 |
+
def aggregate_pars(sent_level_data):
|
327 |
+
# TODO: the code is a copypaste of the document aggregation, with an additional groupby - could use a refactor
|
328 |
+
uniq_idx_par = 0
|
329 |
+
for idx_doc, (curr_id, curr_group) in enumerate(groupby(sent_level_data, key=lambda tup: tup[1]["id_doc"])):
|
330 |
+
curr_instances = list(map(lambda tup: tup[1], curr_group)) # remove the redundant index info from datasets
|
331 |
+
|
332 |
+
# Some sentences have no `idx_src_par` because they are added by the teacher (not present in the source)
|
333 |
+
for idx_par, curr_par_group in groupby(
|
334 |
+
curr_instances,
|
335 |
+
key=lambda _inst: _inst["idx_src_par"][0] if len(_inst["idx_src_par"]) > 0 else
|
336 |
+
_inst["idx_tgt_par"][0]
|
337 |
+
):
|
338 |
+
src_tokens, tgt_tokens, mapped_corrections = [], [], []
|
339 |
+
src_ling_anns = {"lemma": [], "ana": [], "msd": [], "ne_tag": [], "space_after": []}
|
340 |
+
tgt_ling_anns = {"lemma": [], "ana": [], "msd": [], "ne_tag": [], "space_after": []}
|
341 |
+
seen_src_tokens, seen_tgt_tokens = {}, {}
|
342 |
+
src_base, tgt_base = 0, 0
|
343 |
+
prev_src_base, prev_tgt_base = 0, 0
|
344 |
+
|
345 |
+
doc_title, is_validated = None, None
|
346 |
+
for curr_inst in curr_par_group:
|
347 |
+
doc_title, is_validated = curr_inst["doc_title"], curr_inst["is_manually_validated"]
|
348 |
+
|
349 |
+
id_src_toks, id_tgt_toks = curr_inst["id_src_tokens"], curr_inst["id_tgt_tokens"]
|
350 |
+
curr_src_toks, curr_tgt_toks = curr_inst["src_tokens"], curr_inst["tgt_tokens"]
|
351 |
+
curr_src_anns, curr_tgt_anns = curr_inst["src_ling_annotations"], curr_inst["tgt_ling_annotations"]
|
352 |
+
curr_corrs = curr_inst["corrections"]
|
353 |
+
|
354 |
+
num_added_src, num_added_tgt = 0, 0
|
355 |
+
for idx_position, (id_tok, tok) in enumerate(zip(id_src_toks, curr_src_toks)):
|
356 |
+
if id_tok not in seen_src_tokens:
|
357 |
+
src_tokens.append(tok)
|
358 |
+
src_ling_anns["lemma"].append(curr_src_anns["lemma"][idx_position])
|
359 |
+
src_ling_anns["ana"].append(curr_src_anns["ana"][idx_position])
|
360 |
+
src_ling_anns["msd"].append(curr_src_anns["msd"][idx_position])
|
361 |
+
src_ling_anns["ne_tag"].append(curr_src_anns["ne_tag"][idx_position])
|
362 |
+
src_ling_anns["space_after"].append(curr_src_anns["space_after"][idx_position])
|
363 |
+
|
364 |
+
seen_src_tokens[id_tok] = tok
|
365 |
+
num_added_src += 1
|
366 |
+
|
367 |
+
for idx_position, (id_tok, tok) in enumerate(zip(id_tgt_toks, curr_tgt_toks)):
|
368 |
+
if id_tok not in seen_tgt_tokens:
|
369 |
+
tgt_tokens.append(tok)
|
370 |
+
tgt_ling_anns["lemma"].append(curr_tgt_anns["lemma"][idx_position])
|
371 |
+
tgt_ling_anns["ana"].append(curr_tgt_anns["ana"][idx_position])
|
372 |
+
tgt_ling_anns["msd"].append(curr_tgt_anns["msd"][idx_position])
|
373 |
+
tgt_ling_anns["ne_tag"].append(curr_tgt_anns["ne_tag"][idx_position])
|
374 |
+
tgt_ling_anns["space_after"].append(curr_tgt_anns["space_after"][idx_position])
|
375 |
+
|
376 |
+
seen_tgt_tokens[id_tok] = tok
|
377 |
+
num_added_tgt += 1
|
378 |
+
|
379 |
+
if num_added_src == 0:
|
380 |
+
src_base, prev_src_base = prev_src_base, src_base
|
381 |
+
|
382 |
+
if num_added_tgt == 0:
|
383 |
+
tgt_base, prev_tgt_base = prev_tgt_base, tgt_base
|
384 |
+
|
385 |
+
for corr in curr_corrs:
|
386 |
+
mapped_corrections.append({
|
387 |
+
"idx_src": list(map(lambda _i: src_base + _i, corr["idx_src"])),
|
388 |
+
"idx_tgt": list(map(lambda _i: tgt_base + _i, corr["idx_tgt"])),
|
389 |
+
"corr_types": corr["corr_types"]
|
390 |
+
})
|
391 |
+
|
392 |
+
src_base += num_added_src
|
393 |
+
tgt_base += num_added_tgt
|
394 |
+
|
395 |
+
if num_added_src == 0:
|
396 |
+
src_base, prev_src_base = prev_src_base, src_base
|
397 |
+
|
398 |
+
if num_added_tgt == 0:
|
399 |
+
tgt_base, prev_tgt_base = prev_tgt_base, tgt_base
|
400 |
+
|
401 |
+
yield uniq_idx_par, {
|
402 |
+
"id_doc": curr_id,
|
403 |
+
"doc_title": doc_title,
|
404 |
+
"is_manually_validated": is_validated,
|
405 |
+
"src_tokens": src_tokens,
|
406 |
+
"src_ling_annotations": src_ling_anns,
|
407 |
+
"tgt_tokens": tgt_tokens,
|
408 |
+
"tgt_ling_annotations": tgt_ling_anns,
|
409 |
+
"corrections": mapped_corrections
|
410 |
+
}
|
411 |
+
uniq_idx_par += 1
|
412 |
+
|
413 |
@staticmethod
|
414 |
def aggregate_docs(sent_level_data):
|
415 |
# NOTE: assuming here that `sent_level_data` is pre-sorted by id_doc, which is done in the raw data
|
|
|
498 |
sent_level_data = list(Solar3.generate_sentences(source_path, target_path, links_path))
|
499 |
|
500 |
if self.config.name == "sentence_level":
|
501 |
+
# Remove IDs and indices that are only useful for aggregating the document-level data
|
502 |
for i, instance in sent_level_data:
|
503 |
+
yield i, {_k: _v for _k, _v in instance.items() if _k not in {"id_src_tokens", "id_tgt_tokens",
|
504 |
+
"idx_src_par", "idx_tgt_par"}}
|
505 |
+
elif self.config.name == "paragraph_level":
|
506 |
+
yield from list(Solar3.aggregate_pars(sent_level_data))
|
507 |
else:
|
508 |
yield from list(Solar3.aggregate_docs(sent_level_data))
|
509 |
|