Upload ukrainian-treebank-lm.py

ukrainian-treebank-lm.py

@@ -13,7 +13,6 @@
 # limitations under the License.
 """Ukrainian Treebank (Language Modeling) - dataset by Universal Dependencies preprocessed for language modeling"""
 
-
 import itertools
 import operator
 import conllu
@@ -55,22 +54,18 @@ _URLS = {
 class UkrainianTreebankLMConfig(datasets.BuilderConfig):
     """BuilderConfig for UkrainianTreebankLM"""
 
-    def __init__(self, *args, split_by: str = 'paragraph', split: str = 'all', **kwargs):
+    def __init__(self, *args, split_by: str = 'paragraph', **kwargs):
         """BuilderConfig for UkrainianTreebankLM.
         Args:
           *args: keyword arguments forwarded to super.
           split_by: one of 'document', 'paragraph', 'sentence'
-          split: one of 'all', 'train', 'dev', 'test'
           **kwargs: keyword arguments forwarded to super.
         """
         super(UkrainianTreebankLMConfig, self).__init__(*args, **kwargs)
         assert split_by in ['document', 'paragraph', 'sentence'], \
             "split_by should be one of 'document', 'paragraph', 'sentence'"
-        assert split in ['all', 'train', 'dev', 'test'], \
-            "split should be one of 'all', 'train', 'dev', 'test'"
 
         self.split_by = split_by
-        self.split = split
 
 
 class UkrainianTreebankLM(datasets.GeneratorBasedBuilder):
@@ -80,16 +75,14 @@ class UkrainianTreebankLM(datasets.GeneratorBasedBuilder):
     BUILDER_CONFIG_CLASS = UkrainianTreebankLMConfig
     BUILDER_CONFIGS = [
         UkrainianTreebankLMConfig(
-            name=split + '_' + split_by,
-            description=f"{split.capitalize()} {'parts' if split == 'all' else 'part'} split by {split_by}",
+            name=split_by,
+            description=f"Ukrainian Treebank split by {split_by}",
             split_by=split_by,
-            split=split
         )
-        for split, split_by in itertools.product(['all', 'train', 'dev', 'test'],
-                                                 ['document', 'paragraph', 'sentence'])
+        for split_by in ['document', 'paragraph', 'sentence']
     ]
 
-    DEFAULT_CONFIG_NAME = "all_paragraph"
+    DEFAULT_CONFIG_NAME = "paragraph"
 
     def _info(self):
         if self.config.split_by == "document":
@@ -131,51 +124,66 @@ class UkrainianTreebankLM(datasets.GeneratorBasedBuilder):
         )
 
     def _split_generators(self, dl_manager):
-        if self.config.split == "all":
-            urls = list(_URLS.values())
-        else:
-            urls = [_URLS[self.config.split]]
-        filepaths = dl_manager.download_and_extract(urls)
+        filepaths = dl_manager.download_and_extract(_URLS)
         return [
             datasets.SplitGenerator(
-                name=split_name,
-                gen_kwargs={"filepaths": filepaths},
-            )
-            for split_name in [datasets.Split.TRAIN,
-                               datasets.Split.VALIDATION,
-                               datasets.Split.TEST,
-                               datasets.Split.ALL]
+                name=datasets.Split.TRAIN,
+                gen_kwargs={
+                    "filepath": filepaths['train'],
+                    "split": "train"
+                },
+            ),
+            datasets.SplitGenerator(
+                name=datasets.Split.VALIDATION,
+                gen_kwargs={
+                    "filepath": filepaths['dev'],
+                    "split": "dev"
+                },
+            ),
+            datasets.SplitGenerator(
+                name=datasets.Split.TEST,
+                gen_kwargs={
+                    "filepath": filepaths['test'],
+                    "split": "test"
+                },
+            ),
         ]
 
     def _parse(self, filepath):
         with open(filepath, 'r') as f:
-            document_id = None
-            document_title = None
-            paragraph_id = None
-            for sentence in conllu.parse_incr(f):
-                document_id = sentence.metadata.get("newdoc id", document_id)
-                document_title = sentence.metadata.get("doc title", document_title)
-                paragraph_id = sentence.metadata.get("newpar id", paragraph_id)
-
-                yield {
-                    "text": sentence.metadata["text"],
-                    "document_id": document_id,
-                    "document_title": document_title,
-                    "paragraph_id": paragraph_id,
-                    "sentence_id": sentence.metadata["sent_id"]  # sentence id is always present
-                }
-
-    def _parse_filepaths(self, filepaths):
-        for filepath in filepaths:
-            yield from self._parse(filepath)
+            yield from conllu.parse_incr(f, metadata_parsers={"annotation_gap": lambda key, value: (key, value)})
+
+    def _sentence_iterator(self, filepath, split):
+        document_id = None
+        document_title = None
+        paragraph_id = None
+        gap_number = 0
+        for sentence in self._parse(filepath):
+            # TODO how do we handle annotation gaps?
+            # if "annotation_gap" in sentence.metadata:
+            #     document_id = f"unknown_{gap_number}"
+            #     document_title = f"unknown_{gap_number}"
+            #     paragraph_id = f"unknown_{gap_number}"
+            #     gap_number += 1
+
+            document_id = sentence.metadata.get("newdoc id", document_id)
+            document_title = sentence.metadata.get("doc_title", document_title)
+            paragraph_id = sentence.metadata.get("newpar id", paragraph_id)
+
+            yield {
+                "text": sentence.metadata["text"],
+                "document_id": document_id + '_' + split,
+                "document_title": document_title,
+                "paragraph_id": paragraph_id + '_' + split,
+                "sentence_id": sentence.metadata["sent_id"]  # sentence id is always present
+            }
 
     def _generate_examples_sentence(self, sentence_iterator):
         for sentence in sentence_iterator:
             yield sentence["sentence_id"], sentence
 
     def _generate_examples_paragraph(self, sentence_iterator):
-        sorted_sentences = sorted(sentence_iterator, key=operator.itemgetter("sentence_id"))
-        for key, group in itertools.groupby(sorted_sentences,
+        for key, group in itertools.groupby(sentence_iterator,
                                             operator.itemgetter("paragraph_id")):
             try:
                 sentence = next(group)
@@ -192,8 +200,7 @@ class UkrainianTreebankLM(datasets.GeneratorBasedBuilder):
             }
 
     def _generate_examples_document(self, sentence_iterator):
-        sorted_sentences = sorted(sentence_iterator, key=operator.itemgetter("sentence_id"))
-        for key, group in itertools.groupby(sorted_sentences,
+        for key, group in itertools.groupby(sentence_iterator,
                                             operator.itemgetter("document_id")):
             try:
                 sentence = next(group)
@@ -208,8 +215,8 @@ class UkrainianTreebankLM(datasets.GeneratorBasedBuilder):
                 "document_title": sentence["document_title"],
             }
 
-    def _generate_examples(self, filepaths):
-        sentence_iterator = self._parse_filepaths(filepaths)
+    def _generate_examples(self, filepath, split):
+        sentence_iterator = self._sentence_iterator(filepath, split)
 
         if self.config.split_by == "document":
             yield from self._generate_examples_document(sentence_iterator)