Add first version of the ssj500k script

dataset_infos.json

@@ -0,0 +1 @@
+{"all_data": {"description": "The ssj500k training corpus contains about 500 000 tokens manually annotated on the levels of tokenisation,\nsentence segmentation, morphosyntactic tagging, and lemmatisation. About half of the corpus is also manually annotated \nwith syntactic dependencies, named entities, and verbal multiword expressions. About a quarter of the corpus is also \nannotated with semantic role labels. The morphosyntactic tags and syntactic dependencies are included both in the \nJOS/MULTEXT-East framework, as well as in the framework of Universal Dependencies.\n", "citation": "@InProceedings{krek2020ssj500k,\ntitle = {The ssj500k Training Corpus for Slovene Language Processing},\nauthor={Krek, Simon and Erjavec, Toma\u017e and Dobrovoljc, Kaja and Gantar, Polona and Arhar Holdt, Spela and \u010cibej, Jaka and Brank, Janez},\nbooktitle={Proceedings of the Conference on Language Technologies and Digital Humanities},\nyear={2020},\npages={24-33}\n}\n", "homepage": "http://hdl.handle.net/11356/1434", "license": "Creative Commons - Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)", "features": {"id_doc": {"dtype": "string", "id": null, "_type": "Value"}, "idx_par": {"dtype": "int32", "id": null, "_type": "Value"}, "idx_sent": {"dtype": "int32", "id": null, "_type": "Value"}, "id_words": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "words": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "lemmas": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "msds": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "has_ne_ann": {"dtype": "bool", "id": null, "_type": "Value"}, "has_ud_dep_ann": {"dtype": "bool", "id": null, "_type": "Value"}, "has_jos_dep_ann": {"dtype": "bool", "id": null, "_type": "Value"}, "has_srl_ann": {"dtype": "bool", "id": null, "_type": "Value"}, "has_mwe_ann": {"dtype": "bool", "id": null, "_type": "Value"}, "ne_tags": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ud_dep_head": {"feature": {"dtype": "int32", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ud_dep_rel": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "jos_dep_head": {"feature": {"dtype": "int32", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "jos_dep_rel": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "srl_info": [{"idx_arg": {"dtype": "uint32", "id": null, "_type": "Value"}, "idx_head": {"dtype": "uint32", "id": null, "_type": "Value"}, "role": {"dtype": "string", "id": null, "_type": "Value"}}], "mwe_info": [{"type": {"dtype": "string", "id": null, "_type": "Value"}, "word_indices": {"feature": {"dtype": "uint32", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}]}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "ssj500k", "config_name": "all_data", "version": {"version_str": "2.3.0", "description": null, "major": 2, "minor": 3, "patch": 0}, "splits": {"train": {"name": "train", "num_bytes": 69372407, "num_examples": 27829, "dataset_name": "ssj500k"}}, "download_checksums": {"https://www.clarin.si/repository/xmlui/bitstream/handle/11356/1434/ssj500k-en.TEI.zip": {"num_bytes": 13021836, "checksum": "08ac4d6cf74a45bc81f6e9ca53e7406c96c906c218cbb8ff2f7365e96655c460"}}, "download_size": 13021836, "post_processing_size": null, "dataset_size": 69372407, "size_in_bytes": 82394243}, "named_entity_recognition": {"description": "The ssj500k training corpus contains about 500 000 tokens manually annotated on the levels of tokenisation,\nsentence segmentation, morphosyntactic tagging, and lemmatisation. About half of the corpus is also manually annotated \nwith syntactic dependencies, named entities, and verbal multiword expressions. About a quarter of the corpus is also \nannotated with semantic role labels. The morphosyntactic tags and syntactic dependencies are included both in the \nJOS/MULTEXT-East framework, as well as in the framework of Universal Dependencies.\n", "citation": "@InProceedings{krek2020ssj500k,\ntitle = {The ssj500k Training Corpus for Slovene Language Processing},\nauthor={Krek, Simon and Erjavec, Toma\u017e and Dobrovoljc, Kaja and Gantar, Polona and Arhar Holdt, Spela and \u010cibej, Jaka and Brank, Janez},\nbooktitle={Proceedings of the Conference on Language Technologies and Digital Humanities},\nyear={2020},\npages={24-33}\n}\n", "homepage": "http://hdl.handle.net/11356/1434", "license": "Creative Commons - Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)", "features": {"id_doc": {"dtype": "string", "id": null, "_type": "Value"}, "idx_par": {"dtype": "int32", "id": null, "_type": "Value"}, "idx_sent": {"dtype": "int32", "id": null, "_type": "Value"}, "id_words": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "words": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "lemmas": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "msds": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ne_tags": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "ssj500k", "config_name": "named_entity_recognition", "version": {"version_str": "2.3.0", "description": null, "major": 2, "minor": 3, "patch": 0}, "splits": {"train": {"name": "train", "num_bytes": 17651305, "num_examples": 9489, "dataset_name": "ssj500k"}}, "download_checksums": {"https://www.clarin.si/repository/xmlui/bitstream/handle/11356/1434/ssj500k-en.TEI.zip": {"num_bytes": 13021836, "checksum": "08ac4d6cf74a45bc81f6e9ca53e7406c96c906c218cbb8ff2f7365e96655c460"}}, "download_size": 13021836, "post_processing_size": null, "dataset_size": 17651305, "size_in_bytes": 30673141}, "dependency_parsing_ud": {"description": "The ssj500k training corpus contains about 500 000 tokens manually annotated on the levels of tokenisation,\nsentence segmentation, morphosyntactic tagging, and lemmatisation. About half of the corpus is also manually annotated \nwith syntactic dependencies, named entities, and verbal multiword expressions. About a quarter of the corpus is also \nannotated with semantic role labels. The morphosyntactic tags and syntactic dependencies are included both in the \nJOS/MULTEXT-East framework, as well as in the framework of Universal Dependencies.\n", "citation": "@InProceedings{krek2020ssj500k,\ntitle = {The ssj500k Training Corpus for Slovene Language Processing},\nauthor={Krek, Simon and Erjavec, Toma\u017e and Dobrovoljc, Kaja and Gantar, Polona and Arhar Holdt, Spela and \u010cibej, Jaka and Brank, Janez},\nbooktitle={Proceedings of the Conference on Language Technologies and Digital Humanities},\nyear={2020},\npages={24-33}\n}\n", "homepage": "http://hdl.handle.net/11356/1434", "license": "Creative Commons - Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)", "features": {"id_doc": {"dtype": "string", "id": null, "_type": "Value"}, "idx_par": {"dtype": "int32", "id": null, "_type": "Value"}, "idx_sent": {"dtype": "int32", "id": null, "_type": "Value"}, "id_words": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "words": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "lemmas": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "msds": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ud_dep_head": {"feature": {"dtype": "int32", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "ud_dep_rel": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "ssj500k", "config_name": "dependency_parsing_ud", "version": {"version_str": "2.3.0", "description": null, "major": 2, "minor": 3, "patch": 0}, "splits": {"train": {"name": "train", "num_bytes": 14048597, "num_examples": 8000, "dataset_name": "ssj500k"}}, "download_checksums": {"https://www.clarin.si/repository/xmlui/bitstream/handle/11356/1434/ssj500k-en.TEI.zip": {"num_bytes": 13021836, "checksum": "08ac4d6cf74a45bc81f6e9ca53e7406c96c906c218cbb8ff2f7365e96655c460"}}, "download_size": 13021836, "post_processing_size": null, "dataset_size": 14048597, "size_in_bytes": 27070433}, "dependency_parsing_jos": {"description": "The ssj500k training corpus contains about 500 000 tokens manually annotated on the levels of tokenisation,\nsentence segmentation, morphosyntactic tagging, and lemmatisation. About half of the corpus is also manually annotated \nwith syntactic dependencies, named entities, and verbal multiword expressions. About a quarter of the corpus is also \nannotated with semantic role labels. The morphosyntactic tags and syntactic dependencies are included both in the \nJOS/MULTEXT-East framework, as well as in the framework of Universal Dependencies.\n", "citation": "@InProceedings{krek2020ssj500k,\ntitle = {The ssj500k Training Corpus for Slovene Language Processing},\nauthor={Krek, Simon and Erjavec, Toma\u017e and Dobrovoljc, Kaja and Gantar, Polona and Arhar Holdt, Spela and \u010cibej, Jaka and Brank, Janez},\nbooktitle={Proceedings of the Conference on Language Technologies and Digital Humanities},\nyear={2020},\npages={24-33}\n}\n", "homepage": "http://hdl.handle.net/11356/1434", "license": "Creative Commons - Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)", "features": {"id_doc": {"dtype": "string", "id": null, "_type": "Value"}, "idx_par": {"dtype": "int32", "id": null, "_type": "Value"}, "idx_sent": {"dtype": "int32", "id": null, "_type": "Value"}, "id_words": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "words": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "lemmas": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "msds": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "jos_dep_head": {"feature": {"dtype": "int32", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "jos_dep_rel": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "ssj500k", "config_name": "dependency_parsing_jos", "version": {"version_str": "2.3.0", "description": null, "major": 2, "minor": 3, "patch": 0}, "splits": {"train": {"name": "train", "num_bytes": 23027788, "num_examples": 11411, "dataset_name": "ssj500k"}}, "download_checksums": {"https://www.clarin.si/repository/xmlui/bitstream/handle/11356/1434/ssj500k-en.TEI.zip": {"num_bytes": 13021836, "checksum": "08ac4d6cf74a45bc81f6e9ca53e7406c96c906c218cbb8ff2f7365e96655c460"}}, "download_size": 13021836, "post_processing_size": null, "dataset_size": 23027788, "size_in_bytes": 36049624}, "semantic_role_labeling": {"description": "The ssj500k training corpus contains about 500 000 tokens manually annotated on the levels of tokenisation,\nsentence segmentation, morphosyntactic tagging, and lemmatisation. About half of the corpus is also manually annotated \nwith syntactic dependencies, named entities, and verbal multiword expressions. About a quarter of the corpus is also \nannotated with semantic role labels. The morphosyntactic tags and syntactic dependencies are included both in the \nJOS/MULTEXT-East framework, as well as in the framework of Universal Dependencies.\n", "citation": "@InProceedings{krek2020ssj500k,\ntitle = {The ssj500k Training Corpus for Slovene Language Processing},\nauthor={Krek, Simon and Erjavec, Toma\u017e and Dobrovoljc, Kaja and Gantar, Polona and Arhar Holdt, Spela and \u010cibej, Jaka and Brank, Janez},\nbooktitle={Proceedings of the Conference on Language Technologies and Digital Humanities},\nyear={2020},\npages={24-33}\n}\n", "homepage": "http://hdl.handle.net/11356/1434", "license": "Creative Commons - Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)", "features": {"id_doc": {"dtype": "string", "id": null, "_type": "Value"}, "idx_par": {"dtype": "int32", "id": null, "_type": "Value"}, "idx_sent": {"dtype": "int32", "id": null, "_type": "Value"}, "id_words": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "words": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "lemmas": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "msds": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "srl_info": [{"idx_arg": {"dtype": "uint32", "id": null, "_type": "Value"}, "idx_head": {"dtype": "uint32", "id": null, "_type": "Value"}, "role": {"dtype": "string", "id": null, "_type": "Value"}}]}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "ssj500k", "config_name": "semantic_role_labeling", "version": {"version_str": "2.3.0", "description": null, "major": 2, "minor": 3, "patch": 0}, "splits": {"train": {"name": "train", "num_bytes": 9901320, "num_examples": 5523, "dataset_name": "ssj500k"}}, "download_checksums": {"https://www.clarin.si/repository/xmlui/bitstream/handle/11356/1434/ssj500k-en.TEI.zip": {"num_bytes": 13021836, "checksum": "08ac4d6cf74a45bc81f6e9ca53e7406c96c906c218cbb8ff2f7365e96655c460"}}, "download_size": 13021836, "post_processing_size": null, "dataset_size": 9901320, "size_in_bytes": 22923156}, "multiword_expressions": {"description": "The ssj500k training corpus contains about 500 000 tokens manually annotated on the levels of tokenisation,\nsentence segmentation, morphosyntactic tagging, and lemmatisation. About half of the corpus is also manually annotated \nwith syntactic dependencies, named entities, and verbal multiword expressions. About a quarter of the corpus is also \nannotated with semantic role labels. The morphosyntactic tags and syntactic dependencies are included both in the \nJOS/MULTEXT-East framework, as well as in the framework of Universal Dependencies.\n", "citation": "@InProceedings{krek2020ssj500k,\ntitle = {The ssj500k Training Corpus for Slovene Language Processing},\nauthor={Krek, Simon and Erjavec, Toma\u017e and Dobrovoljc, Kaja and Gantar, Polona and Arhar Holdt, Spela and \u010cibej, Jaka and Brank, Janez},\nbooktitle={Proceedings of the Conference on Language Technologies and Digital Humanities},\nyear={2020},\npages={24-33}\n}\n", "homepage": "http://hdl.handle.net/11356/1434", "license": "Creative Commons - Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)", "features": {"id_doc": {"dtype": "string", "id": null, "_type": "Value"}, "idx_par": {"dtype": "int32", "id": null, "_type": "Value"}, "idx_sent": {"dtype": "int32", "id": null, "_type": "Value"}, "id_words": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "words": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "lemmas": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "msds": {"feature": {"dtype": "string", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}, "mwe_info": [{"type": {"dtype": "string", "id": null, "_type": "Value"}, "word_indices": {"feature": {"dtype": "uint32", "id": null, "_type": "Value"}, "length": -1, "id": null, "_type": "Sequence"}}]}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "ssj500k", "config_name": "multiword_expressions", "version": {"version_str": "2.3.0", "description": null, "major": 2, "minor": 3, "patch": 0}, "splits": {"train": {"name": "train", "num_bytes": 24215008, "num_examples": 13516, "dataset_name": "ssj500k"}}, "download_checksums": {"https://www.clarin.si/repository/xmlui/bitstream/handle/11356/1434/ssj500k-en.TEI.zip": {"num_bytes": 13021836, "checksum": "08ac4d6cf74a45bc81f6e9ca53e7406c96c906c218cbb8ff2f7365e96655c460"}}, "download_size": 13021836, "post_processing_size": null, "dataset_size": 24215008, "size_in_bytes": 37236844}}

ssj500k.py

@@ -0,0 +1,365 @@
+"""ssj500k is a partially annotated training corpus for multiple syntactic and semantic tasks."""
+import re
+import xml.etree.ElementTree as ET
+import os
+
+import datasets
+
+
+_CITATION = """\
+@InProceedings{krek2020ssj500k,
+title = {The ssj500k Training Corpus for Slovene Language Processing},
+author={Krek, Simon and Erjavec, Tomaž and Dobrovoljc, Kaja and Gantar, Polona and Arhar Holdt, Spela and Čibej, Jaka and Brank, Janez},
+booktitle={Proceedings of the Conference on Language Technologies and Digital Humanities},
+year={2020},
+pages={24-33}
+}
+"""
+
+_DESCRIPTION = """\
+The ssj500k training corpus contains about 500 000 tokens manually annotated on the levels of tokenisation,
+sentence segmentation, morphosyntactic tagging, and lemmatisation. About half of the corpus is also manually annotated 
+with syntactic dependencies, named entities, and verbal multiword expressions. About a quarter of the corpus is also 
+annotated with semantic role labels. The morphosyntactic tags and syntactic dependencies are included both in the 
+JOS/MULTEXT-East framework, as well as in the framework of Universal Dependencies.
+"""
+
+_HOMEPAGE = "http://hdl.handle.net/11356/1434"
+
+_LICENSE = "Creative Commons - Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)"
+
+_URLS = {
+    "ssj500k-en.tei": "https://www.clarin.si/repository/xmlui/bitstream/handle/11356/1434/ssj500k-en.TEI.zip"
+}
+
+
+XML_NAMESPACE = "{http://www.w3.org/XML/1998/namespace}"
+IDX_ROOT_WORD = -1
+IDX_NA_HEAD = -2
+NA_TAG = "N/A"
+
+
+def namespace(element):
+    # https://stackoverflow.com/a/12946675
+    m = re.match(r'\{.*\}', element.tag)
+    return m.group(0) if m else ''
+
+
+def word_information(w_or_pc_el):
+    if w_or_pc_el.tag.endswith("pc"):
+        id_word = w_or_pc_el.attrib[f"{XML_NAMESPACE}id"]
+        form = w_or_pc_el.text.strip()
+        lemma = w_or_pc_el.text.strip()
+        msd = w_or_pc_el.attrib[f"msd"]
+    else:  # word - w
+        id_word = w_or_pc_el.attrib[f"{XML_NAMESPACE}id"]
+        form = w_or_pc_el.text.strip()
+        lemma = w_or_pc_el.attrib["lemma"]
+        msd = w_or_pc_el.attrib[f"msd"]
+
+    return id_word, form, lemma, msd
+
+
+class Ssj500k(datasets.GeneratorBasedBuilder):
+    """ssj500k is a partially annotated training corpus for multiple syntactic and semantic tasks."""
+
+    VERSION = datasets.Version("2.3.0")
+
+    BUILDER_CONFIGS = [
+        datasets.BuilderConfig(name="all_data", version=VERSION,
+                               description="The entire dataset with all annotations, in some cases partially missing."),
+        datasets.BuilderConfig(name="named_entity_recognition", version=VERSION,
+                               description="The data subset with annotated named entities."),
+        datasets.BuilderConfig(name="dependency_parsing_ud", version=VERSION,
+                               description="The data subset with annotated dependencies (UD schema)."),
+        datasets.BuilderConfig(name="dependency_parsing_jos", version=VERSION,
+                               description="The data subset with annotated dependencies (JOS schema)."),
+        datasets.BuilderConfig(name="semantic_role_labeling", version=VERSION,
+                               description="The data subset with annotated semantic roles."),
+        datasets.BuilderConfig(name="multiword_expressions", version=VERSION,
+                               description="The data subset with annotated named entities.")
+    ]
+
+    DEFAULT_CONFIG_NAME = "all_data"
+
+    def _info(self):
+        features_dict = {
+            "id_doc": datasets.Value("string"),
+            "idx_par": datasets.Value("int32"),
+            "idx_sent": datasets.Value("int32"),
+            "id_words": datasets.Sequence(datasets.Value("string")),
+            "words": datasets.Sequence(datasets.Value("string")),
+            "lemmas": datasets.Sequence(datasets.Value("string")),
+            "msds": datasets.Sequence(datasets.Value("string"))
+        }
+
+        ret_all_data = self.config.name == "all_data"
+        if ret_all_data:
+            features_dict.update({
+                "has_ne_ann": datasets.Value("bool"), "has_ud_dep_ann": datasets.Value("bool"),
+                "has_jos_dep_ann": datasets.Value("bool"), "has_srl_ann": datasets.Value("bool"),
+                "has_mwe_ann": datasets.Value("bool")
+            })
+
+        if ret_all_data or self.config.name == "named_entity_recognition":
+            features_dict["ne_tags"] = datasets.Sequence(datasets.Value("string"))
+
+        if ret_all_data or self.config.name == "dependency_parsing_ud":
+            features_dict.update({
+                "ud_dep_head": datasets.Sequence(datasets.Value("int32")),
+                "ud_dep_rel": datasets.Sequence(datasets.Value("string"))
+            })
+
+        if ret_all_data or self.config.name == "dependency_parsing_jos":
+            features_dict.update({
+                "jos_dep_head": datasets.Sequence(datasets.Value("int32")),
+                "jos_dep_rel": datasets.Sequence(datasets.Value("string"))
+            })
+
+        if ret_all_data or self.config.name == "semantic_role_labeling":
+            features_dict.update({
+                "srl_info": [{
+                    "idx_arg": datasets.Value("uint32"),
+                    "idx_head": datasets.Value("uint32"),
+                    "role": datasets.Value("string")
+                }]
+            })
+
+        if ret_all_data or self.config.name == "multiword_expressions":
+            features_dict["mwe_info"] = [{
+                "type": datasets.Value("string"),
+                "word_indices": datasets.Sequence(datasets.Value("uint32"))
+            }]
+
+        features = datasets.Features(features_dict)
+        return datasets.DatasetInfo(
+            description=_DESCRIPTION,
+            features=features,
+            homepage=_HOMEPAGE,
+            license=_LICENSE,
+            citation=_CITATION,
+        )
+
+    def _split_generators(self, dl_manager):
+        urls = _URLS["ssj500k-en.tei"]
+        data_dir = dl_manager.download_and_extract(urls)
+        return [
+            datasets.SplitGenerator(
+                name=datasets.Split.TRAIN,
+                gen_kwargs={"file_path": os.path.join(data_dir, "ssj500k-en.TEI", "ssj500k-en.body.xml")}
+            )
+        ]
+
+    def _generate_examples(self, file_path):
+        ret_all_data = self.config.name == "all_data"
+        ret_ne_only = self.config.name == "named_entity_recognition"
+        ret_ud_dep_only = self.config.name == "dependency_parsing_ud"
+        ret_jos_dep_only = self.config.name == "dependency_parsing_jos"
+        ret_srl_only = self.config.name == "semantic_role_labeling"
+        ret_mwe_only = self.config.name == "multiword_expressions"
+
+        curr_doc = ET.parse(file_path)
+        root = curr_doc.getroot()
+        NAMESPACE = namespace(root)
+
+        idx_example = 0
+        for idx_doc, curr_doc in enumerate(root.iterfind(f"{NAMESPACE}div")):
+            id_doc = curr_doc.attrib[f"{XML_NAMESPACE}id"]
+            doc_metadata = {}
+            metadata_el = curr_doc.find(f"{NAMESPACE}bibl")
+            if metadata_el is not None:
+                for child in metadata_el:
+                    if child.tag.endswith("term"):
+                        if child.attrib[f"{XML_NAMESPACE}lang"] != "en":
+                            continue
+
+                        parts = child.text.strip().split(" / ")
+                        attr_name = parts[0]
+                        attr_value = " / ".join(parts[1:])
+
+                    elif child.tag.endswith("note"):
+                        attr_name = child.attrib["type"]
+                        attr_value = child.text.strip()
+                    else:
+                        attr_name = child.tag[len(NAMESPACE):]
+                        attr_value = child.text.strip()
+
+                    doc_metadata[attr_name] = attr_value
+
+            # IMPORTANT: This is a hack, because it is not clear which documents are annotated with NEs
+            # The numbers of annotated docs are obtained from the paper provided in `_CITATION` (Table 1)
+            has_ne = idx_doc < 498
+            has_mwe = idx_doc < 754
+            has_srl = idx_doc < 228
+
+            for idx_par, curr_par in enumerate(curr_doc.iterfind(f"{NAMESPACE}p")):
+                for idx_sent, curr_sent in enumerate(curr_par.iterfind(f"{NAMESPACE}s")):
+                    id2position = {}
+                    id_words, words, lemmas, msds = [], [], [], []
+
+                    # Optional (partial) annotations
+                    named_ents = []
+                    has_ud_dep, ud_dep_heads, ud_dep_rels = False, [], []
+                    has_jos_dep, jos_dep_heads, jos_dep_rels = False, [], []
+                    srl_info = []
+                    mwe_info = []
+
+                    # Note: assuming that all words of a sentence are observed before processing the optional annotations
+                    # i.e., that <w> and <pc> elements come first, then the optional <linkGroup> annotations
+                    for curr_el in curr_sent:
+                        # Words
+                        if curr_el.tag.endswith(("w", "pc")):
+                            id_word, word, lemma, msd = word_information(curr_el)
+
+                            id2position[id_word] = len(id2position)
+                            id_words.append(id_word)
+                            words.append(word)
+                            lemmas.append(lemma)
+                            msds.append(msd)
+                            named_ents.append("O")
+
+                        # Named entities
+                        elif curr_el.tag.endswith("seg"):
+                            has_ne = True
+                            ne_type = curr_el.attrib["subtype"]  # {"per", "loc", "org", "misc", "deriv-per"}
+                            if ne_type.startswith("deriv-"):
+                                ne_type = ne_type[len("deriv-"):]
+                            ne_type = ne_type.upper()
+
+                            num_ne_tokens = 0
+                            for curr_child in curr_el:
+                                num_ne_tokens += 1
+                                id_word, word, lemma, msd = word_information(curr_child)
+
+                                id2position[id_word] = len(id2position)
+                                id_words.append(id_word)
+                                words.append(word)
+                                lemmas.append(lemma)
+                                msds.append(msd)
+
+                            assert num_ne_tokens > 0
+                            nes = [f"B-{ne_type.upper()}"] + [f"I-{ne_type.upper()}" for _ in range(num_ne_tokens - 1)]
+                            named_ents.extend(nes)
+
+                        elif curr_el.tag.endswith("linkGrp"):
+                            # UD dependencies
+                            if curr_el.attrib["type"] == "UD-SYN":
+                                has_ud_dep = True
+                                ud_dep_heads = [None for _ in range(len(words))]
+                                ud_dep_rels = [None for _ in range(len(words))]
+
+                                for link in curr_el:
+                                    dep_rel = link.attrib["ana"].split(":")[-1]
+                                    id_head_word, id_dependant = tuple(map(
+                                        lambda _t_id: _t_id[1:] if _t_id.startswith("#") else _t_id,
+                                        link.attrib["target"].split(" ")
+                                    ))
+
+                                    idx_head_word = id2position[id_head_word] if dep_rel != "root" else IDX_ROOT_WORD
+                                    idx_dep_word = id2position[id_dependant]
+
+                                    ud_dep_heads[idx_dep_word] = idx_head_word
+                                    ud_dep_rels[idx_dep_word] = dep_rel
+
+                            # JOS dependencies
+                            elif curr_el.attrib["type"] == "JOS-SYN":
+                                has_jos_dep = True
+                                jos_dep_heads = [None for _ in range(len(words))]
+                                jos_dep_rels = [None for _ in range(len(words))]
+
+                                for link in curr_el:
+                                    dep_rel = link.attrib["ana"].split(":")[-1]
+                                    id_head_word, id_dependant = tuple(map(
+                                        lambda _t_id: _t_id[1:] if _t_id.startswith("#") else _t_id,
+                                        link.attrib["target"].split(" ")
+                                    ))
+
+                                    idx_head_word = id2position[id_head_word] if dep_rel != "Root" else IDX_ROOT_WORD
+                                    idx_dep_word = id2position[id_dependant]
+
+                                    jos_dep_heads[idx_dep_word] = idx_head_word
+                                    jos_dep_rels[idx_dep_word] = dep_rel
+
+                            # Semantic role labels
+                            elif curr_el.attrib["type"] == "SRL":
+                                for link in curr_el:
+                                    sem_role = link.attrib["ana"].split(":")[-1]
+                                    id_head_word, id_arg_word = tuple(map(
+                                        lambda _t_id: _t_id[1:] if _t_id.startswith("#") else _t_id,
+                                        link.attrib["target"].split(" ")
+                                    ))
+                                    idx_head_word = id2position[id_head_word]
+                                    idx_arg_word = id2position[id_arg_word]
+
+                                    srl_info.append({
+                                        "idx_arg": idx_arg_word,
+                                        "idx_head": idx_head_word,
+                                        "role": sem_role
+                                    })
+
+                            # Multi-word expressions
+                            elif curr_el.attrib["type"] == "MWE":
+                                has_mwe = True
+                                # Follow the KOMET/G-KOMET format, i.e. list of {"type": ..., "word_indices": ...}
+                                for link in curr_el:
+                                    mwe_type = link.attrib["ana"].split(":")[-1]
+                                    involved_words = list(map(
+                                        lambda _t_id: _t_id[1:] if _t_id.startswith("#") else _t_id,
+                                        link.attrib["target"].split(" "))
+                                    )
+                                    word_indices = [id2position[_curr_tok] for _curr_tok in involved_words]
+                                    mwe_info.append({"type": mwe_type, "word_indices": word_indices})
+
+                    # Specified config expects only annotated instances, but there are none for the current instance
+                    if (ret_ne_only and not has_ne) or (ret_ud_dep_only and not has_ud_dep) or \
+                        (ret_jos_dep_only and not has_jos_dep) or (ret_srl_only and not has_srl) or \
+                            (ret_mwe_only and not has_mwe):
+                        continue
+
+                    instance_dict = {
+                        "id_doc": id_doc,
+                        "idx_par": idx_par,
+                        "idx_sent": idx_sent,
+                        "id_words": id_words,
+                        "words": words,
+                        "lemmas": lemmas,
+                        "msds": msds
+                    }
+
+                    if ret_ne_only or ret_all_data:
+                        if not has_ne:
+                            named_ents = [NA_TAG for _ in range(len(words))]
+
+                        instance_dict["ne_tags"] = named_ents
+
+                    if ret_ud_dep_only or ret_all_data:
+                        if not has_ud_dep:
+                            ud_dep_heads = [IDX_NA_HEAD for _ in range(len(words))]
+                            ud_dep_rels = [NA_TAG for _ in range(len(words))]
+
+                        instance_dict["ud_dep_head"] = ud_dep_heads
+                        instance_dict["ud_dep_rel"] = ud_dep_rels
+
+                    if ret_jos_dep_only or ret_all_data:
+                        if not has_jos_dep:
+                            jos_dep_heads = [IDX_NA_HEAD for _ in range(len(words))]
+                            jos_dep_rels = [NA_TAG for _ in range(len(words))]
+
+                        instance_dict["jos_dep_head"] = jos_dep_heads
+                        instance_dict["jos_dep_rel"] = jos_dep_rels
+
+                    if ret_srl_only or ret_all_data:
+                        instance_dict["srl_info"] = srl_info
+
+                    if ret_mwe_only or ret_all_data:
+                        instance_dict["mwe_info"] = mwe_info
+
+                    # When all data is returned, some instances are unannotated or partially annotated, mark instances with flags
+                    if ret_all_data:
+                        instance_dict.update({
+                            "has_ne_ann": has_ne, "has_ud_dep_ann": has_ud_dep, "has_jos_dep_ann": has_jos_dep,
+                            "has_srl_ann": has_srl, "has_mwe_ann": has_mwe
+                        })
+
+                    yield idx_example, instance_dict
+                    idx_example += 1