Add passage generation

wikipedia_snippets_streamed.py

@@ -1,5 +1,5 @@
 """Wikipedia snippets in parquet format"""
-import os
+import math
 import pyarrow.parquet as pq
 import datasets
 
@@ -25,16 +25,20 @@ _LICENSE = (
 
 
 class WikipediaSnippetsStreamed(datasets.GeneratorBasedBuilder):
-    """Bengali wikipedia from 03/20/2021"""
 
     def _info(self):
         return datasets.DatasetInfo(
             description=_DESCRIPTION,
             features=datasets.Features(
                 {
-                    "wikidata_id": datasets.Value("string"),
-                    "text": datasets.Value("string"),
-                    "version_id": datasets.Value("string"),
+                    "wiki_id": datasets.Value("string"),
+                    "start_paragraph": datasets.Value("int32"),
+                    "start_character": datasets.Value("int32"),
+                    "end_paragraph": datasets.Value("int32"),
+                    "end_character": datasets.Value("int32"),
+                    "article_title": datasets.Value("string"),
+                    "section_title": datasets.Value("string"),
+                    "passage_text": datasets.Value("string"),
                 }
             ),
             supervised_keys=None,
@@ -49,10 +53,50 @@ class WikipediaSnippetsStreamed(datasets.GeneratorBasedBuilder):
             datasets.SplitGenerator(name=datasets.Split.TRAIN, gen_kwargs={"filepath": downloaded_file}),
         ]
 
+    def wiki40b_article_snippets(self, article, passage_len=100, overlap=0):
+        paragraphs = article["text"].split("\n")
+        aticle_idx = paragraphs.index("_START_ARTICLE_") + 1
+        article_title = paragraphs[aticle_idx] if aticle_idx < len(paragraphs) else ""
+        section_indices = [i + 1 for i, par in enumerate(paragraphs[:-1]) if par == "_START_SECTION_"]
+        par_tabs = [par.split(" ") for par in paragraphs]
+        word_map = [
+            (i, len(" ".join(par[:j])), w)
+            for i, par in enumerate(par_tabs)
+            if not par[0].startswith("_START_")
+            for j, w in enumerate(par)
+            if i > 0
+        ]
+        step_size = passage_len - overlap
+        passages = []
+        for i in range(math.ceil(len(word_map) / step_size)):
+            pre_toks = word_map[i * step_size: i * step_size + passage_len]
+            start_section_id = max([0] + [j for j in section_indices if j <= pre_toks[0][0]])
+            section_ids = [j for j in section_indices if j >= start_section_id and j <= pre_toks[-1][0]]
+            section_ids = section_ids if len(section_ids) > 0 else [0]
+            passage_text = " ".join([w for p_id, s_id, w in pre_toks])
+            passages += [
+                {
+                    "article_title": article_title,
+                    "section_title": " & ".join([paragraphs[j] for j in section_ids]),
+                    "wiki_id": article["wikidata_id"],
+                    "start_paragraph": pre_toks[0][0],
+                    "start_character": pre_toks[0][1],
+                    "end_paragraph": pre_toks[-1][0],
+                    "end_character": pre_toks[-1][1] + len(pre_toks[-1][2]) + 1,
+                    "passage_text": passage_text.replace("_NEWLINE_", "\n"),
+                }
+            ]
+        return passages
+
     def _generate_examples(self, filepath):
         logger.info("generating examples from = %s", filepath)
+        passage_counter = 0
         with open(filepath, "rb") as f:
             pf = pq.ParquetFile(f)
             for i in range(pf.num_row_groups):
-                id_ = f"{filepath}_{i}"
-                yield id_, pf.read_row_group(i).to_pydict()
+                batch_table_dict = pf.read_row_group(i).to_pydict()
+                for idx, article_text in enumerate(batch_table_dict["text"]):
+                    article = {"text": article_text, "wikidata_id": batch_table_dict["wikidata_id"][idx]}
+                    passages = self.wiki40b_article_snippets(article)
+                    for passage in passages:
+                        yield ++passage_counter, passage