Added source files

src/align_corpora.py

@@ -0,0 +1,34 @@
+"""Aligns corpora based on the English text using Murmurhash3"""
+import mmh3
+import numpy as np
+import polars as pl
+
+with open("raw/europarl-v7.da-en.en") as in_file:
+    texts = list(in_file)
+    eng_da = pl.DataFrame(dict(text=texts, da_id=np.arange(len(texts))))
+    eng_da = eng_da.select(
+        pl.col("da_id"), pl.col("text").map_elements(mmh3.hash64).alias("hash")
+    )
+    eng_da = eng_da.select(
+        pl.col("da_id"),
+        # This is necessary as Polars doesn't support 128 bit integers.
+        pl.col("hash").list.first().alias("hash1"),
+        pl.col("hash").list.last().alias("hash2"),
+    )
+
+with open("raw/europarl-v7.sv-en.en") as in_file:
+    texts = list(in_file)
+    eng_sv = pl.DataFrame(dict(text=texts, sv_id=np.arange(len(texts))))
+    eng_sv = eng_sv.select(
+        pl.col("sv_id"), pl.col("text").map_elements(mmh3.hash64).alias("hash")
+    )
+    eng_sv = eng_sv.select(
+        pl.col("sv_id"),
+        # This is necessary as Polars doesn't support 128 bit integers.
+        pl.col("hash").list.first().alias("hash1"),
+        pl.col("hash").list.last().alias("hash2"),
+    )
+
+alignment = eng_sv.join(eng_da, on=["hash1", "hash2"], how="inner")
+alignment = alignment.select(pl.col("da_id", "sv_id"))
+alignment.write_parquet("alignment.parquet")

src/build_corpus.py

@@ -0,0 +1,23 @@
+"""Builds corpus based on the alignment."""
+import polars as pl
+
+with open("raw/europarl-v7.da-en.da") as in_file:
+    danish_texts = pl.Series(list(in_file))
+
+with open("raw/europarl-v7.da-en.en") as in_file:
+    english_texts = pl.Series(list(in_file))
+
+with open("raw/europarl-v7.sv-en.sv") as in_file:
+    swedish_texts = pl.Series(list(in_file))
+
+alignment = pl.read_parquet("filtered_alignment.parquet")
+
+corpus = pl.DataFrame(
+    dict(
+        da=danish_texts.take(alignment["da_id"]),
+        en=english_texts.take(alignment["da_id"]),
+        sv=swedish_texts.take(alignment["sv_id"]),
+    )
+)
+
+corpus.write_parquet("corpus.parquet")

src/filter_alignment.py

@@ -0,0 +1,18 @@
+"""Filters the alignment by removing all ids that occur multiple times."""
+import polars as pl
+
+alignment = pl.read_parquet("alignment.parquet")
+
+# Selecting entries without hash collisions
+good_da_ids = (
+    alignment["da_id"].value_counts().filter(pl.col("counts") == 1)["da_id"]
+)
+good_sv_ids = (
+    alignment["sv_id"].value_counts().filter(pl.col("counts") == 1)["sv_id"]
+)
+
+alignment = alignment.filter(
+    pl.col("da_id").is_in(good_da_ids), pl.col("sv_id").is_in(good_sv_ids)
+)
+
+alignment.write_parquet("filtered_alignment.parquet")

src/push_corpus.py

@@ -0,0 +1,11 @@
+"""Pushes corpus to hub."""
+from functools import partial
+
+import polars as pl
+from datasets import Dataset
+
+corpus = pl.read_parquet("corpus.parquet")
+
+dataset = Dataset.from_generator(partial(corpus.iter_rows, named=True))
+dataset = dataset.train_test_split(test_size=0.2, shuffle=True)
+dataset.push_to_hub("kardosdrur/europarl-scandinavian")