More comments

opusparcus.py

@@ -41,15 +41,16 @@ _HOMEPAGE = ""
 
 _LICENSE = ""
 
-# The HuggingFace dataset library doesn't host the datasets but only point to the original files
-# This can be an arbitrary nested dict/list of URLs (see below in `_split_generators` method)
-
+# The HuggingFace dataset library doesn't host the datasets but only
+# points to the original files. This can be an arbitrary nested
+# dict/list of URLs (see below in `_split_generators` method):
 _URLs = {
     "validation": "validation.jsonl",
     "test": "test.jsonl",
     "validation.full": "validation.jsonl",
     "test.full": "test.jsonl",
-    # NB: the "train" split file is defined dynamically inside the `_split_generators` method
+    # NB: the "train" split file is defined dynamically inside the
+    # `_split_generators` method
 }
 
 _VERSION = datasets.Version("1.0.0", "")
@@ -73,8 +74,22 @@ class OpusparcusConfig(datasets.BuilderConfig):
         self.lang = lang
         self.quality = quality
 
+
+# Languages in Opusparcus: German (de), English (en), Finnish (fi),
+# French (fr), Russian (ru), Swedish (sv):
 LANGS = [ "de", "en", "fi", "fr", "ru", "sv" ]
 
+# The training sets (train splits) come in eight sizes (95 .. 60),
+# where the number indicates the estimated proportion [%] of true
+# paraphrases in the set. The higher the number the smaller (but
+# ideally cleaner) the set. The lower the number, the larger (but
+# noisier) the set is. The smaller sets are included as subsets of
+# larger sets. The special value 100 matches no training data at all,
+# so if you are only interested in validation and test sets, you can
+# use the value 100 in order to save time and space. (The quality
+# value is irrelevant for the validation and test sets, which have
+# been annotated manually, and each example has an annotation score
+# attached to it.)
 QUALITIES = [ 100, 95, 90, 85, 80, 75, 70, 65, 60 ]
     
 class Opusparcus(datasets.GeneratorBasedBuilder):
@@ -95,7 +110,7 @@ class Opusparcus(datasets.GeneratorBasedBuilder):
     ]
     
     # There is no default configuration. User always needs to specify one:
-    #DEFAULT_CONFIG_NAME = None
+    # DEFAULT_CONFIG_NAME = None
 
     def _info(self):
         # This method specifies the datasets.DatasetInfo object which
@@ -139,7 +154,8 @@ class Opusparcus(datasets.GeneratorBasedBuilder):
         # self.config.lang and self.config.quality.
 
         if self.config.lang is None:
-            # This is an error, nothing to do here
+            # This is an error: nothing to do here if no language
+            # has been defined:
             return []
         
         # Select which file of the training data contains the matching data:
@@ -212,7 +228,7 @@ class Opusparcus(datasets.GeneratorBasedBuilder):
 
         # If the desired quality value is 100, no subset of the
         # training set is good enough, and we only produce validation
-        # and test sets, in order to save space and time.
+        # and test sets, in order to save space and time:
             
         if self.config.quality <= 95:
             # In this case there is matching training data, so we produce
@@ -243,7 +259,7 @@ class Opusparcus(datasets.GeneratorBasedBuilder):
         # The `key` is here for legacy reason (tfds) and is not important in itself.
 
         if split == datasets.Split.TRAIN:
-            # Training sets are in compressed bz2 files.
+            # Training sets are in jsonl files that have been compressed using bzip2.
             # They contain a field "quality" missing from the validation and test sets.
             # We also know that this file only contains the desired language,
             # because for the training sets the languages are in separate
@@ -264,10 +280,11 @@ class Opusparcus(datasets.GeneratorBasedBuilder):
                     }
         else:
             # The validation and test sets are in jsonl files.
-            # They contain the fields "lang" and "annot_score" that we filter on.
-            # If we ask for the full sets, we will keep all data entries, also
-            # the sentence pairs that were not considered paraphrases by the
-            # annotators:
+            # They contain the fields "lang" and "annot_score" that we
+            # filter on.  If we ask for the full sets, we will keep
+            # all data entries for the desired language, also the
+            # sentence pairs that were not considered paraphrases by
+            # the annotators:
             keep_all = (split == "validation.full" or split == "test.full")
             with open(filepath, encoding="utf-8") as f:
                 for id_, row in enumerate(f):
@@ -276,8 +293,8 @@ class Opusparcus(datasets.GeneratorBasedBuilder):
                         if keep_all or data["annot_score"] >= 3.0:
                             # for full sets keep all;
                             # for standard test and validation sets, keep only
-                            # the paraphrases (annot_score >= 3.0 means "good
-                            # or mostly good example of paraphrases")
+                            # the actual paraphrases (annot_score >= 3.0 means
+                            # "good or mostly good example of paraphrases")
                             yield id_, {
                                 "lang": data["lang"],
                                 "sent1": data["sent1"],