Interpret CAR correctly

requirements.txt

@@ -0,0 +1 @@
+trec_car

trec-cast-2019-multi-turn.py

@@ -17,6 +17,7 @@ from collections import defaultdict
 
 import datasets
 import csv
+from trec_car import read_data
 
 # Find for instance the citation on arxiv or on the dataset repo/website
 _CITATION = """\
@@ -104,7 +105,6 @@ class TrecCast2019MultiTurn(datasets.GeneratorBasedBuilder):
             })
             download_size = 8010
         else: # for self.config.name == 'test_collection':
-
             features = datasets.Features({
                 "docno": datasets.Value("string"),
                 "text": datasets.Value("string"),
@@ -182,18 +182,17 @@ class TrecCast2019MultiTurn(datasets.GeneratorBasedBuilder):
                     qid = f"{conversation_id}_{str(idx+1)}"
                     yield qid, ({'query': query, 'history': queries[:idx], 'qid': qid})
         elif split == 'test_collection':
-            with open(file['msmarco']+"/collection.tsv") as f:
+            car_file = file['car'] + "/paragraphCorpus/dedup.articles-paragraphs.cbor"
+            msmarco_file = file['msmarco']+"/collection.tsv"
+            for para in read_data.iter_paragraphs(open(car_file, 'rb')):
+                docid = f"CAR_{para.para_id}"
+                yield docid, ({"docno": docid, "text": para.get_text()})
+
+            with open(msmarco_file) as f:
                 msmarco = csv.reader(f, delimiter="\t")
                 for line in msmarco:
                     docid, text = line
                     docid = f"MARCO_{docid}"
                     yield docid, ({"docno": docid, "text": text})
-
-            with open(file['car']+"/collection.tsv", 'rb') as f:
-                car = csv.reader(f, delimiter="\t")
-                for line in car:
-                    docid, text = line.split(" ")
-                    docid = f"CAR_{docid}"
-                    yield docid, ({"docno": docid, "text": text})
         else:
             raise NotImplementedError(f"'{split}' is not yet implemented")