Bugfixes, answer mapping, comments

lm_eval/tasks/coqa.py

 # REMINDER: this code needs to be rewritten for the new framework. Remove this comment when the code is fully converted.
 
 import json
-import random
-import numpy as np
 from lm_eval.base import Task, rf, mean
 from ..utils import sh
 from itertools import zip_longest
 import transformers.data.metrics.squad_metrics as squad_metrics
+import collections
+import datasets
+import numpy as np
+from lm_eval.base import rf, mean
+from . common import HFTask
+from tqdm import tqdm
+import string, re
 
 class CoQA(Task):
+
     def download(self):
         pass
         # -N only overwrites if the remote file has changed
@@ -28,7 +34,11 @@ class CoQA(Task):
         return False
 
     def training_docs(self):
-        return json.load(open('data/coqa/coqa-train-v1.0.json'))['data']
+        doc_data = json.load(open('data/coqa/coqa-train-v1.0.json'))['data']
+        for doc in doc_data:
+            for answer in doc['answers']:
+                answer['input_text'] = self.get_answer_choice(answer['input_text'])
+        return doc_data
 
     def validation_docs(self):
         return json.load(open('data/coqa/coqa-dev-v1.0.json'))['data']
@@ -40,9 +50,10 @@ class CoQA(Task):
         return "Given a passage and a conversation so far, answer the next question in the conversation."
     
     def doc_to_text(self, doc):
-        # Each "doc" is a story and conversation (Q and A pairs).
+        # Given a passage p, the conversation history {q1, a1, . . . qi−1, ai−1} 
+        # and a question qi, the task is to predict the answer ai
         doc_text = doc["story"] + '\n\n'
-        for (q, a) in zip_longest(doc["questions"], doc["answers"][:-1]):   # omit target answer
+        for (q, a) in zip_longest(doc["questions"], doc["answers"][:-1]):   # omit target answer ai
             question = f"Q: {q['input_text']}" + '\n\n'
             answer = f"A: {a['input_text']}" + '\n\n' if a is not None else "A: "
             doc_text += question + answer
@@ -51,33 +62,43 @@ class CoQA(Task):
         
     @classmethod
     def get_answers(cls, doc, turn_id):
-        # This function returns an answer and valid alternatives.
+        # Returns unique answers and valid alternatives (Some questions in CoQA have multiple valid answers).
         answers = []
         answer_forturn = doc["answers"][turn_id - 1]["input_text"]
         answers.append(answer_forturn)
         
-        additionals = doc.get("additional_answers")
-        if additionals:
-            for key in additionals:
-                additional_answer_for_turn = additionals[key][turn_id - 1]["input_text"]
-                if additional_answer_for_turn.upper() not in map(str.upper, answers):
+        additional_answers = doc.get("additional_answers")
+        if additional_answers:
+            for key in additional_answers:
+                additional_answer_for_turn = additional_answers[key][turn_id - 1]["input_text"]
+                if additional_answer_for_turn.lower() not in map(str.lower, answers):
                     answers.append(additional_answer_for_turn)
         return answers
     
-    def doc_to_target(self, doc, turnid=None):
-        # Default to predict last turn.
-        if turnid is None:
-            turnid = len(doc["questions"])
-        all_answers = self.get_answers(doc, turnid)
-        return all_answers[0]   # ignore alternative answers for now
+    @classmethod
+    def get_answer_choice(self, raw_text):
+        # Function maps answers to CoQA answer categories
+        # ~ 1/5 of the CoQA answers are Yes/No 
+        # ~ 2/3 of the CoQA answers are span-based
+        # (answers overlap with the passage ignoring punctuation and case mismatch)
+        if raw_text == "unknown":
+            return '0'
+        if squad_metrics.normalize_answer(raw_text) == "yes":
+            return '1'
+        if squad_metrics.normalize_answer(raw_text) == "no":
+            return '2'
+        return '3' # Not a yes/no question
 
     @staticmethod
     def compute_scores(gold_list, pred):
+        # tests for exact match and on the normalised answer (compute_exact)
+        # test for overlap (compute_f1)
         f1_sum = 0.0
         em_sum = 0.0
         if len(gold_list) > 1:
             for i in range(len(gold_list)):
                 gold_answers = gold_list[0:i] + gold_list[i + 1:]
+                # predictions compared against (n) golds and take maximum
                 em_sum += max(squad_metrics.compute_exact(a, pred) for a in gold_answers)
                 f1_sum += max(squad_metrics.compute_f1(a, pred) for a in gold_answers)
         else:
@@ -86,6 +107,14 @@ class CoQA(Task):
 
         return {'em': em_sum / max(1, len(gold_list)), 'f1': f1_sum / max(1, len(gold_list))}
 
+    def doc_to_target(self, doc, turnid=None):
+        # Default to prediction of last turn.
+        if turnid is None:
+            turnid = len(doc["questions"])
+
+        raw_text = doc['answers'][turnid - 1]["input_text"]
+        return self.get_answer_choice(raw_text)
+
     def construct_requests(self, doc, ctx):
         """ Uses RequestFactory to construct Requests and returns an iterable of 
         Requests which will be sent to the LM.
@@ -97,11 +126,11 @@ class CoQA(Task):
             language description, as well as the few shot examples, and the question
             part of the document for `doc`. 
         """
-        requests = []
-        for answers in self.get_answers(doc, len(doc["questions"])):
-            for a in answers:
-                requests.append(rf.loglikelihood(ctx, " " + a)) 
-        return requests
+        ll_requests = [
+            rf.loglikelihood(ctx, " " + i)
+            for i in ['0', '1', '2', '3']
+        ]
+        return ll_requests
 
     def process_results(self, doc, results):
         """Take a single document and the LM results and evaluates, returning a 
@@ -113,16 +142,15 @@ class CoQA(Task):
         :param results:
             The results of the requests created in construct_requests.
         """
-
         turn_id = len(doc["questions"])
-        gold_list = self.get_answers(doc, turn_id)
-        pred = np.argmax(results)
+        gold_list = [self.get_answer_choice(r_text) for r_text in self.get_answers(doc, turn_id)]
+        pred = str(np.argmax(results))
 
-        (em, f1) = self.compute_scores(gold_list, pred)
+        scores = self.compute_scores(gold_list, pred)
 
         return {
-            "f1": f1,
-            "em": em,
+            "f1": scores['f1'],
+            "em": scores['em'],
         }
 
     def higher_is_better(self):