Coreference Resolution for WikiQA dataset

.gitignore

@@ -1,3 +1,5 @@
 *.json
 */__pycache__/*
 *.pyc
+*.tsv
+*.csv

DataSets.md

@@ -8,6 +8,8 @@ We are planning to use the following datasets to train the models.
 
 ## Question Answering
 
+[WikiQA (Wikipedia Open-Domain Question Answering](https://paperswithcode.com/dataset/wikiqa)
+
 ## Reasoning
 
 [Avicenna: Syllogistic Commonsense Reasoning](https://github.com/ZeinabAghahadi/Syllogistic-Commonsense-Reasoning)

qarac/utils/CoreferenceResolver.py

@@ -0,0 +1,64 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Created on Mon Sep 11 09:46:51 2023
+
+@author: peter
+"""
+
+from allennlp.predictors.predictor import Predictor
+import pandas
+
+def clean(sentence):
+    return sentence if sentence.strip().endswith('.') else sentence+'.'
+
+class CoreferenceResolver(object):
+    
+    def __init__(self):
+        model_url = "https://storage.googleapis.com/allennlp-public-models/coref-spanbert-large-2020.02.27.tar.gz"
+        self.predictor = Predictor.from_path(model_url)
+        
+    def __call__(self,group):
+        tokenized = group.apply(clean).str.split()
+        line_breaks = tokenized.apply(len).cumsum()
+        doc = []
+        for line in tokenized:
+            doc.extend(line)
+        clusters = self.predictor.predict_tokenized(doc)
+        resolutions = {}
+        for cluster in clusters['clusters']:
+            starts = []
+            longest = -1
+            canonical = None
+            for [start_pos,end_pos] in cluster:
+                resolutions[start_pos]={'end':end_pos+1}
+                starts.append(start_pos)
+                length = end_pos - start_pos
+                if length > longest:
+                    longest = length
+                    canonical = doc[start_pos:end_pos+1]
+            for start in starts:
+                resolutions[start]['canonical']=canonical
+        doc_pos = 0
+        line = 0
+        results = []
+        current = []
+        while doc_pos < len(doc):
+            if doc_pos in resolutions:
+                current.extend(resolutions[doc_pos]['canonical'])
+                doc_pos=resolutions[doc_pos]['end']
+            else:
+                current.append(doc[doc_pos])
+                doc_pos+=1
+            if doc_pos>=line_breaks.iloc[line]:
+                results.append(' '.join(current))
+                line+=1
+                current = []
+        return pandas.Series(results,
+                             index=group.index)
+        
+            
+            
+                
+        
+        

requirements.txt

@@ -8,3 +8,5 @@ transformers
 spacy
 spacy-experimental
 pandas
+allennlp
+allennlp-models

scripts.py

@@ -1,5 +1,6 @@
 
 import os
+import re
 import argparse
 import pickle
 import tokenizers
@@ -9,8 +10,10 @@ import qarac.models.qarac_base_model
 import keras
 import tensorflow
 import spacy
-import spacy_experimental
 import pandas
+import qarac.utils.CoreferenceResolver
+
+
 
 def decoder_loss(y_true,y_pred):
     return keras.losses.sparse_categorical_crossentropy(y_true,
@@ -28,13 +31,17 @@ def clean_question(doc):
 
 def prepare_wiki_qa(filename,outfilename):
     data = pandas.read_csv(filename,sep='\t')
+    data['QNum']=data['QuestionID'].apply(lambda x: int(x[1:]))
     nlp = spacy.load('en_core_web_trf')
-    nlp.add_pipe('experimental_coref')
-    data['Resolved_answer'] = pandas.Series([sent.text 
-                                              for doc in nlp.pipe(data.groupby('DocumentID')['Sentence'].apply(lambda x: ' '.join(x)))
-                                              for sent in doc.sentences])
-    data['Cleaned_questions']=pandas.Series([clean_question(doc) for doc in nlp.pipe(data)])
-    data[['Cleaned_questions','Resolved_answers','Label']].to_csv(outfilename)
+    predictor = qarac.utils.CoreferenceResolver.CoreferenceResolver()
+    data['Resolved_answer'] = data.groupby('QNum')['Sentence'].transform(predictor)
+    unique_questions = data.groupby('QNum')['Question'].first()
+    cleaned_questions = pandas.Series([clean_question(doc)
+                                       for doc in nlp.pipe(unique_questions)],
+                                      index = unique_questions.index)
+    for (i,question) in cleaned_questions.items():
+        data.loc[data['QNum']==i,'Cleaned_question']=question
+    data[['Cleaned_question','Resolved_answer','Label']].to_csv(outfilename)
 
         
 def train_base_model(task,filename):
@@ -71,7 +78,10 @@ if __name__ == '__main__':
     parser.add_argument('task')
     parser.add_argument('-f','--filename')
     parser.add_argument('-t','--training-task')
+    parser.add_argument('-o','--outputfile')
     args = parser.parse_args()
     if args.task == 'train_base_model':
         train_base_model(args.training_task,args.filename)
+    elif args.task == 'prepare_wiki_qa':
+        prepare_wiki_qa(args.filename,args.outputfile)