Kenlm model

kenlm.sh

@@ -0,0 +1,5 @@
+#!/bin/bash
+
+KENLM_BUILD_PATH='/home/zary/Desktop/kenlm/build'
+$KENLM_BUILD_PATH/bin/lmplz -o 3 < input_train.txt > model.arpa
+$KENLM_BUILD_PATH/bin/build_binary   model.arpa model.binary

run.py

@@ -1,68 +1,74 @@
 import pandas as pd
 import csv
-from collections import Counter, defaultdict
 from nltk.tokenize import RegexpTokenizer
+from english_words import english_words_set
 from nltk import trigrams
+import os
+import kenlm
+from math import log10
 
 
 class WordGapPrediction:
 
     def __init__(self):
         self.tokenizer = RegexpTokenizer(r"\w+")
-        self.model = defaultdict(lambda: defaultdict(lambda: 0))
+        self.model = None
         self.vocab = set()
-        self.alpha = 0.001
+        self.alpha = 0.6
 
     def read_train_data(self, file):
         data = pd.read_csv(file, sep="\t", error_bad_lines=False, index_col=0, header=None)
-        for index, row in data[:100000].iterrows():
-            text = str(row[6]) + ' ' + str(row[7])
-            tokens = self.tokenizer.tokenize(text)
-            for w1, w2, w3 in trigrams(tokens, pad_right=True, pad_left=True):
-                if w1 and w2 and w3:
-                    self.model[(w2, w3)][w1] += 1
-                    self.model[(w1, w2)][w3] += 1
-                    self.vocab.add(w1)
-                    self.vocab.add(w2)
-                    self.vocab.add(w3)
+        with open('input_train.txt', 'w') as f:
+            for index, row in data[:500000].iterrows():
+                first_part = str(row[6])
+                sec_part = str(row[7])
+                if first_part != 'nan':
+                    f.write(first_part + '\n')
+                if sec_part != 'nan':
+                    f.write(sec_part + '\n')
+        os.system('sh ./kenlm.sh')
+        self.model = kenlm.Model("model.binary")
         
-        for word_pair in self.model:
-            num_n_grams = float(sum(self.model[word_pair].values()))
-            for word in self.model[word_pair]:
-                self.model[word_pair][word] = (self.model[word_pair][word] + self.alpha) / (num_n_grams + self.alpha*len(self.vocab))
-
     def generate_outputs(self, input_file, output_file):
         data = pd.read_csv(input_file, sep='\t', error_bad_lines=False, index_col=0, header=None, quoting=csv.QUOTE_NONE)
         with open(output_file, 'w') as f:
             for index, row in data.iterrows():
-                text = str(row[7])
-                tokens = self.tokenizer.tokenize(text)
-                if len(tokens) < 4:
+                first_context = row[6]
+                sec_context = row[7]
+                first_context_tokens = self.tokenizer.tokenize(first_context)
+                sec_context_tokens = self.tokenizer.tokenize(sec_context)
+                if len(first_context_tokens) + len(sec_context_tokens) < 4:
                     prediction = 'the:0.2 be:0.2 to:0.2 of:0.1 and:0.1 a:0.1 :0.1'
                 else:
-                    prediction = word_gap_prediction.predict_probs(tokens[0], tokens[1])
+                    prediction = word_gap_prediction.predict_probs(first_context_tokens[-1], sec_context_tokens[0])
                 f.write(prediction + '\n')
 
     def predict_probs(self, word1, word2):
-        predictions = dict(self.model[word1, word2])
-        most_common = dict(Counter(predictions).most_common(6))
-        
-        total_prob = 0.0
-        str_prediction = ''
 
-        for word, prob in most_common.items():
-            total_prob += prob
-            str_prediction += f'{word}:{prob} '
+        predictions = []
+        for word in english_words_set:
+            sentence = word1 + ' ' + word + ' ' + word2
+            text_score = self.model.score(sentence, bos=False, eos=False)
 
-        if total_prob == 0.0:
-            return 'the:0.2 be:0.2 to:0.2 of:0.1 and:0.1 a:0.1 :0.1'
-
-        if 1 - total_prob >= 0.01:
-            str_prediction += f":{1-total_prob}"
-        else:
-            str_prediction += f":0.01"
-        
-        return str_prediction
+            if len(predictions) < 12:
+                predictions.append((word, text_score))
+            else:
+                worst_score = None
+                for score in predictions:
+                    if not worst_score:
+                        worst_score = score
+                    else:
+                        if worst_score[1] > score[1]:
+                            worst_score = score
+                if worst_score[1] < text_score:
+                    predictions.remove(worst_score)
+                    predictions.append((word, text_score))
+        probs = sorted(predictions, key=lambda tup: tup[1], reverse=True)
+        pred_str = ''
+        for word, prob in probs:
+            pred_str += f'{word}:{prob} '
+        pred_str += f':{log10(0.99)}'
+        return pred_str
 
 word_gap_prediction = WordGapPrediction()
 word_gap_prediction.read_train_data('./train/in.tsv')