def predict_error_analysis(test_data_lst, weight_dict):
results_lst = []
for sent in test_data_lst:
words, gs_label, sentence = extract_features_from_sentence(sent)
selected_features = get_features(words, weight_dict)
classify_label = classify(selected_features, sentence)
if classify_label != gs_label:
results_lst.append((sentence, classify_label, gs_label))
return results_lst
def hinge_loss_bigram(weight_dict, sentence):
words, gs_label, sent = extract_features_from_sentence_bigram(sentence)
selected_features = get_features(words, weight_dict)
classify_label = costClassify_bigram(selected_features, sent, gs_label)
for feature_temp in selected_features:
bigram = feature_temp[0]
feature_weight = weight_dict[bigram][feature_temp]
f_j_1 = feature_function_bigram(feature_temp, sent, gs_label)
f_j_2 = feature_function_bigram(feature_temp, sent, classify_label)
weight_update = feature_weight + 0.01*f_j_1 - 0.01*f_j_2
weight_dict[bigram][feature_temp] = weight_update
def predict_bigram(test_data_lst, w_dict):
"""
Given a sentence and the trained weights for our features - predict the sentiment of a sentence.
Same for Hinge Loss and for Perceptron Loss ?
"""
actual_labels = []
predicted_labels = []
for sent in test_data_lst:
words, gs_label, sentence = extract_features_from_sentence_bigram(sent)
selected_features = get_features(words, w_dict)
classify_label = classify_bigram(selected_features, sentence)
actual_labels.append(gs_label)
predicted_labels.append(classify_label)
return accuracy_score(actual_labels, predicted_labels)
def hinge_loss(weight_dict, sentence):
"""
This code represents the hinge loss function as specified in the assignment.
Note how the function does not return anything. This is because out weights are stored in a dictionary.
Manipulations to the weight dictionary below will occur in-place as the dictionary has a global shared memory location.
"""
words, gs_label, sent = extract_features_from_sentence(sentence)
selected_features = get_features(words, weight_dict)
classify_label = costClassify(
selected_features, sent, gs_label
) # classify label only calculated once at the beginning of a training example. Note use of costClassify and not classify
for feature_temp in selected_features:
word = feature_temp[0]
feature_weight = weight_dict[word][feature_temp]
f_j_1 = feature_function(feature_temp, sent, gs_label)
f_j_2 = feature_function(feature_temp, sent, classify_label)
weight_update = feature_weight + 0.01 * f_j_1 - 0.01 * f_j_2 #sub-gradient descent step. Step size is fixed to 0.01
weight_dict[word][feature_temp] = weight_update