More than Bags of Words: Sentiment Analysis with Word Embeddings

Figures & data

Figure 1. Supervised sentiment analysis approach with distributed word embeddings.

Figure 2. Illustrative example of mapping sentiment to word embedding dimensions.

Figure 3. Negativity distribution of 20,600 training sentences ranging from 0 (not negative) to 4 (very negative), divided into three classes: not/slightly negative, negative and very negative.

Table 1. Precision, recall, and F1 score for the bag of words approach.

	Actual	Predicted	Precision	Recall	F1 Score
not/slightly negative	524.3	205.6	0.33	0.83	0.47
negative	805.7	1188.7	0.71	0.48	0.57
very negative	730	665.7	0.53	0.58	0.56

Table 2. Precision, recall, and F1 score for the Word Embeddings approach.

	Actual	Predicted	Precision	Recall	F1 Score
not/slightly negative	522.4	575	0.65	0.59	0.61
negative	799.2	771.6	0.52	0.53	0.53
very negative	739.4	714.4	0.55	0.57	0.56

Figure 4. Negativity evolution in the Austrian parliament from 1996 to 2013 showing those four parties that were present over the whole period.

Figure 5. Negativity distinction per legislative term and parliamentary role: Average scores of cabinet members, MPs, and parliamentary party group leaders.

Figure 6. Negativity distinction per legislative term and type of debate: Average scores of Urgent Question Debates and all other types of debates.

Table B1 Average confusion matrix for unseen test data: Bag-of-words

Predicted
		not/slightly negative	Negative	very negative
Actual	not/slightly negative	170.6	284.6	69.1
	negative	27.3	569.7	208.7
	very negative	7.7	334.4	387.9

Table B2. Average confusion matrix for unseen test data: word embeddings.

Predicted
		not/slightly negative	Negative	very negative
Actual	not/slightly negative	337.2	120.3	64.9
	negative	146.5	412.7	240.0
	very negative	91.3	238.6	409.5