Exploiting Syntactic and Semantic Information for Textual Similarity Estimation

Table 3

Comparison results of the proposed method and other state-of-the-art methods on the STS tasks (2012–2015).

Year	Dataset	Compared methods					The method
		Glove	Word2vec	RNN	LSTM	iRNN	WVD-tree
2012	MSRpar	47.8	49.2	18.6	16.1	43.5	64.3
	MSRvid	63.8	65.6	66.5	71.3	73.3	80.6
	SMTeuroparl	46.2	45.1	40.9	41.8	47.2	51.7
	OnWN	55.3	55.9	63.1	65.2	70.2	68.8
	SMTnews	50.3	52.3	51.3	51.0	58.3	53.2
2013	Headlines	63.9	65.5	59.5	57.4	72.7	78.1
	OnWN	49.1	52.4	54.6	50.4	69.3	80.6
	FNWN	34.3	37.9	30.9	38.4	45.2	50.3
2014	Deft-forum	27.2	38.7	41.5	46.1	49.1	49.4
	Deft-news	68.3	69.3	53.7	39.1	72.3	74.1
	Headlines	59.8	63.8	57.5	50.9	70.3	71.7
	Images	61.2	61.7	67.6	62.8	78.3	74.2
	OnWN	58.3	57.4	67.7	61.5	78.7	85.5
	Tweet-news	51.1	53.4	58.0	48.3	76.7	73.5
2015	Answers-forums	30.7	32.3	32.8	51.4	67.2	70.2
	Belief	40.7	42.3	51.9	52.7	75.8	73.4
	Headlines	61.9	63.2	65.3	56.8	75.0	80.8
	Images	67.6	68.5	71.4	64.3	81.2	83.4

Bold values denote the optimal results among the horizontal experimental results.