Hybrid Attention-based Approach for Arabic Paraphrase Detection

Figures & data

Figure 1. Proposed architecture for Arabic paraphrase detection.

Table 1. Arabic paraphrased sentence generation

Source	Arabic	اتمنى لك النجاح في العمل
	Translation	I wish you success in business
Tokens		Synonyms
Arabic	Translation	POS	Arabic, Cos	Translation
اتمنى	wish	Verb	ارجو (0.83) | امل (0.84) | طلب (0.78)	hope |trust |request
لك	you	Preposition	لك	you
النجاح	the success	Noun	نجح (0.86) |التوفيق (0.83) | وفاق (0.77)	success |success |accord
في	in	Preposition	في	in
العمل	the business	Noun	مهنة (0.85) | الشغل(0.83) | وظيفة(0.87)	the career |the work |the job
$P_1$	Arabic	امل لك التوفيق في وظيفة
	Translation	I wish you would succeed in a job
$P_2$	Arabic	ارجو لك التوفيق في الشغل
	Translation	I hope you success in the job

Table 2. Experimental datasets

Corpora	Models	# Total Sentence-pairs	# Paraphrased Sentence-pairs	# Different Sentence-pairs
Proposed corpus	Train	720	480	240
	Test	300	200	100
SemEval	Train	510	357	153
	Test	210	147	63

Table 3. Parameter settings

Models	Parameters	Values
GloVe	Size of co-occurrence matrix	1,119,436 × 1,119,436 words
	Vector dimension	300
	Window size	3
	Min_count	20
	Learning rate	0.05
	Batch size	512
CNN	Filters number	64
	Kernel size	3
	Activation function	ReLU
	Pooling size	4
	Dropout	0.5

Figure 2. Experimental results of the proposed models.

Figure 3. Attention-CNN-based approach performances according to the window sizes using OSAC corpus.

Figure 4. Attention-CNN-based Approach performances according to the pooling operations using OSAC Corpus.

Figure 5. Summary of experimental results according to F1 score.

Table 4. Comparison between the proposed approach and the state-of-the-art-based methods

Models	Dataset	Model	Precision	Recall	F1 score
Ours	SemEval	CNN	0.731	0.710	0.720
		Attention CNN	0.775	0.745	0.759
	Paraphrased OSAC	CNN	0.790	0.770	0.779
		Attention CNN	0.805	0.785	0.794
Xie et al. (2020)	817,216 codes sources pairs	Word2vec + Cosine	0.790	0.420	0.550
		WICE	0.690	0.600	0.640
		WICE-SNN	0.670	0.830	0.740

Xie, C., X. Wang, C. Qian, and M. Wang. 2020. A source code similarity based on Siamese neural network. Applied Sciences 10 (21):1–12. doi:https://doi.org/10.3390/app10217519.

 Google Scholar