Multi-level Community-awareness Graph Neural Networks for Neural Machine Translation

Proceedings of the 29th International Conference on Computational Linguistics, pages 5021–5028 October 12–17, 2022. 5021 Multi-level Community-awareness Graph Neural Networks for Neural Machine Translation Binh Nguyen1,2, Long Nguyen1,2∗, Dien Dinh1,2 1Faculty of Information Technology, University of Science, Ho Chi Minh City, Vietnam 2Vietnam National University, Ho Chi Minh City, Vietnam nqbinh17@apcs.fitus.edu.vn, {nhblong,ddien}@fit.hcmus.edu.vn Abstract Neural Machine Translation (NMT) aims to translate the source- to the target-language while preserving the original meaning. Lin- guistic information such as morphology, syn- tactic, and semantics shall be grasped in token embeddings to produce a high-quality transla- tion. Recent works have leveraged the power- ful Graph Neural Networks (GNNs) to encode such language knowledge into token embed- dings. Specifically, they use a trained parser to construct semantic graphs given sentences and then apply GNNs. However, most semantic graphs are tree-shaped and too sparse for GNNs which cause the over-smoothing problem. To alleviate this problem, we propose a novel Multi-level Community-awareness Graph Neu- ral Network (MC-GNN) layer to jointly model local and global relationships between words and their linguistic roles in multiple communi- ties. Intuitively, the MC-GNN layer substitutes a self-attention layer at the encoder side of a transformer-based machine translation model. Extensive experiments on four language-pair datasets with common evaluation metrics show the remarkable improvements of our method while reducing the time complexity in very long sentences. 1 Introduction Self-attention mechanisms, introduced by Trans- formers (Vaswani et al., 2017), have been ubiq- uitous in Neural Machine Translation (NMT). It encodes the relative information of a position based on other positions in the same sequence. Recent works have seen the rising trend of improving the self-attention module such as (Choromanski et al., 2021; Katharopoulos et al., 2020). One of the serious self-attention weaknesses is its inefficiency at computing long sentences due to quadratic complexity. Following this problem, Per- former (Choromanski et al., 2021) estimates a full- ∗Corresponding author rank-attention matrix with Fast Attention Via pos- itive Orthogonal Random features approach (FA- VOR+), that trade-offs the time complexity from O(T 2d) to O(Td2logd). However, as they scale the model larger such as GPT-3 (Brown et al., 2020), the quadratic dimensional space d exponentially scales much faster than that of the sequence length T. In addition, Linear Transformer (Katharopoulos et al., 2020) reported a comparable performance with the Transformer, while speeding up on autore- gressive inference time with a linear dot-product of kernel feature maps. Several studies replace self-attention with an attention-free module, which significantly speeds up the training time. Atten- tion Free Transformer (ATF) (Zhai et al., 2021) eliminates the dot-product attention with weighted element-wise multiplication, which linearizes the time complexity to O(Td). FNet (Lee-Thorp et al., 2021) is a pre-trained model with Discrete Fourier Transform that speeds up 80% training time on GPU but still achieves 92-97% of the accuracy of BERT (Devlin et al., 2018). These works make a trade-off between the accuracy and the perfor- mance to reduce the computational cost. Another line of work constructs a content-based sparse graph to select keys that have …

Content truncated to 500 words