GEM: A General Evaluation Benchmark for Multimodal Tasks

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 256–268 August 1–6, 2021. ©2021 Association for Computational Linguistics 256 How does Attention Affect the Model? Cheng Zhang1,Qiuchi Li2, Lingyu Hua3, Dawei Song 3∗ 1College of Intelligence and Computing, Tianjin University, China 2Department of Information Engineering, University of Padua, Italy 3School of Computer Science, Beijing Institute of Technology, China cheng.zhang@tju.edu.cn qiuchili@dei.unipd.it hualingyu@hotmail.com dwsong@bit.edu.cn Abstract The attention layer has become a prevalent component in improving the effectiveness of neural network models for NLP tasks. Figur- ing out why attention is effective and its inter- pretability has attracted a widespread deliber- ation. Current studies mostly investigate the effect of attention mechanism based on the at- tention distribution it generates with one single neural network structure. However they do not consider the changes in semantic capability of different components in the model due to the attention mechanism, which can vary across different network structures. In this paper, we propose a comprehensive analytical frame- work that exploits a convex hull representation of sequence semantics in an n-dimensional Se- mantic Euclidean Space and defines a series of indicators to capture the impact of attention on sequence semantics. Through a series of experiments on various NLP tasks and three representative recurrent units, we analyze why and how attention benefits the semantic capac- ity of different types of recurrent neural net- works based on the indicators defined in the proposed framework. 1 Introduction and Motivation The first appearance of the attention mechanism in natural language processing (NLP) can be traced back to its successful application in Neural Ma- chine Translation (NMT). Bahdanau et al. (2014a) proposed an attention mechanism in an Encoder- Decoder model, which achieved great success, and showed that attention weight produced in this mech- anism improved the interpretability of the model by providing a way of aligning the source and target languages through a simple quantitative analysis. Subsequently, the assumption that attention could improve the interpretability and transparency of a model was acquiesced by many later works, such ∗ Corresponding Author: Dawei Song as AEN (Song et al., 2019) (applied to targeted sen- timent classification), ATAE-LSTM (Wang et al., 2016) (applied to aspect-level sentiment classifica- tion), and CMLA (Wang et al., 2017) (applied to semantic sentiment analysis). More recently, this hypothetical premise has aroused controversies. For example, Serrano and Smith (2019) and Jain and Wallace (2019) used an erasure-based approach and advocated the at- tention weight does not necessarily correspond to importance. Wiegreffe and Pinter (2019) and Vashishth et al. (2019) considered attention to be interpretable, using a more model-driven approach and manual verification. These investigations fo- cus on whether the attention distribution is unique and the correlation between attention weight and model prediction results, based on a similar neural network setting that consists of an embedding layer, a specific Recurrent Neural Network (RNN) and an attention component. Complex components such as the encoder-decoder structure were removed from the network as they may bias the analysis on the effect of attention weights. However, these …

Content truncated to 500 words