Attention Is All You Need

Provided proper attribution is provided, Google hereby grants permission to reproduce the tables and figures in this paper solely for use in journalistic or scholarly works.

在给予适当署名的前提下，Google 特此允许复制本文中的表格和图片，但仅限用于新闻报道或学术作品。

Attention Is All You Need

注意力就是你所需要的一切

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, Illia Polosukhin

Abstract

摘要

The dominant sequence transduction models are based on complex recurrent or convolutional neural networks that include an encoder and a decoder. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles, by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data.

目前主流的序列转换模型基于复杂的循环神经网络或卷积神经网络，包含编码器和解码器。表现最好的模型还通过注意力机制连接编码器与解码器。我们提出了一种新的简洁网络架构 Transformer，它完全基于注意力机制，彻底舍弃循环结构和卷积。在两个机器翻译任务上的实验表明，这类模型不仅质量更高，而且更易于并行化，所需训练时间也显著缩短。我们的模型在 WMT 2014 英语到德语翻译任务上取得了 28.4 BLEU，比包括集成模型在内的已有最佳结果提高了超过 2 BLEU。在 WMT 2014 英语到法语翻译任务上，我们的模型使用 8 块 GPU 训练 3.5 天后，取得了 41.8 BLEU，创下新的单模型最佳水平；其训练成本仅为文献中最佳模型的一小部分。我们还在训练数据充足和有限两种条件下，将 Transformer 成功应用于英语成分句法分析，表明它能够很好地泛化到其他任务。

Equal contribution. Listing order is random. Jakob proposed replacing RNNs with self-attention and started the effort to evaluate this idea. Ashish, with Illia, designed and implemented the first Transformer models and has been crucially involved in every aspect of this work. Noam proposed scaled dot-product attention, multi-head attention and the parameter-free position representation and became the other person involved in nearly every detail. Niki designed, implemented, tuned and evaluated countless model variants in our original codebase and tensor2tensor. Llion also experimented with novel model variants, was responsible for our initial codebase, and efficient inference and visualizations. Lukasz and Aidan spent countless long days designing various parts of and implementing tensor2tensor, replacing our earlier codebase, greatly improving results and massively accelerating our research.

各位作者贡献相同，排名顺序随机。Jakob 提出用自注意力替代 RNN，并开始验证这一想法。Ashish 与 Illia 一起设计并实现了最早的 Transformer 模型，且在这项工作的每个方面都发挥了关键作用。Noam 提出了缩放点积注意力、多头注意力以及无参数的位置表示，是另一位几乎参与所有细节的作者。Niki 在我们的原始代码库和 tensor2tensor 中设计、实现、调优并评估了无数模型变体。Llion 也试验了新的模型变体，负责我们的初始代码库、高效推理和可视化。Lukasz 和 Aidan 投入了无数个漫长的工作日，设计并实现 tensor2tensor 的各个部分，用它替换了早期代码库，大幅改善了结果，并极大加快了我们的研究进程。

31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.

第 31 届神经信息处理系统大会（NIPS 2017），美国加利福尼亚州长滩。

arXiv:1706.03762v7 [cs.CL] 2 Aug 2023

arXiv:1706.03762v7 [cs.CL]，2023 年 8 月 2 日。

1 Introduction

1 引言

Recurrent neural networks, long short-term memory [13] and gated recurrent [7] neural networks in particular, have been firmly established as state of the art approaches in sequence modeling and transduction problems such as language modeling and machine translation [35, 2, 5]. Numerous efforts have since continued to push the boundaries of recurrent language models and encoder-decoder architectures [38, 24, 15].

循环神经网络，尤其是长短期记忆网络 [13] 和门控循环神经网络 [7]，已经成为语言建模、机器翻译等序列建模与序列转换问题中公认的最先进方法 [35, 2, 5]。此后，大量研究持续推进循环语言模型和编码器—解码器架构的能力边界 [38, 24, 15]。

Recurrent models typically factor computation along the symbol positions of the input and output sequences. Aligning the positions to steps in computation time, they generate a sequence of hidden states $h_t$, as a function of the previous hidden state $h_{t-1}$ and the input for position $t$. This inherently sequential nature precludes parallelization within training examples, which becomes critical at longer sequence lengths, as memory constraints limit batching across examples. Recent work has achieved significant improvements in computational efficiency through factorization tricks [21] and conditional computation [32], while also improving model performance in case of the latter. The fundamental constraint of sequential computation, however, remains.

循环模型通常沿着输入序列和输出序列中的符号位置来分解计算。它们将这些位置对应到计算时间上的各个步骤，并根据前一个隐藏状态 $h_{t-1}$ 和位置 $t$ 的输入生成隐藏状态序列 $h_t$。这种内在的顺序性使单个训练样本内部无法并行计算；在序列较长时，这一问题尤为突出，因为内存限制了跨样本的批量处理规模。近期研究通过分解技巧 [21] 和条件计算 [32] 显著提高了计算效率，后者还改善了模型性能。然而，顺序计算这一根本限制仍然存在。

Attention mechanisms have become an integral part of compelling sequence modeling and transduction models in various tasks, allowing modeling of dependencies without regard to their distance in the input or output sequences [2, 19]. In all but a few cases [27], however, such attention mechanisms are used in conjunction with a recurrent network.

在多种任务中，注意力机制已成为表现优异的序列建模与序列转换模型不可或缺的一部分，使模型能够对依赖关系建模，而不受相关元素在输入或输出序列中距离远近的限制 [2, 19]。然而，除了少数情况 [27]，这类注意力机制都与循环网络配合使用。

In this work we propose the Transformer, a model architecture eschewing recurrence and instead relying entirely on an attention mechanism to draw global dependencies between input and output. The Transformer allows for significantly more parallelization and can reach a new state of the art in translation quality after being trained for as little as twelve hours on eight P100 GPUs.

在这项工作中，我们提出 Transformer：一种舍弃循环结构、完全依靠注意力机制来建立输入与输出之间全局依赖关系的模型架构。Transformer 支持显著更高的并行程度，仅需在 8 块 P100 GPU 上训练 12 小时，就能在翻译质量上达到新的最先进水平。

2 Background

2 背景

The goal of reducing sequential computation also forms the foundation of the Extended Neural GPU [16], ByteNet [18] and ConvS2S [9], all of which use convolutional neural networks as basic building block, computing hidden representations in parallel for all input and output positions. In these models, the number of operations required to relate signals from two arbitrary input or output positions grows in the distance between positions, linearly for ConvS2S and logarithmically for ByteNet. This makes it more difficult to learn dependencies between distant positions [12]. In the Transformer this is reduced to a constant number of operations, albeit at the cost of reduced effective resolution due to averaging attention-weighted positions, an effect we counteract with Multi-Head Attention as described in section 3.2.

减少顺序计算也是 Extended Neural GPU [16]、ByteNet [18] 和 ConvS2S [9] 的基本出发点。这些模型都以卷积神经网络为基本构建模块，并行计算所有输入和输出位置的隐藏表示。在这些模型中，将任意两个输入位置或输出位置的信号关联起来所需的运算次数，会随着位置间距离增加：ConvS2S 呈线性增长，ByteNet 呈对数增长。这使得学习远距离位置之间的依赖关系更加困难 [12]。在 Transformer 中，所需运算次数被降为常数，但代价是对各位置按注意力权重进行平均会降低有效分辨率；我们用第 3.2 节介绍的多头注意力来抵消这一影响。

Self-attention, sometimes called intra-attention is an attention mechanism relating different positions of a single sequence in order to compute a representation of the sequence. Self-attention has been used successfully in a variety of tasks including reading comprehension, abstractive summarization, textual entailment and learning task-independent sentence representations [4, 27, 28, 22].

自注意力有时也称为内部注意力（intra-attention），是一种将同一序列内不同位置关联起来，以计算该序列表示的注意力机制。自注意力已成功应用于多种任务，包括阅读理解、生成式摘要、文本蕴含，以及学习与任务无关的句子表示 [4, 27, 28, 22]。

End-to-end memory networks are based on a recurrent attention mechanism instead of sequence-aligned recurrence and have been shown to perform well on simple-language question answering and language modeling tasks [34].

端到端记忆网络采用循环注意力机制，而非与序列位置对齐的循环结构；已有研究表明，它们在使用简单语言的问答任务和语言建模任务上表现良好 [34]。

To the best of our knowledge, however, the Transformer is the first transduction model relying entirely on self-attention to compute representations of its input and output without using sequence-aligned RNNs or convolution. In the following sections, we will describe the Transformer, motivate self-attention and discuss its advantages over models such as [17, 18] and [9].

然而，据我们所知，Transformer 是首个完全依靠自注意力计算输入和输出表示的序列转换模型，它不使用与序列位置对齐的 RNN，也不使用卷积。在接下来的章节中，我们将介绍 Transformer，阐述采用自注意力的动机，并讨论它相对于 [17, 18] 和 [9] 等模型的优势。

3 Model Architecture

3 模型架构

Most competitive neural sequence transduction models have an encoder-decoder structure [5, 2, 35]. Here, the encoder maps an input sequence of symbol representations $(x_1, ..., x_n)$ to a sequence of continuous representations $\mathbf{z} = (z_1, ..., z_n)$. Given $\mathbf{z}$, the decoder then generates an output sequence $(y_1, ..., y_m)$ of symbols one element at a time. At each step the model is auto-regressive [10], consuming the previously generated symbols as additional input when generating the next.

大多数有竞争力的神经序列转换模型都采用编码器—解码器结构 [5, 2, 35]。其中，编码器将由符号表示构成的输入序列 $(x_1, ..., x_n)$ 映射为连续表示序列 $\mathbf{z} = (z_1, ..., z_n)$。给定 $\mathbf{z}$ 后，解码器逐个元素地生成符号输出序列 $(y_1, ..., y_m)$。模型在每一步都采用自回归方式 [10]，即在生成下一个符号时，将此前生成的符号作为额外输入。

Figure 1: The Transformer - model architecture.

图 1：Transformer 模型架构。

The Transformer follows this overall architecture using stacked self-attention and point-wise, fully connected layers for both the encoder and decoder, shown in the left and right halves of Figure 1, respectively.

Transformer 遵循这一总体架构，编码器和解码器都使用堆叠的自注意力层与逐位置全连接层，分别如图 1 的左半部分和右半部分所示。

3.1 Encoder and Decoder Stacks

3.1 编码器和解码器的堆叠结构

Encoder: The encoder is composed of a stack of $N = 6$ identical layers. Each layer has two sub-layers. The first is a multi-head self-attention mechanism, and the second is a simple, position-wise fully connected feed-forward network. We employ a residual connection [11] around each of the two sub-layers, followed by layer normalization [1]. That is, the output of each sub-layer is $\operatorname{LayerNorm}(x + \operatorname{Sublayer}(x))$, where $\operatorname{Sublayer}(x)$ is the function implemented by the sub-layer itself. To facilitate these residual connections, all sub-layers in the model, as well as the embedding layers, produce outputs of dimension $d_{\text{model}} = 512$.

编码器： 编码器由 $N = 6$ 个结构相同的层堆叠而成。每层包含两个子层：第一个是多头自注意力机制，第二个是简单的逐位置全连接前馈网络。我们在这两个子层的每一个外部都使用残差连接 [11]，随后进行层归一化 [1]。也就是说，每个子层的输出为 $\operatorname{LayerNorm}(x + \operatorname{Sublayer}(x))$，其中 $\operatorname{Sublayer}(x)$ 是该子层本身实现的函数。为便于使用这些残差连接，模型中的所有子层以及嵌入层都输出维度为 $d_{\text{model}} = 512$ 的表示。

Decoder: The decoder is also composed of a stack of $N = 6$ identical layers. In addition to the two sub-layers in each encoder layer, the decoder inserts a third sub-layer, which performs multi-head attention over the output of the encoder stack. Similar to the encoder, we employ residual connections around each of the sub-layers, followed by layer normalization. We also modify the self-attention sub-layer in the decoder stack to prevent positions from attending to subsequent positions. This masking, combined with fact that the output embeddings are offset by one position, ensures that the predictions for position $i$ can depend only on the known outputs at positions less than $i$.

解码器： 解码器同样由 $N = 6$ 个结构相同的层堆叠而成。除了编码器每层中的两个子层，解码器还插入了第三个子层，对编码器堆叠结构的输出执行多头注意力。与编码器类似，我们在每个子层外部使用残差连接，随后进行层归一化。我们还修改了解码器堆叠结构中的自注意力子层，防止当前位置关注后续位置。这种掩码机制，加上输出嵌入错开一个位置的做法，确保位置 $i$ 的预测只能依赖位置小于 $i$ 的已知输出。

3.2 Attention

3.2 注意力

An attention function can be described as mapping a query and a set of key-value pairs to an output, where the query, keys, values, and output are all vectors. The output is computed as a weighted sum of the values, where the weight assigned to each value is computed by a compatibility function of the query with the corresponding key.

注意力函数可以描述为将一个查询和一组键值对映射为一个输出，其中查询、键、值和输出都是向量。输出通过对各个值进行加权求和得到，而分配给每个值的权重，则由查询与对应键之间的兼容性函数计算得出。

Figure 2: (left) Scaled Dot-Product Attention. (right) Multi-Head Attention consists of several attention layers running in parallel.

图 2：（左）缩放点积注意力。（右）多头注意力由多个并行运行的注意力层组成。

3.2.1 Scaled Dot-Product Attention

3.2.1 缩放点积注意力

We call our particular attention "Scaled Dot-Product Attention" (Figure 2). The input consists of queries and keys of dimension $d_k$, and values of dimension $d_v$. We compute the dot products of the query with all keys, divide each by $\sqrt{d_k}$, and apply a softmax function to obtain the weights on the values.

我们将所采用的这一注意力机制称为“缩放点积注意力”（图 2）。输入由维度为 $d_k$ 的查询和键，以及维度为 $d_v$ 的值组成。我们计算查询与所有键的点积，将每个点积除以 $\sqrt{d_k}$，再应用 softmax 函数，得到各个值的权重。

In practice, we compute the attention function on a set of queries simultaneously, packed together into a matrix $Q$. The keys and values are also packed together into matrices $K$ and $V$. We compute the matrix of outputs as:

$$\operatorname{Attention}(Q, K, V) = \operatorname{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)V \tag{1}$$

在实际计算中，我们将一组查询组合成矩阵 $Q$，同时计算它们的注意力函数。键和值也分别组合成矩阵 $K$ 和 $V$。输出矩阵的计算方式为：

The two most commonly used attention functions are additive attention [2], and dot-product (multiplicative) attention. Dot-product attention is identical to our algorithm, except for the scaling factor of $\frac{1}{\sqrt{d_k}}$. Additive attention computes the compatibility function using a feed-forward network with a single hidden layer. While the two are similar in theoretical complexity, dot-product attention is much faster and more space-efficient in practice, since it can be implemented using highly optimized matrix multiplication code.

最常用的两种注意力函数是加性注意力 [2] 和点积（乘性）注意力。除缩放因子 $\frac{1}{\sqrt{d_k}}$ 外，点积注意力与我们的算法相同。加性注意力使用一个仅有单个隐藏层的前馈网络来计算兼容性函数。虽然两者在理论复杂度上相近，但在实际运行中，点积注意力速度更快、空间利用效率更高，因为它可以通过高度优化的矩阵乘法代码来实现。

While for small values of $d_k$ the two mechanisms perform similarly, additive attention outperforms dot product attention without scaling for larger values of $d_k$ [3]. We suspect that for large values of $d_k$, the dot products grow large in magnitude, pushing the softmax function into regions where it has extremely small gradients $^{4}$. To counteract this effect, we scale the dot products by $\frac{1}{\sqrt{d_k}}$.

当 $d_k$ 较小时，两种机制表现相近；但在 $d_k$ 较大时，加性注意力优于未经缩放的点积注意力 [3]。我们推测，当 $d_k$ 较大时，点积的数值幅度会增大，使 softmax 函数进入梯度极小的区域 $^{4}$。为抵消这一影响，我们用 $\frac{1}{\sqrt{d_k}}$ 对点积进行缩放。

$^{4}$ To illustrate why the dot products get large, assume that the components of $q$ and $k$ are independent random variables with mean 0 and variance 1. Then their dot product, $q \cdot k = \sum_{i=1}^{d_k} q_i k_i$, has mean 0 and variance $d_k$.

$^{4}$ 为说明点积为什么会变大，假设 $q$ 和 $k$ 的各分量是均值为 0、方差为 1 的独立随机变量。那么，它们的点积 $q \cdot k = \sum_{i=1}^{d_k} q_i k_i$ 的均值为 0，方差为 $d_k$。

3.2.2 Multi-Head Attention

3.2.2 多头注意力

Instead of performing a single attention function with $d_{\text{model}}$-dimensional keys, values and queries, we found it beneficial to linearly project the queries, keys and values $h$ times with different, learned linear projections to $d_k$, $d_k$ and $d_v$ dimensions, respectively. On each of these projected versions of queries, keys and values we then perform the attention function in parallel, yielding $d_v$-dimensional output values. These are concatenated and once again projected, resulting in the final values, as depicted in Figure 2.

我们发现，相比于使用维度为 $d_{\text{model}}$ 的键、值和查询执行一次注意力函数，更有效的做法是通过不同的可学习线性投影，将查询、键和值分别投影到 $d_k$、$d_k$ 和 $d_v$ 维，重复进行 $h$ 次。随后，我们在这些投影后的查询、键和值上并行执行注意力函数，得到维度为 $d_v$ 的输出值。将这些输出拼接起来，再进行一次投影，就得到最终的值，如图 2 所示。

Multi-head attention allows the model to jointly attend to information from different representation subspaces at different positions. With a single attention head, averaging inhibits this.

$$\begin{aligned} \operatorname{MultiHead}(Q,K,V) &= \operatorname{Concat}(\operatorname{head}_1,\ldots,\operatorname{head}_h)W^O \\ \text{where }\operatorname{head}_i &= \operatorname{Attention}(QW_i^Q,KW_i^K,VW_i^V) \end{aligned}$$

多头注意力使模型能够同时关注不同位置、不同表示子空间中的信息。只有一个注意力头时，求平均的过程会抑制这种能力。

Where the projections are parameter matrices $W_i^Q \in \mathbb{R}^{d_{\text{model}}\times d_k}$, $W_i^K \in \mathbb{R}^{d_{\text{model}}\times d_k}$, $W_i^V \in \mathbb{R}^{d_{\text{model}}\times d_v}$ and $W^O \in \mathbb{R}^{h d_v\times d_{\text{model}}}$.

其中，投影的参数矩阵为 $W_i^Q \in \mathbb{R}^{d_{\text{model}}\times d_k}$、$W_i^K \in \mathbb{R}^{d_{\text{model}}\times d_k}$、$W_i^V \in \mathbb{R}^{d_{\text{model}}\times d_v}$ 和 $W^O \in \mathbb{R}^{h d_v\times d_{\text{model}}}$。

In this work we employ $h = 8$ parallel attention layers, or heads. For each of these we use $d_k = d_v = d_{\text{model}}/h = 64$. Due to the reduced dimension of each head, the total computational cost is similar to that of single-head attention with full dimensionality.

在本研究中，我们使用 $h = 8$ 个并行的注意力层，也称为注意力头。每个头都采用 $d_k = d_v = d_{\text{model}}/h = 64$。由于每个头的维度降低，总计算成本与使用完整维度的单头注意力相近。

3.2.3 Applications of Attention in our Model

3.2.3 注意力在模型中的应用

The Transformer uses multi-head attention in three different ways:

Transformer 以三种不同的方式使用多头注意力：

• In "encoder-decoder attention" layers, the queries come from the previous decoder layer, and the memory keys and values come from the output of the encoder. This allows every position in the decoder to attend over all positions in the input sequence. This mimics the typical encoder-decoder attention mechanisms in sequence-to-sequence models such as [38, 2, 9].

• 在“编码器—解码器注意力”层中，查询来自前一层解码器，而记忆中的键和值来自编码器的输出。这使解码器中的每个位置都能关注输入序列中的所有位置。这种方式仿照了 [38, 2, 9] 等序列到序列模型中常见的编码器—解码器注意力机制。

• The encoder contains self-attention layers. In a self-attention layer all of the keys, values and queries come from the same place, in this case, the output of the previous layer in the encoder. Each position in the encoder can attend to all positions in the previous layer of the encoder.

• 编码器包含自注意力层。在自注意力层中，所有键、值和查询都来自同一处，这里就是编码器前一层的输出。编码器中的每个位置都可以关注编码器前一层的所有位置。

• Similarly, self-attention layers in the decoder allow each position in the decoder to attend to all positions in the decoder up to and including that position. We need to prevent leftward information flow in the decoder to preserve the auto-regressive property. We implement this inside of scaled dot-product attention by masking out (setting to $-\infty$) all values in the input of the softmax which correspond to illegal connections. See Figure 2.

• 类似地，解码器中的自注意力层允许每个位置关注解码器中直到当前位置为止的所有位置，包括当前位置本身。为了保持自回归特性，我们需要阻止信息在解码器中向左流动。我们在缩放点积注意力内部实现这一点：将 softmax 输入中对应非法连接的所有值屏蔽掉，即设为 $-\infty$。见图 2。

3.3 Position-wise Feed-Forward Networks

3.3 逐位置前馈网络

In addition to attention sub-layers, each of the layers in our encoder and decoder contains a fully connected feed-forward network, which is applied to each position separately and identically. This consists of two linear transformations with a ReLU activation in between.

$$\operatorname{FFN}(x) = \max(0,xW_1+b_1)W_2+b_2 \tag{2}$$

除了注意力子层，我们的编码器和解码器中的每一层还包含一个全连接前馈网络，以相同的方式分别作用于每个位置。该网络由两次线性变换组成，两者之间使用 ReLU 激活函数。

While the linear transformations are the same across different positions, they use different parameters from layer to layer. Another way of describing this is as two convolutions with kernel size 1. The dimensionality of input and output is $d_{\text{model}} = 512$, and the inner-layer has dimensionality $d_{\text{ff}} = 2048$.

虽然不同位置使用相同的线性变换，但不同层使用不同的参数。也可以将其描述为两次卷积核大小为 1 的卷积。输入和输出的维度为 $d_{\text{model}} = 512$，内部层的维度为 $d_{\text{ff}} = 2048$。

3.4 Embeddings and Softmax

3.4 嵌入与 Softmax

Similarly to other sequence transduction models, we use learned embeddings to convert the input tokens and output tokens to vectors of dimension $d_{\text{model}}$. We also use the usual learned linear transformation and softmax function to convert the decoder output to predicted next-token probabilities. In our model, we share the same weight matrix between the two embedding layers and the pre-softmax linear transformation, similar to [30]. In the embedding layers, we multiply those weights by $\sqrt{d_{\text{model}}}$.

与其他序列转换模型类似，我们使用可学习的嵌入，将输入词元和输出词元转换为维度为 $d_{\text{model}}$ 的向量。我们还使用常规的可学习线性变换和 softmax 函数，将解码器输出转换为对下一个词元概率的预测。与 [30] 类似，在我们的模型中，两个嵌入层与 softmax 之前的线性变换共享同一个权重矩阵。在嵌入层中，我们将这些权重乘以 $\sqrt{d_{\text{model}}}$。

Table 1: Maximum path lengths, per-layer complexity and minimum number of sequential operations for different layer types. n is the sequence length, d is the representation dimension, k is the kernel size of convolutions and r the size of the neighborhood in restricted self-attention.

表 1：不同类型的层所对应的最大路径长度、每层计算复杂度，以及所需串行操作的最少数量。其中，$n$ 为序列长度，$d$ 为表示维度，$k$ 为卷积核大小，$r$ 为受限自注意力中的邻域大小。

3.5 Positional Encoding

3.5 位置编码

Since our model contains no recurrence and no convolution, in order for the model to make use of the order of the sequence, we must inject some information about the relative or absolute position of the tokens in the sequence. To this end, we add "positional encodings" to the input embeddings at the bottoms of the encoder and decoder stacks. The positional encodings have the same dimension $d_{\text{model}}$ as the embeddings, so that the two can be summed. There are many choices of positional encodings, learned and fixed [9].

由于我们的模型既没有循环结构，也没有卷积，为了让模型利用序列中的顺序，我们必须注入一些关于词元在序列中相对位置或绝对位置的信息。为此，我们在编码器和解码器堆叠的底部，将“位置编码”加到输入嵌入中。位置编码与嵌入的维度相同，均为 $d_{\text{model}}$，因此两者可以相加。位置编码有多种选择，既有可学习的，也有固定的 [9]。

In this work, we use sine and cosine functions of different frequencies:

$$\begin{aligned} PE_{(pos,2i)} &= \sin\left(pos/10000^{2i/d_{\text{model}}}\right) \\ PE_{(pos,2i+1)} &= \cos\left(pos/10000^{2i/d_{\text{model}}}\right) \end{aligned}$$

在本研究中，我们使用不同频率的正弦函数和余弦函数：

where $pos$ is the position and $i$ is the dimension. That is, each dimension of the positional encoding corresponds to a sinusoid. The wavelengths form a geometric progression from $2\pi$ to $10000 \cdot 2\pi$. We chose this function because we hypothesized it would allow the model to easily learn to attend by relative positions, since for any fixed offset $k$, $PE_{pos+k}$ can be represented as a linear function of $PE_{pos}$.

其中，$pos$ 表示位置，$i$ 表示维度。也就是说，位置编码的每个维度都对应一个正弦波。波长从 $2\pi$ 到 $10000 \cdot 2\pi$ 构成等比数列。我们选择这个函数，是因为我们假设它能够让模型容易地学会根据相对位置分配注意力：对于任意固定的偏移量 $k$，$PE_{pos+k}$ 都可以表示为 $PE_{pos}$ 的线性函数。

We also experimented with using learned positional embeddings [9] instead, and found that the two versions produced nearly identical results (see Table 3 row (E)). We chose the sinusoidal version because it may allow the model to extrapolate to sequence lengths longer than the ones encountered during training.

我们也尝试改用可学习的位置嵌入 [9]，发现两个版本的结果几乎相同（见表 3 的 (E) 行）。我们选择正弦形式的版本，是因为它可能让模型外推到比训练时所遇到的序列更长的序列。

4 Why Self-Attention

4 为什么使用自注意力

In this section we compare various aspects of self-attention layers to the recurrent and convolutional layers commonly used for mapping one variable-length sequence of symbol representations $(x_1,\ldots,x_n)$ to another sequence of equal length $(z_1,\ldots,z_n)$, with $x_i,z_i \in \mathbb{R}^d$, such as a hidden layer in a typical sequence transduction encoder or decoder. Motivating our use of self-attention we consider three desiderata.

在本节中，我们从多个方面比较自注意力层，以及常用于将一个变长符号表示序列 $(x_1,\ldots,x_n)$ 映射为另一个等长序列 $(z_1,\ldots,z_n)$ 的循环层和卷积层，其中 $x_i,z_i \in \mathbb{R}^d$。典型序列转换编码器或解码器中的隐藏层就是这种映射的例子。我们从三个期望达到的目标出发，说明使用自注意力的动机。

One is the total computational complexity per layer. Another is the amount of computation that can be parallelized, as measured by the minimum number of sequential operations required.

第一个目标涉及每层的总计算复杂度。第二个目标涉及能够并行化的计算量，我们通过所需串行操作的最少数量来衡量它。

The third is the path length between long-range dependencies in the network. Learning long-range dependencies is a key challenge in many sequence transduction tasks. One key factor affecting the ability to learn such dependencies is the length of the paths forward and backward signals have to traverse in the network. The shorter these paths between any combination of positions in the input and output sequences, the easier it is to learn long-range dependencies [12]. Hence we also compare the maximum path length between any two input and output positions in networks composed of the different layer types.

第三个目标涉及网络中长程依赖之间的路径长度。在许多序列转换任务中，学习长程依赖是一项关键挑战。影响模型学习这类依赖能力的一个关键因素，是前向信号和反向信号在网络中必须经过的路径长度。输入和输出序列中任意位置组合之间的路径越短，学习长程依赖就越容易 [12]。因此，我们还比较了由不同类型的层组成的网络中，任意输入位置与输出位置之间的最大路径长度。

As noted in Table 1, a self-attention layer connects all positions with a constant number of sequentially executed operations, whereas a recurrent layer requires $O(n)$ sequential operations. In terms of computational complexity, self-attention layers are faster than recurrent layers when the sequence length $n$ is smaller than the representation dimensionality $d$, which is most often the case with sentence representations used by state-of-the-art models in machine translations, such as word-piece [38] and byte-pair [31] representations. To improve computational performance for tasks involving very long sequences, self-attention could be restricted to considering only a neighborhood of size $r$ in the input sequence centered around the respective output position. This would increase the maximum path length to $O(n/r)$. We plan to investigate this approach further in future work.

如表 1 所示，自注意力层只需常数个串行执行的操作就能连接所有位置，而循环层需要 $O(n)$ 个串行操作。从计算复杂度来看，当序列长度 $n$ 小于表示维度 $d$ 时，自注意力层比循环层更快。当时最先进的机器翻译模型使用的句子表示通常符合这种情况，例如词片段（word-piece）[38] 和字节对（byte-pair）[31] 表示。为了改善涉及超长序列任务的计算性能，可以将自注意力限制为仅考虑输入序列中以对应输出位置为中心、大小为 $r$ 的邻域。这样会将最大路径长度增加到 $O(n/r)$。我们计划在后续工作中进一步研究这种方法。

A single convolutional layer with kernel width $k < n$ does not connect all pairs of input and output positions. Doing so requires a stack of $O(n/k)$ convolutional layers in the case of contiguous kernels, or $O(\log_k(n))$ in the case of dilated convolutions [18], increasing the length of the longest paths between any two positions in the network. Convolutional layers are generally more expensive than recurrent layers, by a factor of $k$. Separable convolutions [6], however, decrease the complexity considerably, to $O(k \cdot n \cdot d + n \cdot d^2)$. Even with $k = n$, however, the complexity of a separable convolution is equal to the combination of a self-attention layer and a point-wise feed-forward layer, the approach we take in our model.

卷积核宽度为 $k < n$ 的单个卷积层无法连接所有输入位置与输出位置的组合。要做到这一点，连续卷积核需要堆叠 $O(n/k)$ 个卷积层，空洞卷积 [18] 则需要 $O(\log_k(n))$ 个卷积层，这会增加网络中任意两个位置之间的最长路径。卷积层的计算成本通常为循环层的 $k$ 倍。不过，可分离卷积 [6] 能显著降低复杂度，将其降至 $O(k \cdot n \cdot d + n \cdot d^2)$。然而，即使在 $k = n$ 时，可分离卷积的复杂度也只是与一个自注意力层加上一个逐位置前馈层的复杂度相同，而这正是我们的模型所采用的组合。

As side benefit, self-attention could yield more interpretable models. We inspect attention distributions from our models and present and discuss examples in the appendix. Not only do individual attention heads clearly learn to perform different tasks, many appear to exhibit behavior related to the syntactic and semantic structure of the sentences.

自注意力还可能带来一个额外好处：让模型更具可解释性。我们检查了模型中的注意力分布，并在附录中展示和讨论了一些例子。各个注意力头不仅明显学会了执行不同的任务，其中许多头还表现出似乎与句子句法结构和语义结构相关的行为。

5 Training

5 训练

This section describes the training regime for our models.

本节介绍模型的训练方案。

5.1 Training Data and Batching

5.1 训练数据与批次构建

We trained on the standard WMT 2014 English-German dataset consisting of about 4.5 million sentence pairs. Sentences were encoded using byte-pair encoding [3], which has a shared source-target vocabulary of about 37000 tokens. For English-French, we used the significantly larger WMT 2014 English-French dataset consisting of 36M sentences and split tokens into a 32000 word-piece vocabulary [38]. Sentence pairs were batched together by approximate sequence length. Each training batch contained a set of sentence pairs containing approximately 25000 source tokens and 25000 target tokens.

我们在标准的 WMT 2014 英德数据集上进行训练，该数据集包含约 450 万个句子对。句子使用字节对编码 [3] 进行编码，源语言和目标语言共享一个约有 37000 个词元的词表。对于英法翻译，我们使用规模大得多的 WMT 2014 英法数据集，其中包含 3600 万个句子，并将词元切分为包含 32000 个词片段的词表 [38]。我们根据近似的序列长度将句子对组成批次。每个训练批次包含一组句子对，总计约有 25000 个源语言词元和 25000 个目标语言词元。

5.2 Hardware and Schedule

5.2 硬件与训练安排

We trained our models on one machine with 8 NVIDIA P100 GPUs. For our base models using the hyperparameters described throughout the paper, each training step took about 0.4 seconds. We trained the base models for a total of 100,000 steps or 12 hours. For our big models, (described on the bottom line of table 3), step time was 1.0 seconds. The big models were trained for 300,000 steps (3.5 days).

我们在一台配有 8 块 NVIDIA P100 GPU 的机器上训练模型。对于采用本文所述超参数的基础模型，每个训练步骤耗时约 0.4 秒。基础模型共训练了 100000 步，耗时 12 小时。对于大型模型（见表 3 最后一行），每步耗时 1.0 秒。大型模型训练了 300000 步，耗时 3.5 天。

5.3 Optimizer

5.3 优化器

We used the Adam optimizer [20] with $\beta_1 = 0.9$, $\beta_2 = 0.98$ and $\epsilon = 10^{-9}$. We varied the learning rate over the course of training, according to the formula:

$$\mathit{lrate} = d_{\text{model}}^{-0.5} \cdot \min\left(\mathit{step\_num}^{-0.5},\mathit{step\_num} \cdot \mathit{warmup\_steps}^{-1.5}\right) \tag{3}$$

我们使用 Adam 优化器 [20]，参数为 $\beta_1 = 0.9$、$\beta_2 = 0.98$ 和 $\epsilon = 10^{-9}$。训练期间，我们按以下公式调整学习率：

This corresponds to increasing the learning rate linearly for the first $\mathit{warmup\_steps}$ training steps, and decreasing it thereafter proportionally to the inverse square root of the step number. We used $\mathit{warmup\_steps} = 4000$.

这相当于在最初的 $\mathit{warmup\_steps}$ 个训练步骤中线性提高学习率，此后再按训练步数平方根的倒数成比例降低学习率。我们设置 $\mathit{warmup\_steps} = 4000$。

5.4 Regularization

5.4 正则化

We employ three types of regularization during training:

训练期间，我们采用三种正则化方法：

Table 2: The Transformer achieves better BLEU scores than previous state-of-the-art models on the English-to-German and English-to-French newstest2014 tests at a fraction of the training cost.

表 2：在英译德和英译法的 newstest2014 测试中，Transformer 以远低于此前最先进模型的训练成本，取得了更高的 BLEU 分数。

Residual Dropout We apply dropout [33] to the output of each sub-layer, before it is added to the sub-layer input and normalized. In addition, we apply dropout to the sums of the embeddings and the positional encodings in both the encoder and decoder stacks. For the base model, we use a rate of $P_{\text{drop}} = 0.1$.

残差 Dropout 我们对每个子层的输出施加 dropout [33]，然后再将其与该子层的输入相加并进行归一化。此外，在编码器和解码器堆叠中，我们都对嵌入与位置编码相加后的结果施加 dropout。对于基础模型，我们使用的 dropout 概率为 $P_{\text{drop}} = 0.1$。

Label Smoothing During training, we employed label smoothing of value $\epsilon_{\text{ls}} = 0.1$ [36]. This hurts perplexity, as the model learns to be more unsure, but improves accuracy and BLEU score.

标签平滑 训练期间，我们采用取值为 $\epsilon_{\text{ls}} = 0.1$ 的标签平滑 [36]。由于模型学会了降低预测的确定性，这会使困惑度变差，但能提高准确率和 BLEU 分数。

6 Results

6 结果

6.1 Machine Translation

6.1 机器翻译

On the WMT 2014 English-to-German translation task, the big transformer model (Transformer (big) in Table 2) outperforms the best previously reported models (including ensembles) by more than 2.0 BLEU, establishing a new state-of-the-art BLEU score of 28.4. The configuration of this model is listed in the bottom line of Table 3. Training took 3.5 days on 8 P100 GPUs. Even our base model surpasses all previously published models and ensembles, at a fraction of the training cost of any of the competitive models.

在 WMT 2014 英译德任务中，大型 Transformer 模型（表 2 中的 Transformer (big)）比此前报道的最佳模型（包括集成模型）高出超过 2.0 个 BLEU 分数，取得了 28.4 BLEU 的新最佳成绩。该模型的配置列于表 3 的最后一行。训练使用 8 块 P100 GPU，耗时 3.5 天。即便是我们的基础模型，也超过了此前发表的所有模型和集成模型，而训练成本仅为任一有竞争力模型的一小部分。

On the WMT 2014 English-to-French translation task, our big model achieves a BLEU score of 41.0, outperforming all of the previously published single models, at less than 1/4 the training cost of the previous state-of-the-art model. The Transformer (big) model trained for English-to-French used dropout rate $P_{\text{drop}} = 0.1$, instead of 0.3.

在 WMT 2014 英译法任务中，我们的大型模型取得了 41.0 的 BLEU 分数，超过此前发表的所有单模型，而训练成本不到此前最先进模型的四分之一。用于英译法训练的 Transformer (big) 模型采用的 dropout 概率为 $P_{\text{drop}} = 0.1$，而不是 0.3。

For the base models, we used a single model obtained by averaging the last 5 checkpoints, which were written at 10-minute intervals. For the big models, we averaged the last 20 checkpoints. We used beam search with a beam size of 4 and length penalty $\alpha = 0.6$ [38]. These hyperparameters were chosen after experimentation on the development set. We set the maximum output length during inference to input length + 50, but terminate early when possible [38].

对于基础模型，我们将最后 5 个检查点的参数取平均，得到一个单模型；这些检查点每隔 10 分钟保存一次。对于大型模型，我们对最后 20 个检查点取平均。我们使用束宽为 4、长度惩罚 $\alpha = 0.6$ 的束搜索 [38]。这些超参数是在开发集上进行实验后选定的。推理时，我们将最大输出长度设置为输入长度加 50，但在可以结束时提前终止生成 [38]。

Table 2 summarizes our results and compares our translation quality and training costs to other model architectures from the literature. We estimate the number of floating point operations used to train a model by multiplying the training time, the number of GPUs used, and an estimate of the sustained single-precision floating-point capacity of each GPU [Footnote 5].

表 2 汇总了我们的结果，并将我们的翻译质量和训练成本与文献中的其他模型架构进行了比较。我们通过将训练时间、使用的 GPU 数量，以及每块 GPU 持续单精度浮点运算能力的估计值相乘，来估算训练一个模型所需的浮点运算次数 [脚注 5]。

Footnote 5. We used values of 2.8, 3.7, 6.0 and 9.5 TFLOPS for K80, K40, M40 and P100, respectively.

脚注 5。 对于 K80、K40、M40 和 P100，我们采用的数值分别为 2.8、3.7、6.0 和 9.5 TFLOPS。

6.2 Model Variations

6.2 模型变体

To evaluate the importance of different components of the Transformer, we varied our base model in different ways, measuring the change in performance on English-to-German translation on the development set, newstest2013. We used beam search as described in the previous section, but no checkpoint averaging. We present these results in Table 3.

为了评估 Transformer 各个组件的重要性，我们以不同方式调整基础模型，并在开发集 newstest2013 上测量英译德性能的变化。我们采用了上一节所述的束搜索，但未进行检查点平均。结果见表 3。

Table 3: Variations on the Transformer architecture. Unlisted values are identical to those of the base model. All metrics are on the English-to-German translation development set, newstest2013. Listed perplexities are per-wordpiece, according to our byte-pair encoding, and should not be compared to per-word perplexities.

表 3：Transformer 架构的变体。未列出的数值与基础模型相同。所有指标均在英译德开发集 newstest2013 上测得。表中的困惑度按我们使用的字节对编码所划分的词片段计算，不应与按词计算的困惑度进行比较。

In Table 3 rows (A), we vary the number of attention heads and the attention key and value dimensions, keeping the amount of computation constant, as described in Section 3.2.2. While single-head attention is 0.9 BLEU worse than the best setting, quality also drops off with too many heads.

在表 3 的 (A) 组行中，我们按照第 3.2.2 节所述的方法，在保持计算量不变的情况下，调整注意力头的数量以及注意力键和值的维度。单头注意力的 BLEU 比最佳设置低 0.9，而头数过多也会导致质量下降。

In Table 3 rows (B), we observe that reducing the attention key size $d_k$ hurts model quality. This suggests that determining compatibility is not easy and that a more sophisticated compatibility function than dot product may be beneficial. We further observe in rows (C) and (D) that, as expected, bigger models are better, and dropout is very helpful in avoiding over-fitting. In row (E) we replace our sinusoidal positional encoding with learned positional embeddings [9], and observe nearly identical results to the base model.

在表 3 的 (B) 组行中，我们观察到，减小注意力键的维度 $d_k$ 会损害模型质量。这表明确定匹配程度并不容易，使用比点积更复杂的匹配函数可能有所帮助。我们还在 (C) 和 (D) 组行中观察到，与预期一致，更大的模型表现更好，而 dropout 对避免过拟合很有帮助。在 (E) 行中，我们用学习得到的位置嵌入 [9] 替换正弦位置编码，得到的结果与基础模型几乎相同。

6.3 English Constituency Parsing

6.3 英语成分句法分析

To evaluate if the Transformer can generalize to other tasks we performed experiments on English constituency parsing. This task presents specific challenges: the output is subject to strong structural constraints and is significantly longer than the input. Furthermore, RNN sequence-to-sequence models have not been able to attain state-of-the-art results in small-data regimes [37].

为了评估 Transformer 能否泛化到其他任务，我们进行了英语成分句法分析实验。这项任务有其特有的挑战：输出受到严格的结构约束，而且明显长于输入。此外，RNN 序列到序列模型在小数据量条件下尚未达到最先进的结果 [37]。

We trained a 4-layer transformer with $d_{model} = 1024$ on the Wall Street Journal (WSJ) portion of the Penn Treebank [25], about 40K training sentences. We also trained it in a semi-supervised setting, using the larger high-confidence and BerkleyParser corpora from with approximately 17M sentences [37]. We used a vocabulary of 16K tokens for the WSJ only setting and a vocabulary of 32K tokens for the semi-supervised setting.

我们在 Penn Treebank [25] 的《华尔街日报》（WSJ）部分上训练了一个 $d_{model} = 1024$ 的 4 层 Transformer，训练集约有 4 万个句子。我们还在半监督设置下训练了该模型，使用更大的高置信度语料和 BerkleyParser 语料，总计约 1700 万个句子 [37]。仅使用 WSJ 时，我们采用包含 1.6 万个词元的词表；在半监督设置下，则采用包含 3.2 万个词元的词表。

We performed only a small number of experiments to select the dropout, both attention and residual (section 5.4), learning rates and beam size on the Section 22 development set, all other parameters remained unchanged from the English-to-German base translation model. During inference, we increased the maximum output length to input length + 300. We used a beam size of 21 and $α = 0.3$ for both WSJ only and the semi-supervised setting.

我们仅在第 22 节开发集上进行了少量实验，以选择注意力和残差两处的 dropout（第 5.4 节）、学习率及束宽；其他所有参数均与英译德基础翻译模型保持一致。推理时，我们将最大输出长度增加到输入长度加 300。在仅使用 WSJ 和半监督这两种设置下，束宽均设为 21，$α = 0.3$。

Table 4: The Transformer generalizes well to English constituency parsing (Results are on Section 23 of WSJ)

表 4：Transformer 能够很好地泛化到英语成分句法分析任务（结果在 WSJ 的第 23 节上测得）。

Our results in Table 4 show that despite the lack of task-specific tuning our model performs surprisingly well, yielding better results than all previously reported models with the exception of the Recurrent Neural Network Grammar [8].

表 4 的结果表明，尽管没有针对该任务进行专门调优，我们的模型仍表现得出乎意料地好；除了循环神经网络语法模型 [8] 之外，其结果优于此前所有已报告的模型。

In contrast to RNN sequence-to-sequence models [37], the Transformer outperforms the BerkeleyParser [29] even when training only on the WSJ training set of 40K sentences.

与 RNN 序列到序列模型 [37] 不同，即使仅在包含 4 万个句子的 WSJ 训练集上训练，Transformer 也能超越 BerkeleyParser [29]。

7 Conclusion

7 结论

In this work, we presented the Transformer, the first sequence transduction model based entirely on attention, replacing the recurrent layers most commonly used in encoder-decoder architectures with multi-headed self-attention.

在这项工作中，我们提出了 Transformer，这是首个完全基于注意力的序列转换模型，以多头自注意力取代了编码器—解码器架构中最常用的循环层。

For translation tasks, the Transformer can be trained significantly faster than architectures based on recurrent or convolutional layers. On both WMT 2014 English-to-German and WMT 2014 English-to-French translation tasks, we achieve a new state of the art. In the former task our best model outperforms even all previously reported ensembles.

对于翻译任务，Transformer 的训练速度显著快于基于循环层或卷积层的架构。我们在 WMT 2014 英译德和 WMT 2014 英译法两个翻译任务上均达到了新的最先进水平。在前一个任务中，我们表现最好的模型甚至超越了此前所有已报告的集成模型。

We are excited about the future of attention-based models and plan to apply them to other tasks. We plan to extend the Transformer to problems involving input and output modalities other than text and to investigate local, restricted attention mechanisms to efficiently handle large inputs and outputs such as images, audio and video. Making generation less sequential is another research goals of ours.

我们对基于注意力的模型的未来充满期待，并计划将其应用于其他任务。我们计划将 Transformer 扩展到输入和输出包含文本之外模态的问题，并研究局部、受限的注意力机制，以高效处理图像、音频和视频等大型输入与输出。减少生成过程对顺序执行的依赖，也是我们的另一项研究目标。

The code we used to train and evaluate our models is available at https://github.com/tensorflow/tensor2tensor.

我们用于训练和评估模型的代码可在 https://github.com/tensorflow/tensor2tensor 获取。

Acknowledgements

致谢

We are grateful to Nal Kalchbrenner and Stephan Gouws for their fruitful comments, corrections and inspiration.

感谢 Nal Kalchbrenner 和 Stephan Gouws 提供富有成效的意见、修正和启发。

References

参考文献

[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016.

[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. 层归一化。arXiv preprint arXiv:1607.06450, 2016.

[2] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. CoRR, abs/1409.0473, 2014.

[2] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 通过联合学习对齐与翻译实现神经机器翻译。CoRR, abs/1409.0473, 2014.

[3] Denny Britz, Anna Goldie, Minh-Thang Luong, and Quoc V. Le. Massive exploration of neural machine translation architectures. CoRR, abs/1703.03906, 2017.

[3] Denny Britz, Anna Goldie, Minh-Thang Luong, and Quoc V. Le. 对神经机器翻译架构的大规模探索。CoRR, abs/1703.03906, 2017.

[4] Jianpeng Cheng, Li Dong, and Mirella Lapata. Long short-term memory-networks for machine reading. arXiv preprint arXiv:1601.06733, 2016.

[4] Jianpeng Cheng, Li Dong, and Mirella Lapata. 用于机器阅读的长短期记忆网络。arXiv preprint arXiv:1601.06733, 2016.

[5] Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. CoRR, abs/1406.1078, 2014.

[5] Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 使用 RNN 编码器—解码器学习用于统计机器翻译的短语表示。CoRR, abs/1406.1078, 2014.

[6] Francois Chollet. Xception: Deep learning with depthwise separable convolutions. arXiv preprint arXiv:1610.02357, 2016.

[6] Francois Chollet. Xception：使用深度可分离卷积的深度学习。arXiv preprint arXiv:1610.02357, 2016.

[7] Junyoung Chung, Çaglar Gülçehre, Kyunghyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. CoRR, abs/1412.3555, 2014.

[7] Junyoung Chung, Çaglar Gülçehre, Kyunghyun Cho, and Yoshua Bengio. 门控循环神经网络在序列建模中的实证评估。CoRR, abs/1412.3555, 2014.

[8] Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, and Noah A. Smith. Recurrent neural network grammars. In Proc. of NAACL, 2016.

[8] Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, and Noah A. Smith. 循环神经网络语法。In Proc. of NAACL, 2016.

[9] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N. Dauphin. Convolutional sequence to sequence learning. arXiv preprint arXiv:1705.03122v2, 2017.

[9] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N. Dauphin. 卷积序列到序列学习。arXiv preprint arXiv:1705.03122v2, 2017.

[10] Alex Graves. Generating sequences with recurrent neural networks. arXiv preprint arXiv:1308.0850, 2013.

[10] Alex Graves. 使用循环神经网络生成序列。arXiv preprint arXiv:1308.0850, 2013.

[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770–778, 2016.

[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 用于图像识别的深度残差学习。In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770–778, 2016.

[12] Sepp Hochreiter, Yoshua Bengio, Paolo Frasconi, and Jürgen Schmidhuber. Gradient flow in recurrent nets: the difficulty of learning long-term dependencies, 2001.

[12] Sepp Hochreiter, Yoshua Bengio, Paolo Frasconi, and Jürgen Schmidhuber. 循环网络中的梯度流：学习长期依赖的困难，2001。

[13] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735–1780, 1997.

[13] Sepp Hochreiter and Jürgen Schmidhuber. 长短期记忆。Neural computation, 9(8):1735–1780, 1997.

[14] Zhongqiang Huang and Mary Harper. Self-training PCFG grammars with latent annotations across languages. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 832–841. ACL, August 2009.

[14] Zhongqiang Huang and Mary Harper. 对带有隐式标注的 PCFG 语法进行跨语言自训练。In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 832–841. ACL, August 2009.

[15] Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. Exploring the limits of language modeling. arXiv preprint arXiv:1602.02410, 2016.

[15] Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. 探索语言建模的极限。arXiv preprint arXiv:1602.02410, 2016.

[16] Łukasz Kaiser and Samy Bengio. Can active memory replace attention? In Advances in Neural Information Processing Systems, (NIPS), 2016.

[16] Łukasz Kaiser and Samy Bengio. 主动记忆能否取代注意力？In Advances in Neural Information Processing Systems, (NIPS), 2016.

[17] Łukasz Kaiser and Ilya Sutskever. Neural GPUs learn algorithms. In International Conference on Learning Representations (ICLR), 2016.

[17] Łukasz Kaiser and Ilya Sutskever. Neural GPU 学习算法。In International Conference on Learning Representations (ICLR), 2016.

[18] Nal Kalchbrenner, Lasse Espeholt, Karen Simonyan, Aaron van den Oord, Alex Graves, and Koray Kavukcuoglu. Neural machine translation in linear time. arXiv preprint arXiv:1610.10099v2, 2017.

[18] Nal Kalchbrenner, Lasse Espeholt, Karen Simonyan, Aaron van den Oord, Alex Graves, and Koray Kavukcuoglu. 线性时间的神经机器翻译。arXiv preprint arXiv:1610.10099v2, 2017.

[19] Yoon Kim, Carl Denton, Luong Hoang, and Alexander M. Rush. Structured attention networks. In International Conference on Learning Representations, 2017.

[19] Yoon Kim, Carl Denton, Luong Hoang, and Alexander M. Rush. 结构化注意力网络。In International Conference on Learning Representations, 2017.

[20] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015.

[20] Diederik Kingma and Jimmy Ba. Adam：一种随机优化方法。In ICLR, 2015.

[21] Oleksii Kuchaiev and Boris Ginsburg. Factorization tricks for LSTM networks. arXiv preprint arXiv:1703.10722, 2017.

[21] Oleksii Kuchaiev and Boris Ginsburg. LSTM 网络的分解技巧。arXiv preprint arXiv:1703.10722, 2017.

[22] Zhouhan Lin, Minwei Feng, Cicero Nogueira dos Santos, Mo Yu, Bing Xiang, Bowen Zhou, and Yoshua Bengio. A structured self-attentive sentence embedding. arXiv preprint arXiv:1703.03130, 2017.

[22] Zhouhan Lin, Minwei Feng, Cicero Nogueira dos Santos, Mo Yu, Bing Xiang, Bowen Zhou, and Yoshua Bengio. 一种结构化自注意力句子嵌入。arXiv preprint arXiv:1703.03130, 2017.

[23] Minh-Thang Luong, Quoc V. Le, Ilya Sutskever, Oriol Vinyals, and Lukasz Kaiser. Multi-task sequence to sequence learning. arXiv preprint arXiv:1511.06114, 2015.

[23] Minh-Thang Luong, Quoc V. Le, Ilya Sutskever, Oriol Vinyals, and Lukasz Kaiser. 多任务序列到序列学习。arXiv preprint arXiv:1511.06114, 2015.

[24] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-based neural machine translation. arXiv preprint arXiv:1508.04025, 2015.

[24] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. 基于注意力的神经机器翻译的有效方法。arXiv preprint arXiv:1508.04025, 2015.

[25] Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. Building a large annotated corpus of english: The penn treebank. Computational linguistics, 19(2):313–330, 1993.

[25] Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. 构建大型英语标注语料库：Penn Treebank。Computational linguistics, 19(2):313–330, 1993.

[26] David McClosky, Eugene Charniak, and Mark Johnson. Effective self-training for parsing. In Proceedings of the Human Language Technology Conference of the NAACL, Main Conference, pages 152–159. ACL, June 2006.

[26] David McClosky, Eugene Charniak, and Mark Johnson. 用于句法分析的有效自训练。In Proceedings of the Human Language Technology Conference of the NAACL, Main Conference, pages 152–159. ACL, June 2006.

[27] Ankur Parikh, Oscar Täckström, Dipanjan Das, and Jakob Uszkoreit. A decomposable attention model. In Empirical Methods in Natural Language Processing, 2016.

[27] Ankur Parikh, Oscar Täckström, Dipanjan Das, and Jakob Uszkoreit. 一种可分解的注意力模型。In Empirical Methods in Natural Language Processing, 2016.

[28] Romain Paulus, Caiming Xiong, and Richard Socher. A deep reinforced model for abstractive summarization. arXiv preprint arXiv:1705.04304, 2017.

[28] Romain Paulus, Caiming Xiong, and Richard Socher. 用于生成式摘要的深度强化学习模型。arXiv preprint arXiv:1705.04304, 2017.

[29] Slav Petrov, Leon Barrett, Romain Thibaux, and Dan Klein. Learning accurate, compact, and interpretable tree annotation. In Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the ACL, pages 433–440. ACL, July 2006.

[29] Slav Petrov, Leon Barrett, Romain Thibaux, and Dan Klein. 学习准确、紧凑且可解释的树标注。In Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the ACL, pages 433–440. ACL, July 2006.

[30] Ofir Press and Lior Wolf. Using the output embedding to improve language models. arXiv preprint arXiv:1608.05859, 2016.

[30] Ofir Press and Lior Wolf. 使用输出嵌入改进语言模型。arXiv preprint arXiv:1608.05859, 2016.

[31] Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909, 2015.

[31] Rico Sennrich, Barry Haddow, and Alexandra Birch. 使用子词单元进行罕见词的神经机器翻译。arXiv preprint arXiv:1508.07909, 2015.

[32] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017.

[32] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. 极其庞大的神经网络：稀疏门控混合专家层。arXiv preprint arXiv:1701.06538, 2017.

[33] Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(1):1929–1958, 2014.

[33] Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout：一种防止神经网络过拟合的简单方法。Journal of Machine Learning Research, 15(1):1929–1958, 2014.

[34] Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. End-to-end memory networks. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems 28, pages 2440–2448. Curran Associates, Inc., 2015.

[34] Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. 端到端记忆网络。In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems 28, pages 2440–2448. Curran Associates, Inc., 2015.

[35] Ilya Sutskever, Oriol Vinyals, and Quoc VV Le. Sequence to sequence learning with neural networks. In Advances in Neural Information Processing Systems, pages 3104–3112, 2014.

[35] Ilya Sutskever, Oriol Vinyals, and Quoc VV Le. 使用神经网络进行序列到序列学习。In Advances in Neural Information Processing Systems, pages 3104–3112, 2014.

[36] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. CoRR, abs/1512.00567, 2015.

[36] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. 重新思考用于计算机视觉的 Inception 架构。CoRR, abs/1512.00567, 2015.

[37] Vinyals & Kaiser, Koo, Petrov, Sutskever, and Hinton. Grammar as a foreign language. In Advances in Neural Information Processing Systems, 2015.

[37] Vinyals & Kaiser, Koo, Petrov, Sutskever, and Hinton. 将语法视为一门外语。In Advances in Neural Information Processing Systems, 2015.

[38] Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google’s neural machine translation system: Bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144, 2016.

[38] Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google 的神经机器翻译系统：缩小人工翻译与机器翻译之间的差距。arXiv preprint arXiv:1609.08144, 2016.

[39] Jie Zhou, Ying Cao, Xuguang Wang, Peng Li, and Wei Xu. Deep recurrent models with fast-forward connections for neural machine translation. CoRR, abs/1606.04199, 2016.

[39] Jie Zhou, Ying Cao, Xuguang Wang, Peng Li, and Wei Xu. 用于神经机器翻译的带有快速前馈连接的深度循环模型。CoRR, abs/1606.04199, 2016.

[40] Muhua Zhu, Yue Zhang, Wenliang Chen, Min Zhang, and Jingbo Zhu. Fast and accurate shift-reduce constituent parsing. In Proceedings of the 51st Annual Meeting of the ACL (Volume 1: Long Papers), pages 434–443. ACL, August 2013.

[40] Muhua Zhu, Yue Zhang, Wenliang Chen, Min Zhang, and Jingbo Zhu. 快速且准确的移进—归约成分句法分析。In Proceedings of the 51st Annual Meeting of the ACL (Volume 1: Long Papers), pages 434–443. ACL, August 2013.

Attention Visualizations

注意力可视化

Figure 3: An example of the attention mechanism following long-distance dependencies in the encoder self-attention in layer 5 of 6. Many of the attention heads attend to a distant dependency of the verb ‘making’, completing the phrase ‘making...more difficult’. Attentions here shown only for the word ‘making’. Different colors represent different heads. Best viewed in color.

图 3：编码器的 6 层中的第 5 层自注意力中，注意力机制追踪长距离依赖关系的一个例子。许多注意力头关注动词“making”的一个远距离依赖项，补全了“making...more difficult”（使……更加困难）这一短语。此处只展示单词“making”的注意力。不同颜色表示不同的头。以彩色查看效果最佳。

Figure 4: Two attention heads, also in layer 5 of 6, apparently involved in anaphora resolution. Top: Full attentions for head 5. Bottom: Isolated attentions from just the word ‘its’ for attention heads 5 and 6. Note that the attentions are very sharp for this word.

图 4：同样位于 6 层中的第 5 层的两个注意力头，它们似乎参与了回指消解。上图：第 5 个头的完整注意力。下图：单独展示第 5 个和第 6 个注意力头中，来自单词“its”的注意力。请注意，该词的注意力分布非常集中。

Figure 5: Many of the attention heads exhibit behaviour that seems related to the structure of the sentence. We give two such examples above, from two different heads from the encoder self-attention at layer 5 of 6. The heads clearly learned to perform different tasks.

图 5：许多注意力头展现出似乎与句子结构有关的行为。上面给出了两个这样的例子，分别来自编码器的 6 层中的第 5 层自注意力的两个不同头。这些头显然学会了执行不同的任务。