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.
只要给予适当的出处，Google 特此授权复制本文中的表格和图表，且仅限于新闻报道或学术研究中使用。

Attention Is All You Need
注意力机制就是全部所需

\ANDAshish Vaswani
Google Brain
avaswani@google.com
&Noam Shazeer¹¹footnotemark: 1
Google Brain
noam@google.com
&Niki Parmar¹¹footnotemark: 1
Google Research
nikip@google.com
&Jakob Uszkoreit¹¹footnotemark: 1
Google Research
usz@google.com
&Llion Jones¹¹footnotemark: 1
Google Research
llion@google.com
&Aidan N. Gomez¹¹footnotemark: 1
University of Toronto
aidan@cs.toronto.edu &Łukasz Kaiser¹¹footnotemark: 1
Google Brain
lukaszkaiser@google.com
&Illia Polosukhin¹¹footnotemark: 1
illia.polosukhin@gmail.com
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. Work performed while at Google Brain.Work performed while at Google Research.

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通常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结构。表现最优的模型还通过注意力机制将编码器和解码器连接起来。我们提出了一种全新的简单网络架构——Transformer，它完全基于注意力机制，彻底摒弃了循环和卷积操作。在两个机器翻译任务上的实验表明，该模型不仅在翻译质量上表现更优，而且更易于并行处理，训练时间大幅缩短。在 WMT 2014 英德翻译任务中，我们的模型取得了 28.4 的 BLEU 分数，较现有最佳结果（包括集成模型）提升了超过 2 分。在 WMT 2014 英法翻译任务中，经过在八块 GPU 上训练 3.5 天，我们的模型创下了单模型 41.8 的最新 BLEU 记录，训练成本仅为文献中最佳模型的一小部分。我们还证明了 Transformer 在其他任务中的良好泛化能力，成功应用于英语成分句法分析，无论是大规模还是有限的训练数据环境。

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，特别是长短期记忆网络[13]和门控循环神经网络[7]，已被公认为序列建模和转导任务（如语言建模和机器翻译[35, 2, 5]）的最先进方法。此后，许多研究不断推动循环语言模型和编码器-解码器架构的发展和创新[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.
循环模型通常根据输入和输出序列中符号的位置来分步计算。它们将位置对应到计算时间的步骤，生成一系列隐藏状态 $h_{t}$ ，这些状态是前一隐藏状态 $h_{t-1}$ 和当前位置输入 $t$ 的函数。这种固有的顺序特性使得训练过程中无法在单个样本内实现并行化，尤其在序列较长时更为关键，因为内存限制了跨样本的批量处理。近期研究通过分解技巧[21]和条件计算[32]显著提升了计算效率，同时后者还改善了模型性能。但顺序计算的根本限制依然存在。

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.
注意力机制已成为多种任务中强大序列建模和转换模型的重要组成部分，能够在输入或输出序列中无视距离地建模依赖关系[2, 19]。不过，除了少数例外[27]，这种注意力机制通常是与循环网络一起使用的。

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.
本文提出了 Transformer 模型架构，它摒弃了循环结构，完全依靠注意力机制来捕捉输入与输出之间的全局依赖。Transformer 大幅提升了并行处理能力，仅用八块 P100 GPU 训练十二小时，就能达到翻译质量的新突破。

2 Background 2 背景介绍

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.
减少顺序计算的目标也是扩展神经 GPU [16]、ByteNet [18] 和 ConvS2S [9] 的设计基础，这些模型均采用卷积神经网络作为基本构建模块，能够对所有输入和输出位置并行计算隐藏表示。在这些模型中，连接任意两个输入或输出位置的信号所需的操作次数会随着它们之间距离的增加而增长：ConvS2S 是线性增长，ByteNet 是对数增长，这使得学习远距离位置间的依赖关系变得更加困难 [12]。而在 Transformer 中，这一操作次数被降低为常数，虽然由于对注意力加权位置的平均处理导致有效分辨率有所降低，但我们通过第 3.2 节介绍的多头注意力机制来弥补这一缺陷。

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（有时称为内部注意力）是一种通过关联序列中不同位置来计算序列表示的注意力机制。自注意力已在多种任务中取得成功应用，如阅读理解、抽象摘要、文本蕴含以及学习与具体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，而非传统的序列对齐递归，已被证明在简单语言的问答和语言建模任务中表现出色[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].
据我们所知，Transformer 是首个完全依靠自注意力机制来计算输入和输出表示的转导模型，无需使用序列对齐的循环神经网络或卷积。在接下来的章节中，我们将介绍 Transformer，解释自注意力的原理，并探讨它相较于[17, 18]和[9]等模型的优势。

3 Model Architecture 3 模型结构

Refer to caption — Figure 1: The Transformer - model architecture.
图 1：Transformer 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.
大多数先进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，编码器将输入的符号表示序列 $(x_{1},...,x_{n})$ 转换为连续的表示序列 $\mathbf{z}=(z_{1},...,z_{n})$ 。在给定 $\mathbf{z}$ 的情况下，解码器逐步生成输出符号序列 $(y_{1},...,y_{m})$ ，每次生成一个符号。模型采用自回归方式[10]，在生成下一个符号时会利用之前生成的符号作为额外输入。

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 采用这种整体架构，编码器和解码器均由堆叠的自注意力机制和逐点全连接层组成，分别展示在图 1 的左半部分和右半部分。

3.1 Encoder and Decoder Stacks
3.1 编码器与解码器结构层叠

Encoder: 编码器：

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 $\mathrm{LayerNorm}(x+\mathrm{Sublayer}(x))$ , where $\mathrm{Sublayer}(x)$ 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 $d_{\text{model}}=512$ .
编码器由 $N=6$ 个相同的层堆叠而成。每一层包含两个子层：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。我们在每个子层外部采用残差连接[11]，随后进行层归一化[1]。换句话说，每个子层的输出为 $\mathrm{LayerNorm}(x+\mathrm{Sublayer}(x))$ ，其中 $\mathrm{Sublayer}(x)$ 表示该子层自身实现的函数。为了支持这些残差连接，模型中所有子层及嵌入层的输出维度均为 $d_{\text{model}}=512$ 。

Decoder: 解码器：

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 $i$ .
解码器同样由 $N=6$ 个相同的层堆叠而成。除了每个编码器层中的两个子层外，解码器还增加了第三个子层，用于对编码器堆栈的输出进行多头注意力操作。与编码器类似，我们在每个子层周围使用残差连接，随后进行层归一化。同时，我们对解码器中的自注意力子层进行了修改，防止当前位置关注后续位置。通过这种掩码机制以及输出嵌入向后偏移一个位置，确保位置 $i$ 的预测仅依赖于位置小于 $i$ 的已知输出。

3.2 Attention 3.2 注意力机制

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.
注意力函数可以理解为将一个查询和一组键值对映射为一个输出向量，其中查询、键、值以及输出都是向量。输出是值的加权和，权重由查询与对应键之间的兼容性函数计算得出。

3.2.1 Scaled Dot-Product Attention
3.2.1 缩放点积注意力机制

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.
我们将这种特定的注意力机制称为“缩放点积注意力”（见图 2）。输入包括维度为 $d_{k}$ 的查询和键，以及维度为 $d_{v}$ 的值。我们先计算查询与所有键的点积，再将结果除以 $\sqrt{d_{k}}$ ，最后通过 softmax 函数得到对应值的权重。

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:
在实际操作中，我们会将一组查询同时打包成一个矩阵 $Q$ ，并计算其注意力函数。键和值同样被分别打包成矩阵 $K$ 和 $V$ 。然后，我们计算输出矩阵如下：

\mathrm{Attention}(Q,K,V)=\mathrm{softmax}(\frac{QK^{T}}{\sqrt{d_{k}}})V

(1)

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.
两种最常用的注意力函数是加性注意力[2]和点积（乘法）注意力。点积注意力与我们的算法基本相同，只是在缩放因子上有所不同，为 $\frac{1}{\sqrt{d_{k}}}$ 。加性注意力通过一个带有单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，但点积注意力在实际应用中速度更快且更节省内存，因为它可以利用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 ¹¹1To 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 $d_{k}$ .. To counteract this effect, we scale the dot products by $\frac{1}{\sqrt{d_{k}}}$ .
虽然在较小的 $d_{k}$ 值时，这两种机制表现相近，但在较大的 $d_{k}$ 值下，加性注意力优于未缩放的点积注意力[3]。我们推测，当 $d_{k}$ 值较大时，点积的数值会变得很大，使得 softmax 函数进入梯度极其微小的区域 ¹ 。为了解决这个问题，我们对点积进行了乘以 $\frac{1}{\sqrt{d_{k}}}$ 的缩放。

3.2.2 Multi-Head Attention
3.2.2 多头注意力机制

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.
我们没有只执行一次使用 $d_{\text{model}}$ 维键、值和查询的注意力函数，而是发现将查询、键和值分别通过 $h$ 个不同的、学习得到的线性投影，投影到 $d_{k}$ 、 $d_{k}$ 和 $d_{v}$ 维度更为有效。然后，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 $d_{v}$ 维的输出值。接着将这些输出值拼接起来，再次进行线性投影，得到最终的结果，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.
多头注意力机制使模型能够同时关注不同位置上来自不同表示子空间的信息，而单一注意力头的平均操作则会抑制这种能力。

	$\displaystyle\mathrm{MultiHead}(Q,K,V)$	$\displaystyle=\mathrm{Concat}(\mathrm{head_{1}},...,\mathrm{head_{h}})W^{O}$
	$\displaystyle\text{where}~{}\mathrm{head_{i}}$	$\displaystyle=\mathrm{Attention}(QW^{Q}_{i},KW^{K}_{i},VW^{V}_{i})$

Where the projections are parameter matrices $W^{Q}_{i}\in\mathbb{R}^{d_{\text{model}}\times d_{k}}$ , $W^{K}_{i}\in\mathbb{R}^{d_{\text{model}}\times d_{k}}$ , $W^{V}_{i}\in\mathbb{R}^{d_{\text{model}}\times d_{v}}$ and $W^{O}\in\mathbb{R}^{hd_{v}\times d_{\text{model}}}$ .
这些投影是由参数矩阵 $W^{Q}_{i}\in\mathbb{R}^{d_{\text{model}}\times d_{k}}$ 、 $W^{K}_{i}\in\mathbb{R}^{d_{\text{model}}\times d_{k}}$ 、 $W^{V}_{i}\in\mathbb{R}^{d_{\text{model}}\times d_{v}}$ 和 $W^{O}\in\mathbb{R}^{hd_{v}\times d_{\text{model}}}$ 构成的。

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，我们采用了 $h=8$ 个并行的注意力层（也称为头）。每个头都使用了 $d_{k}=d_{v}=d_{\text{model}}/h=64$ 。由于每个头的维度较小，整体计算成本与使用全维度的单头注意力相当。

3.2.3 Applications of Attention in our Model
3.2.3 我们模型中注意力机制的应用场景

The Transformer uses multi-head attention in three different ways:
Transformer 在三种不同的场景中使用多头注意力机制：

•

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，查询来自上一层解码器，记忆键和值则来自编码器的输出。这使得解码器的每个位置都能关注输入序列中的所有位置，类似于序列到序列模型中常见的编码器-解码器注意力机制，如文献[38, 2, 9]所示。
•

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.

编码器由自注意力层组成。在自注意力层中，所有的键、值和查询都来自同一来源，这里指的是编码器上一层的输出。编码器中的每个位置都能关注上一层编码器的所有位置。
•

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.

• 同样，解码器中的自注意力层允许解码器的每个位置关注该位置及之前的所有位置。为了保持自回归特性，我们需要阻止解码器中信息向左流动。我们通过在缩放点积注意力机制中，将所有对应非法连接的 softmax 输入值屏蔽（设为 $-\infty$ ）来实现这一点。详见图 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.
除了注意力子层外，我们的编码器和解码器的每一层还包含一个全连接的前馈网络，该网络对每个位置独立且相同地应用。该网络由两个线性变换组成，中间夹有一个 ReLU 激活函数。

\mathrm{FFN}(x)=\max(0,xW_{1}+b_{1})W_{2}+b_{2}

(2)

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 $d_{ff}=2048$ .
虽然线性变换在不同位置保持一致，但每一层使用的参数各不相同。换句话说，这相当于两个卷积核大小为 1 的卷积操作。输入和输出的维度为 $d_{\text{model}}=512$ ，中间层的维度为 $d_{ff}=2048$ 。

3.4 Embeddings and Softmax
3.4 嵌入层与 Softmax 层

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，我们使用学习到的嵌入将输入和输出的标记转换为维度为 $d_{\text{model}}$ 的向量。同时，我们采用常用的学习线性变换和 softmax 函数，将解码器的输出转换为预测的下一个标记的概率。在我们的模型中，两个嵌入层和 softmax 前的线性变换共享同一权重矩阵，这与文献[30]的方法类似。在嵌入层中，我们将这些权重乘以 $\sqrt{d_{\text{model}}}$ 。

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].
由于我们的模型不使用循环和卷积，为了让模型能够利用序列中标记的顺序信息，我们需要注入一些关于标记相对或绝对位置的信息。为此，我们在编码器和解码器堆栈底部的输入嵌入中加入了“位置编码”。位置编码与嵌入的维度相同（ $d_{\text{model}}$ ），这样两者就可以相加。位置编码有多种方式，包括可学习的和固定的[9]。

In this work, we use sine and cosine functions of different frequencies:
在本研究中，我们采用了不同频率的正弦和余弦函数：

	$\displaystyle PE_{(pos,2i)}=sin(pos/10000^{2i/d_{\text{model}}})$
	$\displaystyle PE_{(pos,2i+1)}=cos(pos/10000^{2i/d_{\text{model}}})$

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对应一个正弦函数。波长从 $2\pi$ 到 $10000\cdot 2\pi$ 以几何级数递增。我们选择这个函数，是基于假设模型能够通过相对位置轻松学习注意力机制，因为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.
我们还尝试了使用学习到的位置嵌入[9]，结果发现两种方法的效果几乎相同（见表 3 第(E)行）。我们最终选择了正弦位置编码，因为它可能使模型能够推广到比训练时更长的序列长度。

4 Why Self-Attention 为什么选择自注意力机制

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 $(x_{1},...,x_{n})$ to another sequence of equal length $(z_{1},...,z_{n})$ , 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.
本节中，我们将自注意力层与常用于将一个可变长度的符号表示序列 $(x_{1},...,x_{n})$ 映射到另一个等长序列 $(z_{1},...,z_{n})$ 的循环层和卷积层进行比较，比如典型序列转换编码器或解码器中的隐藏层。为了说明采用自注意力的原因，我们提出了三个关键需求。

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.
一个是每层的总计算复杂度，另一个是可并行计算的程度，通常通过所需的最少顺序操作次数来衡量。

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.
第三个因素是网络中长距离依赖之间的路径长度。学习长距离依赖是许多序列转换任务中的一个重要挑战。影响学习这种依赖能力的关键因素之一，是信号在网络中前向和后向传播所需经过的路径长度。输入和输出序列中任意位置之间的路径越短，学习长距离依赖就越容易[12]。因此，我们还比较了由不同层类型构成的网络中，任意输入和输出位置之间的最大路径长度。

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：不同层类型的最大路径长度、每层计算复杂度及最少顺序操作次数。这里，

n

表示序列长度，

d

表示表示维度，

k

表示卷积核大小，

r

表示受限自注意力中的邻域大小。

Layer Type 网络层类型	Complexity per Layer 每一层的计算复杂度	Sequential 顺序的	Maximum Path Length 最大路径长度
		Operations 运营
Self-Attention 自我注意力机制	$O(n^{2}\cdot d)$	$O(1)$	$O(1)$
Recurrent 循环神经网络	$O(n\cdot d^{2})$	$O(n)$	$O(n)$
Convolutional 卷积层	$O(k\cdot n\cdot d^{2})$	$O(1)$	$O(log_{k}(n))$
Self-Attention (restricted) 限制性自注意力	$O(r\cdot n\cdot d)$	$O(1)$	$O(n/r)$

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，自注意力层通过固定数量的顺序操作连接所有位置，而循环层则需要 $O(n)$ 个顺序操作。从计算复杂度来看，当序列长度 $n$ 小于表示维度 $d$ 时，自注意力层的速度快于循环层，这种情况在当前机器翻译中常用的句子表示方法（如 word-piece [38]和 byte-pair [31]）中较为普遍。为了提升处理超长序列任务的计算效率，自注意力机制可以限制为只考虑输入序列中以对应输出位置为中心、大小为 $r$ 的邻域范围，这样最大路径长度将增加到 $O(n/r)$ 。我们计划在未来的工作中对这一方法进行深入研究。

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.
单个卷积核宽度为 $k<n$ 的卷积层无法连接所有输入和输出位置对。要实现这一点，连续卷积核情况下需要堆叠 $O(n/k)$ 层卷积，而扩张卷积情况下则需 $O(log_{k}(n))$ 层[18]，这会增加网络中任意两点间最长路径的长度。卷积层通常比循环层计算成本高出 $k$ 倍。但可分离卷积[6]大幅降低了复杂度，降至 $O(k\cdot n\cdot d+n\cdot d^{2})$ 。即便如此，可分离卷积的复杂度仍相当于自注意力层与逐点前馈层的组合，这正是我们模型采用的方案。

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.
作为额外的好处，自注意力机制能够使模型更具可解释性。我们对模型中的注意力分布进行了分析，并在附录中展示和讨论了相关示例。不仅各个注意力头能够清晰地学习不同的任务，许多注意力头的行为还与句子的句法和语义结构密切相关。

5 Training 第 5 章训练

This section describes the training regime for our models.
本节介绍了我们模型的训练方法。

5.1 Training Data and Batching
5.1 训练数据与批量处理

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 英德数据集上进行了训练，该数据集包含约 450 万对句子。句子采用字节对编码（BPE）[3]，该编码在源语言和目标语言之间共享约 37000 个词汇。对于英法任务，我们使用了规模更大的 WMT 2014 英法数据集，包含 3600 万句子，并将标记拆分为包含 32000 个词片的词汇表[38]。句子对根据近似序列长度进行分批，每个训练批次大约包含 25000 个源语言标记和 25000 个目标语言标记。

5.2 Hardware and Schedule 5.2 硬件配置与进度安排

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).
我们在一台配备 8 块 NVIDIA P100 GPU 的机器上训练了模型。对于使用论文中所述超参数的基础模型，每个训练步骤大约耗时 0.4 秒。基础模型总共训练了 100,000 步，约 12 小时。对于大型模型（见表 3 底行），每步耗时约 1.0 秒。大型模型训练了 300,000 步，耗时约 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:
我们采用了 Adam 优化器 [20]，其参数为 $\beta_{1}=0.9$ 、 $\beta_{2}=0.98$ 和 $\epsilon=10^{-9}$ 。在训练过程中，我们根据以下公式动态调整学习率：

lrate=d_{\text{model}}^{-0.5}\cdot\min({step\_num}^{-0.5},{step\_num}\cdot{% warmup\_steps}^{-1.5})

(3)

This corresponds to increasing the learning rate linearly for the first $warmup\_steps$ training steps, and decreasing it thereafter proportionally to the inverse square root of the step number. We used $warmup\_steps=4000$ .
这意味着在前 $warmup\_steps$ 个训练步骤中，学习率会线性增加，之后则根据步骤数的平方根倒数进行递减。我们采用了 $warmup\_steps=4000$ 。

5.4 Regularization 5.4 正则化方法

We employ three types of regularization during training:
我们在训练时使用了三种正则化方法：

Residual Dropout 残差连接中的丢弃法

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 $P_{drop}=0.1$ .
我们在每个子层的输出上应用 dropout [33]，然后再将其与子层输入相加并进行归一化。此外，在编码器和解码器的嵌入与位置编码的和上，我们也应用了 dropout。对于基础模型，dropout 率设为 $P_{drop}=0.1$ 。

Label Smoothing 标签平滑技术

During training, we employed label smoothing of value $\epsilon_{ls}=0.1$ [36]. This hurts perplexity, as the model learns to be more unsure, but improves accuracy and BLEU score.
在训练时，我们使用了值为 $\epsilon_{ls}=0.1$ 的标签平滑[36]。虽然这会使模型的困惑度变差，因为模型变得更加不确定，但却提升了准确率和 BLEU 评分。

6 Results 6 结果

6.1 Machine Translation 6.1 机器翻译

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：Transformer 在 English-to-German 和 English-to-French newstest2014 测试中，以更低的训练成本取得了优于以往最先进模型的 BLEU 分数。

Model	BLEU		Training Cost (FLOPs) 训练计算量（FLOPs）
Model	EN-DE	EN-FR	EN-DE	EN-FR
ByteNet [18] ByteNet [18]	23.75
Deep-Att + PosUnk [39] 深度注意力机制加位置未知处理 [39]		39.2		$1.0\cdot 10^{20}$
GNMT + RL [38] GNMT 加强化学习 [38]	24.6	39.92	$2.3\cdot 10^{19}$	$1.4\cdot 10^{20}$
ConvS2S [9]	25.16	40.46	$9.6\cdot 10^{18}$	$1.5\cdot 10^{20}$
MoE [32]	26.03	40.56	$2.0\cdot 10^{19}$	$1.2\cdot 10^{20}$
Deep-Att + PosUnk Ensemble [39] 深度注意力与位置未知词汇集成模型 [39]		40.4		$8.0\cdot 10^{20}$
GNMT + RL Ensemble [38] GNMT 与强化学习集成 [38]	26.30	41.16	$1.8\cdot 10^{20}$	$1.1\cdot 10^{21}$
ConvS2S Ensemble [9] ConvS2S 集成模型 [9]	26.36	41.29	$7.7\cdot 10^{19}$	$1.2\cdot 10^{21}$
Transformer (base model) Transformer（基础版模型）	27.3	38.1	$3.3\cdot 10^{18}$
Transformer (big) Transformer（大型）	28.4	41.8	$2.3\cdot 10^{19}$

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 英译德翻译任务中，大型 Transformer 模型（表 2 中的 Transformer (big)）比之前所有最佳模型（包括集成模型）高出超过 $2.0$ BLEU，创下了新的最先进 BLEU 记录 $28.4$ 。该模型的配置详见表 3 底部。训练过程在 $8$ 块 P100 GPU 上进行了 $3.5$ 天。即使是我们的基础模型，也以远低于其他竞争模型的训练成本，超越了所有之前发布的模型和集成模型。

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 $P_{drop}=0.1$ , instead of $0.3$ .
在 WMT 2014 英语到法语翻译任务中，我们的大型模型取得了 $41.0$ 的 BLEU 分数，优于所有之前发布的单模型，且训练成本不到之前最先进模型的 $1/4$ 。该用于英语到法语的 Transformer（大）模型采用了 $P_{drop}=0.1$ 的 dropout 率，而非之前的 $0.3$ 。

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，我们采用了将最后 5 个以 10 分钟间隔保存的检查点平均得到的单一模型。对于大型模型，则是平均了最后 20 个检查点。我们使用束搜索，束宽为 $4$ ，长度惩罚系数为 $\alpha=0.6$ [38]。这些超参数是在开发集上经过调试确定的。推理时，最大输出长度设为输入长度加 $50$ ，但在条件允许时会提前结束 [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 ²²2We used values of 2.8, 3.7, 6.0 and 9.5 TFLOPS for K80, K40, M40 and P100, respectively..
表 2 总结了我们的实验结果，并将我们的翻译质量和训练成本与文献中其他模型架构进行了对比。我们通过将训练时间、使用的 GPU 数量以及每个 GPU 的持续单精度浮点运算能力估算值相乘，来估算训练模型所需的浮点运算量 ² 。

6.2 Model Variations 6.2 模型的不同变体

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 架构的不同变体。未特别说明的参数与基础模型保持一致。所有指标均来自英译德的新测试集 2013。所列困惑度为每个字片的困惑度，基于我们的字节对编码，不应与按词计算的困惑度直接比较。

	$N$	$d_{\text{model}}$	$d_{\text{ff}}$	$h$	$d_{k}$	$d_{v}$	$P_{drop}$	$\epsilon_{ls}$	train 训练过程	PPL	BLEU	params 参数设置
	$N$	$d_{\text{model}}$	$d_{\text{ff}}$	$h$	$d_{k}$	$d_{v}$	$P_{drop}$	$\epsilon_{ls}$	steps 步骤	(dev) （开发环境）	(dev) （开发环境）	$\times 10^{6}$
base 基础	6	512	2048	8	64	64	0.1	0.1	100K	4.92	25.8	65
(A) （A）				1	512	512				5.29	24.9
				4	128	128				5.00	25.5
				16	32	32				4.91	25.8
				32	16	16				5.01	25.4
(B) （B）					16					5.16	25.1	58
(B) （B）					32					5.01	25.4	60
(C) （版权符号）	2									6.11	23.7	36
	4									5.19	25.3	50
	8									4.88	25.5	80
		256			32	32				5.75	24.5	28
		1024			128	128				4.66	26.0	168
			1024							5.12	25.4	53
			4096							4.75	26.2	90
(D) （D）							0.0			5.77	24.6
							0.2			4.95	25.5
								0.0		4.67	25.3
								0.2		5.47	25.7
(E) （E）		positional embedding instead of sinusoids 使用位置嵌入代替正弦波编码								4.92	25.7
big 大型的	6	1024	4096	16			0.3		300K	4.33	26.4	213

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，我们对基础模型进行了不同的调整，并在开发集 newstest2013 上测量了英译德翻译性能的变化。我们采用了前文介绍的束搜索方法，但未使用检查点平均。相关结果见表 3。

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）行中，我们调整了注意力头的数量以及注意力键和值的维度，同时保持计算量不变，具体如第 3.2.2 节所述。虽然单头注意力的表现比最佳设置低 0.9 BLEU，但注意力头过多时，质量也会下降。

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）行中，我们发现减少注意力键的大小 $d_{k}$ 会降低模型的性能。这表明判断兼容性并不简单，使用比点积更复杂的兼容性函数可能会更有效。我们还在（C）和（D）行看到，正如预期的那样，模型越大效果越好，且 dropout 对防止过拟合非常有帮助。在（E）行，我们用学习得到的位置嵌入[9]替换了正弦位置编码，结果与基础模型几乎相同。

6.3 English Constituency Parsing
6.3 英语成分句法解析

Table 4: The Transformer generalizes well to English constituency parsing (Results are on Section 23 of WSJ)
表 4：Transformer 模型在英语成分句法分析任务中表现出良好的泛化能力（结果来自 WSJ 语料库第 23 节）

Parser 语法分析器	Training 培训	WSJ 23 F1
Vinyals & Kaiser el al. (2014) [37] Vinyals 和 Kaiser 等人（2014）[37]	WSJ only, discriminative 仅使用 WSJ 数据，采用判别式方法	88.3
Petrov et al. (2006) [29] 彼得罗夫等人（2006 年）[29]	WSJ only, discriminative 仅使用 WSJ 数据，采用判别式方法	90.4
Zhu et al. (2013) [40] 朱等人（2013 年）[40]	WSJ only, discriminative 仅使用 WSJ 数据，采用判别式方法	90.4
Dyer et al. (2016) [8] 迪尔等人（2016 年）[8]	WSJ only, discriminative 仅使用 WSJ 数据，采用判别式方法	91.7
Transformer (4 layers) Transformer 模型（4 层）	WSJ only, discriminative 仅使用 WSJ 数据，采用判别式方法	91.3
Zhu et al. (2013) [40] 朱等人（2013 年）[40]	semi-supervised 半监督学习	91.3
Huang & Harper (2009) [14] 黄和哈珀（2009 年）[14]	semi-supervised 半监督学习	91.3
McClosky et al. (2006) [26] 麦克洛斯基等人（2006 年）[26]	semi-supervised 半监督学习	92.1
Vinyals & Kaiser el al. (2014) [37] Vinyals 和 Kaiser 等人（2014）[37]	semi-supervised 半监督学习	92.1
Transformer (4 layers) Transformer 模型（4 层）	semi-supervised 半监督学习	92.7
Luong et al. (2015) [23] Luong 等人（2015）[23]	multi-task 多任务处理	93.0
Dyer et al. (2016) [8] 迪尔等人（2016 年）[8]	generative 生成式	93.3

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 是否能推广到其他任务，我们在英语成分句法分析任务上进行了实验。该任务具有特殊挑战：输出具有严格的结构限制，且长度远超输入。此外，循环神经网络的序列到序列模型在小数据集条件下尚未达到最先进的效果 [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.
我们在宾夕法尼亚树库（Penn Treebank）中的华尔街日报（WSJ）部分训练了一个 4 层的 Transformer 模型，训练数据约有 4 万句。我们还在半监督环境下进行了训练，使用了更大规模的高置信度语料库和伯克利解析器语料库，约有 1700 万句。WSJ 单独训练时使用了 1.6 万个词汇表，半监督训练时则使用了 3.2 万个词汇表。

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 $\alpha=0.3$ for both WSJ only and the semi-supervised setting.
我们只做了少量实验来选择 dropout（包括注意力和残差，见第 5.4 节）、学习率以及第 22 节开发集上的束大小，其他参数均沿用英译德基础翻译模型。在推理时，我们将最大输出长度设置为输入长度加 $300$ 。在仅使用 WSJ 和半监督设置中，我们分别使用了束大小为 $21$ 和 $\alpha=0.3$ 。

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，尽管没有进行针对特定任务的调优，我们的模型表现依然出色，优于此前所有已报道的模型，仅次于循环神经网络语法[8]。

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.
与循环神经网络序列到序列模型[37]相比，Transformer 即使只在包含 4 万句子的 WSJ 训练集上训练，也能超过 BerkeleyParser[29]的表现。

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.
在本研究中，我们提出了 Transformer 模型，这是首个完全基于注意力机制的序列转换模型，它用多头自注意力机制取代了编码器-解码器架构中常用的循环层。

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.
在翻译任务中，Transformer 的训练速度远快于基于循环或卷积层的架构。在 WMT 2014 英译德和英译法的翻译任务中，我们取得了新的最佳成绩。在英译德任务中，我们的最佳模型甚至超越了所有之前公布的集成模型。

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.
我们对基于注意力机制的模型未来充满期待，计划将其应用到更多任务中。我们打算将 Transformer 扩展到处理文本以外的输入和输出形式，并探索局部受限的注意力机制，以高效应对图像、音频和视频等大规模数据。让生成过程更加并行化也是我们未来的重要研究方向。

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已公开，地址是 https://github.com/tensorflow/tensor2tensor。

Acknowledgements 鸣谢

We are grateful to Nal Kalchbrenner and Stephan Gouws for their fruitful comments, corrections and inspiration.
衷心感谢 Nal Kalchbrenner 和 Stephan Gouws 的宝贵建议、指正与启发。

References

[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. 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.
[3] Denny Britz, Anna Goldie, Minh-Thang Luong, and Quoc V. Le. Massive exploration of neural machine translation architectures. 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.
[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.
[6] Francois Chollet. Xception: Deep learning with depthwise separable convolutions. 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.
[8] Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, and Noah A. Smith. Recurrent neural network grammars. 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.
[10] Alex Graves. Generating sequences with recurrent neural networks. 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.
[12] Sepp Hochreiter, Yoshua Bengio, Paolo Frasconi, and Jürgen Schmidhuber. Gradient flow in recurrent nets: the difficulty of learning long-term dependencies, 2001.
[13] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. 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.
[15] Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. Exploring the limits of language modeling. 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.
[17] Łukasz Kaiser and Ilya Sutskever. Neural GPUs learn algorithms. 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.
[19] Yoon Kim, Carl Denton, Luong Hoang, and Alexander M. Rush. Structured attention networks. In International Conference on Learning Representations, 2017.
[20] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015.
[21] Oleksii Kuchaiev and Boris Ginsburg. Factorization tricks for LSTM networks. 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.
[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.
[24] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-based neural machine translation. 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.
[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.
[27] Ankur Parikh, Oscar Täckström, Dipanjan Das, and Jakob Uszkoreit. A decomposable attention model. 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.
[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.
[30] Ofir Press and Lior Wolf. Using the output embedding to improve language models. 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.
[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.
[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.
[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.
[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.
[36] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. 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.
[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.
[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.
[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.

Attention Is All You Need注意力机制就是全部所需