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. 只要提供适当的归属，Google 特此授予权限，仅用于新闻或学术作品中复制本文中的表格和图表。
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 Englishto-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. 支配性的序列转换模型是基于复杂的循环神经网络或卷积神经网络的，包括编码器和解码器。最佳性能的模型还通过注意机制将编码器和解码器连接起来。我们提出了一种新的简单的网络架构，Transformer，仅基于注意机制，完全摒弃了循环和卷积。在两个机器翻译任务上的实验证明，这些模型在质量上优于其他模型，并且更易并行化，训练时间显著减少。我们的模型在 WMT 2014 英德翻译任务上达到了 28.4 的 BLEU 值，相比现有最佳结果（包括集成模型），提高了 2 个 BLEU 值。在 WMT 2014 英法翻译任务中，我们的模型在 8 个 GPU 上训练 3.5 天后，取得了 41.8 的单模型最佳 BLEU 得分，训练成本仅为文献中最佳模型的一小部分。我们展示了 Transformer 在大型和有限的训练数据上成功应用于英文成分句法分析，表明它具有良好的泛化能力。
1 Introduction 1 简介
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[13]和门控循环神经网络[7]在序列建模和传导问题方面，如语言建模和机器翻译[35，2，5]已被牢固地确立为最先进的方法。此后，许多努力一直在推动循环语言模型和编码器-解码器架构的边界[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 , as a function of the previous hidden state and the input for position . 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和输出序列的符号位置因子化计算。将位置与计算时间步骤对齐，它们生成一个隐藏状态序列 ，作为前一隐藏状态 和位置 的输入函数。这种固有的序列性质排除了在训练实例内的并行化，这在较长序列长度时变得至关重要，因为在内存限制情况下跨实例的批处理受到限制。最近的工作通过因子化技巧[21]和条件计算[32]实现了计算效率的显著改进，同时在后者的情况下提高了模型性能。然而，顺序计算的基本限制仍然存在。
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. 注意机制已经成为各种任务中引人注目的序列建模和转导模型的一个组成部分，允许对依赖关系进行建模，而不考虑它们在输入或输出序列中的距离[2, 19]。然而，在除了少数情况[27]之外，这些注意机制通常与循环网络一起使用。
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，我们提出了 Transformer，这是一种模型架构，它避免了循环，并完全依赖于注意机制来绘制输入和输出之间的全局依赖关系。Transformer 允许更多的并行化，并且在仅仅在八个 P100 GPU 上训练了十二个小时后，就可以达到翻译质量的最新水平。
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 减少顺序计算的目标也构成了 Extended Neural GPU [16]、ByteNet [18]和 ConvS2S [9]的基础，它们都使用卷积神经网络作为基本构建模块，在所有输入和输出位置并行计算隐藏表示。在这些模型中，需要关联两个任意输入或输出位置的信号的操作数量随着位置之间的距离增加而增加，对于 ConvS2S 是线性增长，对于 ByteNet 是对数增长。这使得学习远距离位置之间的依赖关系变得更加困难。在 Transformer 中，这被减少为一定数量的操作，尽管由于平均关注加权位置而导致有效分辨率降低，我们通过第 3.2 节中描述的 Multi-Head Attention 来抵消这种影响。
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，有时称为内部注意力，是一种关注机制，涉及对单个序列的不同位置进行关联，以计算序列的表示。 自注意力已成功应用于各种任务，包括阅读理解，抽象总结，文本蕴涵和学习任务独立的句子表示[4, 27, 28, 22]。
End-to-end memory networks are based on a recurrent attention mechanism instead of sequencealigned recurrence and have been shown to perform well on simple-language question answering and language modeling tasks [34]. 端到端记忆网络基于循环注意机制，而不是序列对齐循环，并且已被证明在简单语言问答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 sequencealigned 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，然而，变压器是第一个完全依赖自我注意力来计算其输入和输出表示的传导模型，而不使用序列对齐的 RNN 或卷积。 在接下来的章节中，我们将描述变换器，激发自我注意力，并讨论其在模型[17, 18]和[9]方面的优势。
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 to a sequence of continuous representations . Given , the decoder then generates an output sequence 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. 大多数竞争性的神经序列转换模型都具有编码器-解码器结构[5, 2, 35]。在这里，编码器将符号表示的输入序列 映射到连续表示的序列 。给定 ，解码器随后逐个元素地生成一个符号输出序列 。在每个步骤中，模型是自回归的[10]，在生成下一个符号时消耗先前生成的符号作为附加输入。
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 遵循这种整体架构，使用堆叠的自注意力和点对点全连接层作为编码器和解码器，分别显示在图 1 的左半部分和右半部分。
3.1 Encoder and Decoder Stacks 3.1 编码器和解码器堆栈
Encoder: The encoder is composed of a stack of identical layers. Each layer has two sub-layers. The first is a multi-head self-attention mechanism, and the second is a simple, positionwise 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 LayerNorm , where Sublayer 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 . 编码器：编码器由一堆 个相同的层组成。每一层都有两个子层。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。我们在两个子层周围采用残差连接[11]，然后进行层归一化[1]。也就是说，每个子层的输出是 LayerNorm ，其中 Sublayer 是子层本身实现的函数。为了方便这些残差连接，模型中的所有子层以及嵌入层产生维度为 的输出。
Decoder: The decoder is also composed of a stack of 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 can depend only on the known outputs at positions less than . 解码器: 解码器也由一堆 个相同的层组成。除了每个编码器层中的两个子层外，解码器还插入第三个子层，该子层对编码器堆栈的输出执行多头注意力。与编码器类似，我们在每个子层周围采用残差连接，然后进行层归一化。我们还修改了解码器堆栈中的自注意力子层，以防止位置关注后续位置。这种掩码，结合输出嵌入偏移一个位置的事实，确保位置 的预测只能依赖于小于 位置的已知输出。
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 注意力函数可以描述为将查询和一组键-值对映射到输出，其中查询、键、值和输出都是向量。输出被计算为加权和
Figure 2: (left) Scaled Dot-Product Attention. (right) Multi-Head Attention consists of several attention layers running in parallel. 图 2：（左）缩放的点积注意力。（右）多头注意力由几个并行运行的注意力层组成。
of the values, where the weight assigned to each value is computed by a compatibility function of the query with the corresponding key. 对值的加权，其中分配给每个值的权重是通过查询与相应键的兼容性函数计算的。
3.2.1 Scaled Dot-Product Attention 3.2.1 缩放的点积注意力
We call our particular attention "Scaled Dot-Product Attention" (Figure 2). The input consists of queries and keys of dimension , and values of dimension . We compute the dot products of the query with all keys, divide each by , and apply a softmax function to obtain the weights on the values. 我们将我们的特别关注称为“缩放点积注意力”（图 2）。输入由维度为 的查询和键以及维度为 的值组成。我们计算查询与所有键的点积，将每个除以 ，并应用 softmax 函数以获得值的权重。
In practice, we compute the attention function on a set of queries simultaneously, packed together into a matrix . The keys and values are also packed together into matrices and . We compute the matrix of outputs as: 在实践中，我们同时对一组查询进行注意力函数计算，将它们打包成一个矩阵 。键和值也被打包到矩阵 和 中。我们计算输出矩阵如下：
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 . 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（乘法）注意力。点积注意力与我们的算法相同，除了缩放因子 。加性注意力使用具有单隐藏层的前馈网络计算兼容性函数。虽然在理论复杂性上两者相似，但在实践中，点积注意力更快速、更节省空间，因为它可以使用高度优化的矩阵乘法代码实现。
While for small values of the two mechanisms perform similarly, additive attention outperforms dot product attention without scaling for larger values of [3]. We suspect that for large values of , the dot products grow large in magnitude, pushing the softmax function into regions where it has extremely small gradients To counteract this effect, we scale the dot products by . 尽管对于小值的 ，这两种机制表现相似，但对于较大值的 ，加性注意力优于无缩放的点积注意力[3]。我们怀疑对于较大值的 ，点积会变得很大，将 softmax 函数推入具有极小梯度的区域 为了抵消这种影响，我们通过 缩放点积。
3.2.2 Multi-Head Attention 3.2.2 多头注意力
Instead of performing a single attention function with dimensional keys, values and queries, we found it beneficial to linearly project the queries, keys and values times with different, learned linear projections to and dimensions, respectively. On each of these projected versions of queries, keys and values we then perform the attention function in parallel, yielding -dimensional output values. These are concatenated and once again projected, resulting in the final values, as depicted in Figure 2 我们发现，与 维键、值和查询执行单个注意力函数相比，通过使用不同的、学习的线性投影将查询、键和值线性投影 次到 和 维度，有益于在每个投影版本的查询、键和值上并行执行注意力函数，得到 维输出值。然后将这些值进行串联，再次进行投影，得到最终值，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. 多头注意力允许模型同时关注不同位置的不同表示子空间中的信息。使用单个注意力头，平均会抑制这一点。
Where the projections are parameter matrices and . 投影是参数矩阵 和 。
In this work we employ parallel attention layers, or heads. For each of these we use . Due to the reduced dimension of each head, the total computational cost is similar to that of single-head attention with full dimensionality. 在这项工作中，我们使用 个并行注意力层或头。对于每个头，我们使用 。由于每个头的降维，总计算成本类似于具有完整维度的单头注意力。
3.2.3 Applications of Attention in our Model 我们模型中注意力的应用
The Transformer uses multi-head attention in three different ways: Transformer 在三种不同方式中使用多头注意力：
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，查询来自前一个解码器层，而记忆键和值来自编码器的输出。这允许解码器中的每个位置关注输入序列中的所有位置。这模仿了序列到序列模型中典型的编码器-解码器注意力机制，如[38, 2, 9]。
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，所有的键、值和查询都来自同一个地方，即编码器中前一层的输出。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 ) 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解码器中的每个位置关注解码器中直到该位置的所有位置。我们需要阻止解码器中的左向信息流，以保持自回归属性。我们通过在缩放的点积注意力中实现这一点，通过将与非法连接对应的 softmax 输入中的所有值屏蔽掉（设置为 ）。参见图 2。
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. 除了注意力子层之外，我们的编码器和解码器中的每个层都包含一个全连接的前馈网络，该网络分别且相同地应用于每个位置。这由两个线性变换组成，中间有一个 ReLU 激活。
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 , and the inner-layer has dimensionality . 虽然线性变换在不同位置上是相同的，但它们在层与层之间使用不同的参数。另一种描述这一点的方式是两个卷积核大小为 1。输入和输出的维度是 ，内层的维度是 。
3.4 Embeddings and Softmax 3.4 嵌入和 Softmax
Similarly to other sequence transduction models, we use learned embeddings to convert the input tokens and output tokens to vectors of dimension . 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 . 与其他序列转导模型类似，我们使用学习的嵌入将输入标记和输出标记转换为维度为 的向量。我们还使用通常的学习线性变换和 softmax 函数将解码器输出转换为预测的下一个标记概率。在我们的模型中，我们在两个嵌入层和预 softmax 线性变换之间共享相同的权重矩阵，类似于[30]。在嵌入层中，我们将这些权重乘以 。
Table 1: Maximum path lengths, per-layer complexity and minimum number of sequential operations for different layer types. is the sequence length, is the representation dimension, is the kernel size of convolutions and the size of the neighborhood in restricted self-attention. 表 1：不同层类型的最大路径长度、每层复杂度和最小顺序操作数。 是序列长度， 是表示维度， 是卷积的内核大小， 是受限自注意力中邻域的大小。
Layer Type
Complexity per Layer 每层的复杂度
Sequential
Operations
Maximum Path Length 最大路径长度
Self-Attention
Recurrent
Convolutional
Self-Attention (restricted) 自注意力（受限）
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 as the embeddings, so that the two can be summed. There are many choices of positional encodings, learned and fixed [9]. 由于我们的模型不包含任何循环和卷积，为了使模型利用序列的顺序，我们必须注入一些关于序列中令牌的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添加“位置编码”。位置编码与嵌入具有相同的维度 ，因此可以将两者相加。有许多位置编码的选择，可以学习和固定[9]。
In this work, we use sine and cosine functions of different frequencies: 在这项工作中，我们使用不同频率的正弦和余弦函数：
where pos is the position and is the dimension. That is, each dimension of the positional encoding corresponds to a sinusoid. The wavelengths form a geometric progression from to . 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 can be represented as a linear function of . 其中 pos 是位置， 是维度。也就是说，位置编码的每个维度对应于一个正弦波。波长从 到 形成一个几何级数。我们选择这个函数，因为我们假设它可以让模型轻松地学习通过相对位置进行关注，因为对于任何固定的偏移 可以表示为 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. 我们还尝试使用学习到的位置嵌入[9]，发现两个版本产生的结果几乎相同（见表 3 行（E））。我们选择了正弦版本，因为它可能使模型能够推广到比训练过程中遇到的序列长度更长的情况。
4 Why Self-Attention 4 为什么自注意力
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 to another sequence of equal length , with , 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，我们将自注意力层的各个方面与常用于将一个变长符号表示序列 映射到另一个等长序列 的循环和卷积层进行比较，其中 ，例如典型序列转导编码器或解码器中的隐藏层。激励我们使用自注意力的是我们考虑的三个愿望。
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。另一个是可以并行计算的量，由所需的最小顺序操作数量来衡量。
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. 第三个是网络中长程依赖之间的路径长度。学习长程依赖是许多序列转导任务中的关键挑战。影响学习这种依赖能力的一个关键因素是前向和后向信号在网络中必须穿过的路径长度。在输入和输出序列的任意位置之间的路径越短，学习长程依赖就越容易[12]。因此，我们还比较了由不同层类型组成的网络中任意两个输入和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 sequential operations. In terms of computational complexity, self-attention layers are faster than recurrent layers when the sequence 如表 1 中所示，自注意力层将所有位置连接起来，需要恒定数量的顺序执行操作，而循环层需要 个顺序操作。就计算复杂度而言，当序列
length is smaller than the representation dimensionality , 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 in the input sequence centered around the respective output position. This would increase the maximum path length to . We plan to investigate this approach further in future work. 长度 小于表示维度 ，这在最先进的机器翻译模型中使用的句子表示（例如 word-piece [38] 和 byte-pair [31] 表示）中最常见。为了提高涉及非常长序列的任务的计算性能，自注意力可以限制为仅考虑以相应输出位置为中心的输入序列大小为 的邻域。这将增加最大路径长度到 。我们计划在未来的工作中进一步研究这种方法。
A single convolutional layer with kernel width does not connect all pairs of input and output positions. Doing so requires a stack of convolutional layers in the case of contiguous kernels, or 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 . Separable convolutions [6], however, decrease the complexity considerably, to . Even with , 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. 一个宽度为 的单个卷积层不会连接所有输入和输出位置的所有对。要做到这一点，对于连续内核，需要一堆 个卷积层，或者对于扩张卷积[18]，需要 个卷积层，增加网络中任意两个位置之间最长路径的长度。卷积层通常比循环层更昂贵，比例为 。然而，可分离卷积[6]大大降低了复杂性，降至 。然而，即使有 ，可分离卷积的复杂性仍等于自注意力层和逐点前馈层的组合，这是我们模型采用的方法。
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. 作为副产品，自注意力可以产生更具可解释性的模型。我们检查模型的注意力分布，并在附录中呈现和讨论示例。不仅个别注意力头明显学会执行不同任务，许多似乎表现出与句子的句法和语义结构相关的行为。
5 Training 5 训练
This section describes the training regime for our models. 本节描述了我们模型的训练制度。
5.1 Training Data and Batching 5.1 训练数据和批处理
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 sourcetarget 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. 我们使用了标准的 WMT 2014 英德数据集进行训练，该数据集包含约 450 万个句对。句子使用字节对编码[3]进行编码，共享源目标词汇约 37000 个标记。对于英法语言对，我们使用了规模更大的 WMT 2014 英法数据集，包含 3600 万个句子，并将标记分成了一个 32000 个词片词汇[38]。句对按照近似序列长度进行分批处理。每个训练批次包含一组句对，其中包含大约 25000 个源标记和 25000 个目标标记。
5.2 Hardware and Schedule 5.2 硬件和时间表
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 个 NVIDIA P100 GPU 的机器上训练了我们的模型。对于使用本文中描述的超参数的基础模型，每个训练步骤大约需要 0.4 秒。我们为基础模型总共训练了 100,000 步或 12 小时。对于我们的大模型（在表 3 的最后一行描述），步骤时间为 1.0 秒。大模型训练了 300,000 步（3.5 天）。
5.3 Optimizer 5.3 优化器
We used the Adam optimizer [20] with and . We varied the learning rate over the course of training, according to the formula: 我们使用了 Adam 优化器 [20]，其中 和 。我们根据以下公式在训练过程中改变学习率：
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 . 这相当于在前 warmup_steps 训练步骤中线性增加学习率，然后按照步骤数的倒数平方根成比例地减小学习率。我们使用了 warmup_steps 。
5.4 Regularization 5.4 正则化
We employ three types of regularization during training: 我们在训练过程中使用三种类型的正则化：
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 在英德和英法 newstest2014 测试中获得比先前最先进模型更好的 BLEU 分数，训练成本仅为其一小部分。
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 . 残差丢失 我们在每个子层的输出添加丢失[33]，在它被添加到子层输入和归一化之前。此外，我们在编码器和解码器堆栈中的嵌入和位置编码的总和上应用丢失。对于基本模型，我们使用 的速率。
Label Smoothing During training, we employed label smoothing of value [36]. This hurts perplexity, as the model learns to be more unsure, but improves accuracy and BLEU score. 标签平滑 在训练期间，我们采用了价值 的标签平滑[36]。这会降低困惑度，因为模型学习变得更加不确定，但会提高准确性和 BLEU 分数。
6 Results 6 个结果
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 年英语到德语翻译任务中，大型变压器模型（表 2 中的 Transformer（big））的表现优于先前报告的最佳模型（包括集成模型）超过 2.0 BLEU，建立了一个新的最先进的 BLEU 得分为 28.4。该模型的配置列在表 3 的底部行中。在 8 个 P100 GPU 上训练耗时 3.5 天。即使是我们的基础模型也超过了以前发布的所有模型和集成模型，而训练成本只是竞争模型中任何一个的一小部分。
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 the training cost of the previous state-of-the-art model. The Transformer (big) model trained for English-to-French used dropout rate , instead of 0.3 . 在 WMT 2014 年的英法翻译任务中，我们的大型模型获得了 41.0 的 BLEU 分数，在训练成本不到之前先进模型的一半的情况下，表现优于以往发布的所有单一模型。我们为英法翻译训练的 Transformer（大型）模型使用了 的 dropout 率，而非 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 [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，我们使用了由最后 5 个检查点平均而得的单一模型，这些检查点是以 10 分钟间隔写入的。对于大型模型，我们使用了最后 20 个检查点的平均。我们使用了束搜索，束大小为 4，长度惩罚 [38]。这些超参数经过在开发集上的试验选择。我们在推理过程中将最大输出长度设置为输入长度+50，但尽可能提前终止[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 表 2 总结了我们的结果，并将我们的翻译质量和训练成本与文献中其他模型结构进行了比较。我们通过将训练时间、使用的 GPU 数量以及每个 GPU 的持续单精度浮点容量的估计相乘，来估计训练模型所需的浮点运算次数 。
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 为了评估 Transformer 不同组件的重要性，我们以不同方式改变了我们的基础模型，测量了在英语到德语翻译中性能的变化
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。 未列出的值与基础模型相同。 所有指标均基于英语到德语翻译开发集 newstest2013。 列出的困惑度是每个 wordpiece 的困惑度，根据我们的字节对编码，不应与每个单词的困惑度进行比较。
train
steps
PPL
(dev)
BLEU
params
base
6
512
2048
8
64
64
0.1
0.1
4.92
25.8
65
(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)
16
5.16
25.1
58
32
5.01
25.4
60
(C)
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)
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)
positional embedding instead of sinusoids 使用位置嵌入代替正弦函数
4.92
25.7
big
6
1024
4096
16
0.3
4.33
26.4
213
development set, newstest2013. We used beam search as described in the previous section, but no checkpoint averaging. We present these results in Table 3 开发集为 newstest2013。我们使用前一节中描述的波束搜索，但没有进行检查点平均。我们在表 3 中呈现这些结果。
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 worse than the best setting, quality also drops off with too many heads. 在表 3 的(A)行中，我们改变了注意力头的数量和注意力键和值的维度，保持计算量不变，如 3.2.2 节所述。虽然单头注意力比最佳设置差 ，但质量也会随着头的过多而下降。
In Table 3 rows (B), we observe that reducing the attention key size 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. 在表格 3 的 B 行中，我们观察到减小注意力键大小 会损害模型质量。这表明确定兼容性并不容易，比点积更复杂的兼容性函数可能会更有益。我们在(C)和(D)行进一步观察到，正如预期的那样，更大的模型更好，并且辍学非常有助于避免过拟合。在(E)行中，我们用学习的位置嵌入[9]替换了我们的正弦位置编码，并观察到与基础模型几乎相同的结果。
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 是否能够推广到其他任务，我们对英语成分句法分析进行了实验。这个任务提出了特定的挑战：输出受到强烈的结构约束，并且比输入要长得多。此外，RNN 序列到序列模型在小数据情况下无法达到最先进的结果[37]。
We trained a 4-layer transformer with 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 sentences [37]. We used a vocabulary of tokens for the WSJ only setting and a vocabulary of tokens for the semi-supervised setting. 我们在宾州树库的华尔街日报（WSJ）部分训练了一个 4 层的变压器，大约有 40K 个训练句子。我们还在半监督设置中训练它，使用了更大的高置信度和 BerkleyParser 语料库，其中包含大约 个句子。我们在 WSJ 设置中使用了 个标记的词汇表，在半监督设置中使用了 个标记的词汇表。
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 我们只进行了少量实验，以选择在第 22 节开发集上的辍学率、注意力和残差（第 5.4 节）、学习率和波束大小，所有其他参数均保持不变，与英德基础翻译模型相同。在推断过程中，我们
Table 4: The Transformer generalizes well to English constituency parsing (Results are on Section 23 of WSJ) 表 4：变压器在英语成分解析中表现良好（结果在 WSJ 的第 23 节）
Parser
Training
WSJ 23 F1
Vinyals & Kaiser el al. (2014) [37] Vinyals＆Kaiser 等（2014）[37]
WSJ only, discriminative 仅限 WSJ，有区别
88.3
Petrov et al. (2006) [29] Petrov 等（2006）[29]
WSJ only, discriminative 仅限 WSJ，歧视性
90.4
Zhu et al. (2013) [40] 朱等人（2013）[40]
WSJ only, discriminative 仅限 WSJ，歧视性
90.4
Dyer et al. (2016) [8] Dyer 等人（2016 年）[8]
WSJ only, discriminative 仅限 WSJ，具有区分性
91.7
Transformer (4 layers) Transformer（4 层）
WSJ only, discriminative 仅限 WSJ，歧视性
91.3
Zhu et al. (2013) [40] 朱等人（2013）[40]
semi-supervised
91.3
Huang & Harper (2009) [14] 黄＆哈珀（2009）[14]
semi-supervised
91.3
McClosky et al. (2006) [26] McClosky 等人（2006 年）[26]
semi-supervised
92.1
Vinyals & Kaiser el al. (2014) [37] Vinyals＆Kaiser 等人（2014 年）[37]
semi-supervised
92.1
Transformer (4 layers) 变压器（4 层）
semi-supervised
92.7
Luong et al. (2015) [23] Luong 等人（2015）[23]
multi-task
93.0
Dyer et al. (2016) [8] Dyer 等人（2016）[8]
generative
93.3
increased the maximum output length to input length +300 . We used a beam size of 21 and for both WSJ only and the semi-supervised setting. 将最大输出长度增加到输入长度+300。我们在仅使用 WSJ 和半监督设置时都使用了 21 的波束大小和 。
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]. 在表 4 中，我们的结果表明，尽管缺乏特定任务的调整，我们的模型表现出乎意料的好，产生的结果比以往所有报告的模型都要好，除了循环神经网络语法[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 RNN 序列到序列模型[37]相比，即使只在 40K 句子的 WSJ 训练集上训练，Transformer 也优于 BerkeleyParser[29]。
7 Conclusion 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，这是第一个完全基于注意力的序列转导模型，用多头自注意力替代了在编码器-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 的训练速度比基于循环或卷积层的架构快得多。在 WMT 2014 英德翻译和 WMT 2014 英法翻译任务中，我们取得了新的最先进水平。在前一项任务中，我们的最佳模型甚至优于以往所有报告的集成模型。
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感到兴奋，并计划将它们应用于其他任务。我们计划将 Transformer 扩展到涉及文本以外的输入和输出模态的问题，并研究局部、受限的注意力机制，以有效处理大输入和输出，如图像、音频和视频。使生成 less sequential 是我们的另一个研究目标。
The code we used to train and evaluate our models is available at https://github.com// tensorflow/tensor2tensor. 我们用来训练和评估模型的代码可在 https://github.com// tensorflow/tensor2tensor 找到。
Acknowledgements We are grateful to Nal Kalchbrenner and Stephan Gouws for their fruitful comments, corrections and inspiration. 致谢 我们感谢 Nal Kalchbrenner 和 Stephan Gouws 对我们的有益评论、纠正和启发。
References 参考资料
[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. [1] 李宝, Jamie Ryan Kiros, 和 Geoffrey E Hinton. 层归一化. arXiv 预印本 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, 和 Yoshua Bengio. 神经机器翻译通过联合学习对齐和翻译. 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, 和 Quoc V. Le. 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, 和 Mirella Lapata。长短期记忆网络用于机器阅读。arXiv 预印本 arXiv:1601.06733，2016 年。
[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, 和 Yoshua Bengio。使用 RNN 编码器-解码器学习短语表示进行统计机器翻译。CoRR，abs/1406.1078，2014 年。
[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进行深度学习。arXiv 预印本 arXiv:1610.02357，2016 年。
[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 和 Yoshua Bengio。关于序列建模中的门控循环神经网络的实证评估。CoRR，abs/1412.3555，2014 年。
[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 和 Noah A. Smith。递归神经网络语法。在 NAACL 的会议记录中，2016 年。
[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 和 Yann N. Dauphin。卷积序列到序列学习。arXiv 预印本 arXiv:1705.03122v2，2017 年。
[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 预印本 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, 和 Jian Sun. 图像识别的深度残差学习. 在 IEEE 计算机视觉与模式识别会议论文集中, 页码 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, 和 Jürgen Schmidhuber. 循环网络中的梯度流: 学习长期依赖性的困难, 2001.
[13] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. [13] Sepp Hochreiter and Jürgen Schmidhuber. 长期短期记忆神经计算, 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] 黄忠强和玛丽·哈珀. 利用跨语言潜在注释的自我训练 PCFG 语法. 在 2009 年自然语言处理经验方法会议论文集中, 832-841 页. ACL, 2009 年 8 月.
[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, 和 Yonghui Wu. 探索语言建模的极限. arXiv 预印本 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 和 Samy Bengio。主动记忆能否取代注意力？在神经信息处理系统进展（NIPS）中，2016 年。
[17] Łukasz Kaiser and Ilya Sutskever. Neural GPUs learn algorithms. In International Conference on Learning Representations (ICLR), 2016. [17] Łukasz Kaiser 和 Ilya Sutskever。神经 GPU 学习算法。在国际学习表示会议（ICLR）中，2016 年。
[18] Nal Kalchbrenner, Lasse Espeholt, Karen Simonyan, Aaron van den Oord, Alex Graves, and Koray Kavukcuoglu. Neural machine translation in linear time. arXiv preprint v2, 2017. [18] Nal Kalchbrenner，Lasse Espeholt，Karen Simonyan，Aaron van den Oord，Alex Graves 和 Koray Kavukcuoglu。线性时间内的神经机器翻译。arXiv 预印本 v2，2017 年。
[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, 和 Alexander M. Rush. 结构化注意力网络. 在 2017 年国际学习表示会议上。
[20] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. [20] Diederik Kingma 和 Jimmy Ba. Adam：一种随机优化方法。在 ICLR, 2015 年。
[21] Oleksii Kuchaiev and Boris Ginsburg. Factorization tricks for LSTM networks. arXiv preprint arXiv:1703.10722, 2017. [21] Oleksii Kuchaiev 和 Boris Ginsburg. LSTM 网络的分解技巧。arXiv 预印本 arXiv:1703.10722, 2017 年。
[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] 周瀚琳，冯敏伟，西塞罗·桑托斯，余墨，向兵，周博文和 Yoshua Bengio。一种结构化的自注意句子嵌入。arXiv 预印本 arXiv:1703.03130，2017 年。
[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] 龙明堂，乐趣，伊利亚·苏茨凯弗，奥里奥尔·维尼亚尔斯和卢卡什·凯撒。多任务序列到序列学习。arXiv 预印本 arXiv:1511.06114，2015 年。
[24] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attentionbased neural machine translation. arXiv preprint arXiv:1508.04025, 2015. [24] 龙明堂，范显慧和克里斯托弗·D·曼宁。基于注意力的神经机器翻译的有效方法。arXiv 预印本 arXiv:1508.04025，2015 年。
[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] 米切尔 P 马库斯，玛丽·安·马尔钦凯维奇和比阿特丽斯·桑托里尼。构建英语的大型注释语料库：宾夕法尼亚语料树库。计算语言学，19（2）：1993 年的 313-330。
[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] 大卫·麦克洛斯基、尤金·查尼亚克和马克·约翰逊。自我训练的有效解析。在 NAACL 的人类语言技术会议论文集上，主要会议论文集，页码 152-159。 ACL，2006 年 6 月。
[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 和 Jakob Uszkoreit。一个可分解的注意力模型。在自然语言处理的经验方法，2016 年。
[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, 和 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 21 st 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, 和 Dan Klein. Learning accurate, compact, 和 interpretable tree annotation. In Proceedings of the 21 st International Conference on Computational Linguistics 和 44th Annual Meeting of the ACL, 页面 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 和 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. [31] Rico Sennrich, Barry Haddow, 和 Alexandra Birch. 神经机器翻译中使用次词单元翻译稀有词汇. arXiv 预印本 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, 和 Jeff Dean. 极大规模神经网络: 稀疏门控专家混合层. arXiv 预印本 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, 和 Ruslan Salakhutdinov. Dropout: 一种简单的方法防止神经网络过拟合. 机器学习研究杂志, 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, 和 Rob Fergus. 端到端记忆网络. 在 C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, 和 R. Garnett 编辑, Advances in Neural Information Processing Systems 28, 页码 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, 和 Quoc VV Le. 序列到序列学习与神经网络. 在 Advances in Neural Information Processing Systems, 页码 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, 和 Zbigniew Wojna. 重新思考计算机视觉的 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 和 Hinton. 语法作为一门外语。在 2015 年神经信息处理系统进展中。
[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] 吴永辉，迈克·舒斯特，陈志峰，乐文国，穆罕默德·诺鲁兹，沃尔夫冈·马赫雷，马克西姆·克里昆，曹元，高琴，克劳斯·马赫雷等。谷歌的神经机器翻译系统：弥合人类和机器翻译之间的差距。arXiv 预印本 arXiv:1609.08144，2016 年。
[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] 周杰，曹颖，王旭光，李鹏和徐伟。具有快速前向连接的深度递归模型用于神经机器翻译。CoRR，abs/1606.04199，2016 年。
[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] 朱木华，张跃，陈文亮，张敏和朱靖波。快速准确的移位-规约成分解析。在 ACL 第 51 届年会论文集（第 1 卷：长篇论文）中，第 434-443 页。ACL，2013 年 8 月。
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：编码器自注意力中遵循远距离依赖关系的注意力机制示例，在第 5 层中的第 6 层。许多注意力头关注动词'making'的远距离依赖关系，完成短语'making...更困难'。这里仅显示了单词'making'的注意力。不同颜色代表不同的注意力头。最好以彩色查看。
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: 两个注意力头，也在第 5 层的 6 层中，显然参与了指代消解。顶部: 第 5 个头的完整注意力。底部: 仅针对单词 'its' 的注意力头 5 和 6 的孤立注意力。请注意，对于这个单词，注意力非常集中。
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: 许多注意力头表现出似乎与句子结构相关的行为。我们在上面给出了两个这样的例子，来自编码器自注意力的两个不同头的第 5 层的 6 层。这些头显然学会了执行不同的任务。
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. 相等的贡献。列表顺序是随机的。Jakob 提出用自注意力替换 RNN，并开始评估这个想法。Ashish 和 Illia 设计并实现了第一个 Transformer 模型，并在这项工作的每个方面都发挥了关键作用。Noam 提出了缩放的点积注意力，多头注意力和无参数位置表示，并成为几乎每个细节中涉及的另一个人。Niki 在我们的原始代码库和 tensor2tensor 中设计、实现、调整和评估了无数的模型变体。Llion 还尝试了新颖的模型变体，负责我们的初始代码库，以及高效的推理和可视化。Lukasz 和 Aidan 花费了无数的日子设计和实现了 tensor2tensor 的各个部分，取代了我们之前的代码库，极大地改善了结果并大大加速了我们的研究。
Work performed while at Google Brain. 在 Google Brain 期间进行的工作。
Work performed while at Google Research. 在 Google Research 期间进行的工作。
To illustrate why the dot products get large, assume that the components of and are independent random variables with mean 0 and variance 1 . Then their dot product, , has mean 0 and variance . 为了说明为什么点积会变大，假设 和 的分量是独立的随机变量，均值为 0，方差为 1。那么它们的点积 的均值为 0，方差为 。
We used values of 2.8, 3.7, 6.0 and 9.5 TFLOPS for K80, K40, M40 and P100, respectively. 我们分别使用了 2.8、3.7、6.0 和 9.5 TFLOPS 的值来表示 K80、K40、M40 和 P100。
