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.只要提供适当的署名，Google 在此授予许可，允许仅在新闻或学术作品中复制本文中的表格和图形。

Attention Is All You Need注意力即一切

Ashish 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.主流的序列转导模型基于复杂的循环或卷积神经网络，包含编码器和解码器。表现最好的模型还通过注意力机制连接编码器和解码器。我们提出了一种全新的简洁网络架构——Transformer，仅基于注意力机制，完全摒弃循环和卷积。对两个机器翻译任务的实验表明，这些模型在质量上更优，同时更易并行化，训练时间显著更短。我们的模型在WMT 2014英德翻译任务上取得了28.4 BLEU，较包括集成在内的现有最佳结果提升超过2 BLEU。在WMT 2014英法翻译任务上，我们的模型在8块GPU上训练3.5天后，达到了41.8 BLEU的单模型新纪录，训练成本仅为文献中最佳模型的一小部分。我们还展示了Transformer在其他任务上的良好泛化能力，成功应用于大规模和小规模训练数据的英语成分句法分析。

1 Introduction1 引言

Recurrent neural networks, long short-term memory [13] and gated recurrent [7] neural networks in particular, have been firmly established as state of the art approaches in sequence modeling and transduction problems such as language modeling and machine translation [35, 2, 5]. Numerous efforts have since continued to push the boundaries of recurrent language models and encoder-decoder architectures [38, 24, 15].循环神经网络，尤其是长短期记忆网络[13]和门控循环网络[7]，已被牢固确立为序列建模和转导问题（如语言建模和机器翻译[35, 2, 5]）的最先进方法。此后大量工作继续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 $h_{t}$ , as a function of the previous hidden state $h_{t-1}$ and the input for position $t$ . This inherently sequential nature precludes parallelization within training examples, which becomes critical at longer sequence lengths, as memory constraints limit batching across examples. Recent work has achieved significant improvements in computational efficiency through factorization tricks [21] and conditional computation [32], while also improving model performance in case of the latter. The fundamental constraint of sequential computation, however, remains.循环模型通常沿输入和输出序列的符号位置进行计算。将位置对应到计算时间步，它们生成一系列隐藏状态 h_t，作为前一隐藏状态 h_{t-1} 和位置 t 的输入的函数。这种固有的顺序性阻碍了在训练样本内部的并行化，这在序列较长时尤为关键，因为内存限制了跨样本的批处理。近期工作通过因式分解技巧[21]和条件计算[32]显著提升了计算效率，同时在后者情况下也提升了模型性能。但顺序计算的根本约束仍然存在。

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.注意力机制已成为各种任务中引人注目的序列建模和转导模型的组成部分，能够在不考虑输入或输出序列距离的情况下建模依赖关系[2, 19]。然而，除少数情况[27]外，这类注意力机制通常与循环网络结合使用。

In this work we propose the Transformer, a model architecture eschewing recurrence and instead relying entirely on an attention mechanism to draw global dependencies between input and output. The Transformer allows for significantly more parallelization and can reach a new state of the art in translation quality after being trained for as little as twelve hours on eight P100 GPUs.在本工作中，我们提出了Transformer，一种摒弃循环、完全依赖注意力机制来捕获输入输出之间全局依赖的模型架构。Transformer显著提升了并行化程度，并在仅用八块P100 GPU训练十二小时后，就在翻译质量上达到了新的最先进水平。

2 Background2 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.降低顺序计算的目标也是Extended Neural GPU[16]、ByteNet[18]和ConvS2S[9]的基础，这些模型都使用卷积神经网络作为基本构件，对所有输入输出位置并行计算隐藏表示。在这些模型中，将两个任意输入或输出位置的信号关联起来所需的操作次数随位置距离线性增长（ConvS2S）或对数增长（ByteNet），这使得学习远距离位置之间的依赖更加困难[12]。在Transformer中，这一过程被简化为常数次操作，尽管由于对注意力加权位置的平均导致有效分辨率下降，我们通过第3.2节描述的多头注意力来抵消这一影响。

Self-attention, sometimes called intra-attention is an attention mechanism relating different positions of a single sequence in order to compute a representation of the sequence. Self-attention has been used successfully in a variety of tasks including reading comprehension, abstractive summarization, textual entailment and learning task-independent sentence representations [4, 27, 28, 22].自注意力（有时称为内部注意力）是一种将单个序列中不同位置关联起来以计算序列表示的注意力机制。自注意力已成功用于阅读理解、抽象摘要、文本蕴含以及学习任务无关的句子表示等多种任务[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].端到端记忆网络基于循环注意力机制而非序列对齐的循环，已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 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是首个完全依赖自注意力来计算输入输出表示、且不使用序列对齐RNN或卷积的转导模型。接下来的章节将描述Transformer，阐述自注意力的动机，并讨论其相较于[17, 18]和[9]等模型的优势。

3 Model Architecture3 模型架构

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.大多数竞争性的神经序列转导模型采用编码器-解码器结构[5, 2, 35]。编码器将符号表示序列 (x_1,…,x_n) 映射为连续表示序列 \mathbf{z}=(z_1,…,z_n)。给定 \mathbf{z}，解码器随后一次生成一个符号，形成输出序列 (y_1,…,y_m)。在每一步，模型是自回归的[10]，在生成下一个符号时会将先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 使用堆叠的自注意力和逐点全连接层，分别构成编码器和解码器，如图1左、右两半所示。

3.1 Encoder and Decoder Stacks3.1 编码器和解码器堆叠

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 个相同层堆叠而成。每层有两个子层：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。我们在每个子层周围使用残差连接[11]，随后进行层归一化[1]。即每个子层的输出为 LayerNorm(x+Sublayer(x))，其中 Sublayer(x) 是子层本身实现的函数。为实现这些残差连接，模型中所有子层以及嵌入层的输出维度均为 d_{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 个相同层堆叠而成。除每个编码器层的两个子层外，解码器在每层插入第三个子层，对编码器堆叠的输出进行多头注意力。与编码器类似，我们在每个子层周围使用残差连接并进行层归一化。我们还修改了解码器堆叠中的自注意力子层，防止位置关注后续位置。该掩码结合输出嵌入向后偏移一位，确保位置 i 的预测只能依赖于位置小于 i 的已知输出。

3.2 Attention3.2 注意力

An attention function can be described as mapping a query and a set of key-value pairs to an output, where the query, keys, values, and output are all vectors. The output is computed as a weighted sum of the values, where the weight assigned to each value is computed by a compatibility function of the query with the corresponding key.注意力函数可以描述为将查询和一组键-值对映射到输出，其中查询、键、值和输出都是向量。输出通过对值的加权求和得到，权重由查询与对应键的兼容函数计算。

3.2.1 Scaled Dot-Product Attention3.2.1 缩放点积注意力

We call our particular attention "Scaled Dot-Product Attention" (Figure 2). The input consists of queries and keys of dimension $d_{k}$ , and values of dimension $d_{v}$ . We compute the dot products of the query with all keys, divide each by $\sqrt{d_{k}}$ , and apply a softmax function to obtain the weights on the values.我们将我们的特定注意力称为“缩放点积注意力”（图2）。输入由维度 d_k 的查询和键，以及维度 d_v 的值组成。我们计算查询与所有键的点积，将每个结果除以 \sqrt{d_k}，并对得到的值使用 softmax 以获得对值的权重。

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:实际中，我们同时对一组查询进行注意力计算，将它们打包成矩阵 Q。键和值也分别打包成矩阵 K 和 V。我们将输出矩阵计算为：

\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.最常用的两种注意力函数是加性注意力[2]和点积（乘性）注意力。点积注意力与我们的算法相同，只是缩放因子为 1/\sqrt{d_k}。加性注意力使用单隐藏层的前馈网络计算兼容函数。虽然两者在理论复杂度上相似，点积注意力在实践中更快且更节省空间，因为它可以利用高度优化的矩阵乘法实现。

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 $\frac{1}{\sqrt{d_{k}}}$ .当 d_k 较小时，两种机制表现相近；但在 d_k 较大时，加性注意力优于未缩放的点积注意力[3]。我们推测在大 d_k 时，点积的幅度会变大，使 softmax 进入梯度极小的区域。为说明点积为何会变大，假设 q 和 k 的分量是均值为 0、方差为 1 的独立随机变量，则它们的点积 q·k 的均值为 0，方差为 d_k。为抵消这种效应，我们将点积除以 1/\sqrt{d_k}。

3.2.2 Multi-Head Attention3.2.2 多头注意力

Instead of performing a single attention function with $d_{\text{model}}$ -dimensional keys, values and queries, we found it beneficial to linearly project the queries, keys and values $h$ times with different, learned linear projections to $d_{k}$ , $d_{k}$ and $d_{v}$ dimensions, respectively. On each of these projected versions of queries, keys and values we then perform the attention function in parallel, yielding $d_{v}$ -dimensional output values. These are concatenated and once again projected, resulting in the final values, as depicted in Figure 2.我们没有使用 d_{model} 维的键、值和查询一次性完成注意力，而是将查询、键、值分别线性投影 h 次到不同的维度 d_k、d_k、d_v。对每个投影后的查询、键、值并行执行注意力函数，得到 d_v 维的输出值。将这些输出拼接后再次投影，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.多头注意力使模型能够在不同位置同时关注来自不同表示子空间的信息。单头注意力的平均会抑制这种能力。

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

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 $W^{O}\in\mathbb{R}^{hd_{v}\times d_{\text{model}}}$ .其中投影矩阵为参数矩阵 W^{Q}_{i}\in\mathbb{R}^{d_{model}\times d_{k}}、W^{K}_{i}\in\mathbb{R}^{d_{model}\times d_{k}}、W^{V}_{i}\in\mathbb{R}^{d_{model}\times d_{v}}，以及 W^{O}\in\mathbb{R}^{h d_{v}\times d_{model}}。

In this work we employ $h=8$ parallel attention layers, or heads. For each of these we use $d_{k}=d_{v}=d_{\text{model}}/h=64$ . Due to the reduced dimension of each head, the total computational cost is similar to that of single-head attention with full dimensionality.本工作中我们使用 h=8 条并行注意力层（头）。每条的维度为 d_k=d_v=d_{model}/h=64。由于每个头的维度降低，总计算成本与全维度单头注意力相当。

3.2.3 Applications of Attention in our Model3.2.3 注意力在模型中的应用

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，查询来自前一解码器层，键和值来自编码器的输出。这使得解码器的每个位置都能关注输入序列的所有位置，类似于序列到序列模型中的典型编码器-解码器注意力机制[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，所有键、值和查询都来自同一来源，即编码器前一层的输出。编码器的每个位置可以关注编码器前一层的所有位置。
•

Similarly, self-attention layers in the decoder allow each position in the decoder to attend to all positions in the decoder up to and including that position. We need to prevent leftward information flow in the decoder to preserve the auto-regressive property. We implement this inside of scaled dot-product attention by masking out (setting to $-\infty$ ) all values in the input of the softmax which correspond to illegal connections. See Figure 2.类似地，解码器中的自注意力层使得解码器的每个位置可以关注解码器中截至该位置的所有位置。我们需要阻止左侧信息流，以保持自回归属性。我们在缩放点积注意力中通过将非法连接对应的 softmax 输入值设为 -∞ 来实现掩码。见图2。

3.3 Position-wise Feed-Forward Networks3.3 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，编码器和解码器的每一层还包含一个对每个位置独立且相同地应用的全连接前馈网络。该网络由两个线性变换组成，中间使用 ReLU 激活。

\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上相同，但在不同层之间使用不同的参数。也可以将其视为两个 kernel size 为 1 的卷积。输入输出维度为 d_{model}=512，内部层维度为 d_{ff}=2048。

3.4 Embeddings and Softmax3.4 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 $d_{\text{model}}$ . 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 $\sqrt{d_{\text{model}}}$ .与其他序列转导模型类似，我们使用可学习的嵌入将输入和输出标记转换为 d_{model} 维向量。我们还使用常规的可学习线性变换和 softmax 将解码器输出转换为下一个标记的概率。在我们的模型中，两个嵌入层和 pre‑softmax 线性变换共享同一权重矩阵，类似于[30]。在嵌入层中，我们将这些权重乘以 \sqrt{d_{model}}。

3.5 Positional Encoding3.5 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].由于模型没有循环也没有卷积，为了让模型利用序列顺序，需要注入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叠底部的输入嵌入上加入“位置编码”。位置编码的维度与嵌入相同 d_{model}，因此两者可以相加。位置编码有多种选择，包括学习式和固定式[9]。

In this work, we use sine and cosine functions of different frequencies:在本工作中，我们使用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 是位置，i 是维度。即位置编码的每个维度对应一个正弦波。波长从 2π 到 10000·2π 按几何级数递增。我们选择此函数是因为我们假设它能让模型容易学习相对位置，因为对于任意固定偏移 k，PE_{pos+k} 可以表示为 PE_{pos} 的线性函数。

We also experimented with using learned positional embeddings [9] instead, and found that the two versions produced nearly identical results (see Table 3 row (E)). We chose the sinusoidal version because it may allow the model to extrapolate to sequence lengths longer than the ones encountered during training.我们也尝试了学习式位置嵌入[9]，发现两种方式的结果几乎相同（见表3第(E)行）。我们选择正弦式是因为它可能让模型对比训练时更长的序列进行外推。

4 Why Self-Attention4 为什么使用自注意力

In this section we compare various aspects of self-attention layers to the recurrent and convolutional layers commonly used for mapping one variable-length sequence of symbol representations $(x_{1},...,x_{n})$ 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.本节我们比较自注意力层与常用于将一个可变长符号表示序列 (x_1,…,x_n) 映射到等长序列 (z_1,…,z_n) 的循环和卷积层的各方面，xi,zi∈ℝ^d。我们考虑三个动机。

One is the total computational complexity per layer. Another is the amount of computation that can be parallelized, as measured by the minimum number of sequential operations required.其一是每层的总计算复杂度；其二是可并行化的计算量，以所需最小顺序操作数衡量。

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.第三是网络中长程依赖的路径长度。学习长程依赖是许多序列转导任务的关键挑战。影响学习此类依赖的关键因素是信号在网络中前向和后向传播的路径长度。路径越短，学习长程依赖越容易[12]。因此我们还比较了不同层类型网络中任意两个输入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、每层复杂度和最小顺序操作数。n 为序列长度，d 为表示维度，k 为卷积核大小，r 为受限自注意力的邻域大小。

Layer Type	Complexity per Layer	Sequential	Maximum Path Length
		Operations
Self-Attention	$O(n^{2}\cdot d)$	$O(1)$	$O(1)$
Recurrent	$O(n\cdot d^{2})$	$O(n)$	$O(n)$
Convolutional	$O(k\cdot n\cdot d^{2})$	$O(1)$	$O(log_{k}(n))$
Self-Attention (restricted)	$O(r\cdot n\cdot d)$	$O(1)$	$O(n/r)$

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，自注意力层以常数的顺序操作数连接所有位置，而循环层需要 O(n) 的顺序操作。就计算复杂度而言，当序列长度 n 小于表示维度 d 时，自注意力层比循环层更快，这在机器翻译中使用的词片段[38]和字节对[31]表示中最常见。为提升对超长序列任务的计算性能，自注意力可以限制只考虑输入序列中围绕相应输出位置的大小为 r 的邻域。这会将最大路径长度提升至 O(n/r)。我们计划在未来工作中进一步研究此方法。

A single convolutional layer with kernel width $k<n$ does not connect all pairs of input and output positions. Doing so requires a stack of $O(n/k)$ convolutional layers in the case of contiguous kernels, or $O(log_{k}(n))$ in the case of dilated convolutions [18], increasing the length of the longest paths between any two positions in the network. Convolutional layers are generally more expensive than recurrent layers, by a factor of $k$ . Separable convolutions [6], however, decrease the complexity considerably, to $O(k\cdot n\cdot d+n\cdot d^{2})$ . Even with $k=n$ , however, the complexity of a separable convolution is equal to the combination of a self-attention layer and a point-wise feed-forward layer, the approach we take in our model.单个卷积层的核宽度 k<n 并不能连接所有输入输出位置。若要实现，需要 O(n/k) 层的堆叠（连续卷积）或 O(log_k n) 层（空洞卷积）[18]，这会增加网络中任意两位置之间的最长路径。卷积层通常比循环层更昂贵，成本约为 k 倍。可分离卷积[6]则显著降低复杂度至 O(k·n·d + n·d^2)。即使 k=n，可分离卷积的复杂度仍等同于自注意力层加逐点前馈层的组合，这正是我们模型采用的方式。

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，自注意力还能产生更易解释的模型。我们检查了模型的注意力分布，并在附录中展示和讨论示例。许多注意力头不仅学习了不同任务，还表现出与句子句法和语义结构相关的行为。

5 Training5 训练

This section describes the training regime for our models.本节描述我们的模型的训练方案。

5.1 Training Data and Batching5.1 训练数据与批处理

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 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 英德数据集上训练，约 450 万句对。句子使用字节对编码[3]，共享源-目标词表约 37000 个标记。英法任务使用更大的 WMT 2014 英法数据集，约 3600 万句，词表为 32000 词片。句对按近似序列长度分批，每个训练批次约包含 25000 个源标记和 25000 个目标标记。

5.2 Hardware and Schedule5.2 硬件与计划

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 的机器上训练模型。对于本文描述的基础模型，每步约耗时 0.4 秒。基础模型训练 100,000 步（约 12 小时）。对于大模型（表 3 底部行），每步耗时 1.0 秒，训练 300,000 步（约 3.5 天）。

5.3 Optimizer5.3 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:我们使用 Adam 优化器[20]，β₁=0.9，β₂=0.98，ε=10⁻⁹。学习率随训练进程变化，公式如下：

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=4000 步内学习率线性上升，随后按步数的倒数平方根衰减。

5.4 Regularization5.4 正则化

We employ three types of regularization during training:我们在训练期间使用三种正则化方式：

Residual Dropout残差 Dropout

We apply dropout [33] to the output of each sub-layer, before it is added to the sub-layer input and normalized. In addition, we apply dropout to the sums of the embeddings and the positional encodings in both the encoder and decoder stacks. For the base model, we use a rate of $P_{drop}=0.1$ .我们对每个子层的输出在加入子层输入并归一化之前使用 dropout[33]。此外，我们对编码器和解码器堆叠中嵌入与位置编码的求和也使用 dropout。基础模型的 dropout 率为 P_{drop}=0.1。

Label Smoothing标签平滑

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.训练时我们使用标签平滑 ε_{ls}=0.1[36]。这会提升困惑度，因为模型变得更不确定，但能提高准确率和 BLEU 分数。

6 Results6 结果

6.1 Machine Translation6.1 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 测试上取得了比以往最先进模型更高的 BLEU 分数，且训练成本更低。

Model	BLEU		Training Cost (FLOPs)
Model	EN-DE	EN-FR	EN-DE	EN-FR
ByteNet [18]	23.75
Deep-Att + PosUnk [39]		39.2		$1.0\cdot 10^{20}$
GNMT + RL [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]		40.4		$8.0\cdot 10^{20}$
GNMT + RL Ensemble [38]	26.30	41.16	$1.8\cdot 10^{20}$	$1.1\cdot 10^{21}$
ConvS2S Ensemble [9]	26.36	41.29	$7.7\cdot 10^{19}$	$1.2\cdot 10^{21}$
Transformer (base model)	27.3	38.1	$3.3\cdot 10^{18}$
Transformer (big)	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 英德翻译任务上，大型 Transformer（表 2 中的 Transformer (big)）的 BLEU 超过 28.4，较之前最佳模型（包括集成）提升超过 2.0，创下新纪录。该模型配置见表 3 底部行，训练耗时 3.5 天，使用 8 块 P100 GPU。即使是我们的基础模型，也以远低于竞争模型的训练成本超越了所有已发表的模型和集成。

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 英法翻译任务上，我们的大模型取得了 41.0 BLEU，超过所有已发表的单模型，且训练成本不到之前最先进模型的 1/4。大模型在英法任务中使用的 dropout 率为 P_{drop}=0.1，而不是 0.3。

For the base models, we used a single model obtained by averaging the last 5 checkpoints, which were written at 10-minute intervals. For the big models, we averaged the last 20 checkpoints. We used beam search with a beam size of $4$ and length penalty $\alpha=0.6$ [38]. These hyperparameters were chosen after experimentation on the development set. We set the maximum output length during inference to input length + $50$ , but terminate early when possible [38].对于基础模型，我们使用最近 5 个检查点的平均（每 10 分钟保存一次）。对于大模型，平均最近 20 个检查点。我们使用束搜索，束宽为 4，长度惩罚 α=0.6[38]。这些超参数在验证集上实验后确定。推理时最大输出长度设为输入长度 + 50，但在可能时提前终止[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 ²²2We used values of 2.8, 3.7, 6.0 and 9.5 TFLOPS for K80, K40, M40 and P100, respectively..表 2 总结了我们的结果，并将翻译质量和训练成本与文献中的其他模型架构进行比较。我们通过将训练时间、GPU 数量以及每块 GPU 的持续单精度浮点运算能力相乘，估算了训练模型所使用的浮点运算次数（K80、K40、M40、P100 的 TFLOPS 分别为 2.8、3.7、6.0、9.5）。

6.2 Model Variations6.2 模型变体

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 架构的变体。未列出的数值与基础模型相同。所有指标均在 English‑to‑German 翻译验证集 newstest2013 上。列出的困惑度为每个 word‑piece，依据我们的字节对编码，不应与每词困惑度直接比较。

	$N$	$d_{\text{model}}$	$d_{\text{ff}}$	$h$	$d_{k}$	$d_{v}$	$P_{drop}$	$\epsilon_{ls}$	train	PPL	BLEU	params
	$N$	$d_{\text{model}}$	$d_{\text{ff}}$	$h$	$d_{k}$	$d_{v}$	$P_{drop}$	$\epsilon_{ls}$	steps	(dev)	(dev)	$\times 10^{6}$
base	6	512	2048	8	64	64	0.1	0.1	100K	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
(B)					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								4.92	25.7
big	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 各组件的重要性，我们对基础模型进行多种改动，测量在 English‑to‑German 验证集 newstest2013 上的性能变化。使用与前述相同的束搜索，但不进行检查点平均。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.表 3 中的 (A) 行，我们在保持计算量不变的前提下，改变注意力头数及键值维度。如第 3.2.2 节所述。单头注意力比最佳设置低 0.9 BLEU，头数过多也会导致质量下降。

In Table 3 rows (B), we observe that reducing the attention key size $d_{k}$ hurts model quality. This suggests that determining compatibility is not easy and that a more sophisticated compatibility function than dot product may be beneficial. We further observe in rows (C) and (D) that, as expected, bigger models are better, and dropout is very helpful in avoiding over-fitting. In row (E) we replace our sinusoidal positional encoding with learned positional embeddings [9], and observe nearly identical results to the base model.表 3 中的 (B) 行显示，减小注意力键维度 d_k 会削弱模型质量。这表明兼容性判定并不容易，可能需要比点积更复杂的兼容函数。行 (C) 与 (D) 表明，正如预期，更大的模型表现更好，dropout 对防止过拟合非常有帮助。行 (E) 中我们将正弦位置编码替换为学习式位置嵌入[9]，结果与基础模型几乎相同。

6.3 English Constituency Parsing6.3 英语成分句法分析

Table 4: The Transformer generalizes well to English constituency parsing (Results are on Section 23 of WSJ)表 4：Transformer 在英语成分句法分析上表现良好（结果基于 WSJ 第 23 章节）

Parser	Training	WSJ 23 F1
Vinyals & Kaiser el al. (2014) [37]	WSJ only, discriminative	88.3
Petrov et al. (2006) [29]	WSJ only, discriminative	90.4
Zhu et al. (2013) [40]	WSJ only, discriminative	90.4
Dyer et al. (2016) [8]	WSJ only, discriminative	91.7
Transformer (4 layers)	WSJ only, discriminative	91.3
Zhu et al. (2013) [40]	semi-supervised	91.3
Huang & Harper (2009) [14]	semi-supervised	91.3
McClosky et al. (2006) [26]	semi-supervised	92.1
Vinyals & Kaiser el al. (2014) [37]	semi-supervised	92.1
Transformer (4 layers)	semi-supervised	92.7
Luong et al. (2015) [23]	multi-task	93.0
Dyer et al. (2016) [8]	generative	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 能否推广到其他任务，我们在英语成分句法分析上进行实验。该任务具有特定挑战：输出受强结构约束，且长度显著长于输入。此外，RNN 序列到序列模型在小数据场景下未能取得最先进的结果[37]。

We trained a 4-layer transformer with $d_{model}=1024$ on the Wall Street Journal (WSJ) portion of the Penn Treebank [25], about 40K training sentences. We also trained it in a semi-supervised setting, using the larger high-confidence and BerkleyParser corpora from with approximately 17M sentences [37]. We used a vocabulary of 16K tokens for the WSJ only setting and a vocabulary of 32K tokens for the semi-supervised setting.我们在华尔街日报（WSJ）子集的 Penn Treebank 上训练了一个 4 层 Transformer，d_{model}=1024，约 40K 训练句子。我们还在半监督设置下使用更大的高置信度和 BerkeleyParser 语料库进行训练，约 1700 万句子[37]。WSJ 仅设置使用 16K 词表，半监督设置使用 32K 词表。

We performed only a small number of experiments to select the dropout, both attention and residual (section 5.4), learning rates and beam size on the Section 22 development set, all other parameters remained unchanged from the English-to-German base translation model. During inference, we increased the maximum output length to input length + $300$ . We used a beam size of $21$ and $\alpha=0.3$ for both WSJ only and the semi-supervised setting.我们仅在第 22 章节的验证集上进行少量实验，以选择 dropout（注意力和残差）、学习率和束宽，其他参数均保持与 English‑to‑German 基础翻译模型相同。推理时，我们将最大输出长度设为输入长度 + 300。束宽为 2，α=0.3，适用于 WSJ 单独和半监督两种设置。

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 的结果显示，尽管缺乏任务特定调优，我们的模型表现出乎意料地好，优于除 Recurrent Neural Network Grammar[8] 之外的所有已报告模型。

In contrast to RNN sequence-to-sequence models [37], the Transformer outperforms the BerkeleyParser [29] even when training only on the WSJ training set of 40K sentences.与 RNN 序列到序列模型[37] 相比，Transformer 在仅使用 WSJ 40K 句子训练时也超越了 BerkeleyParser[29]。

7 Conclusion7 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.在本工作中，我们提出了 Transformer——首个完全基于注意力的序列转导模型，用多头自注意力取代了编码器‑解码器架构中最常用的循环层。

For translation tasks, the Transformer can be trained significantly faster than architectures based on recurrent or convolutional layers. On both WMT 2014 English-to-German and WMT 2014 English-to-French translation tasks, we achieve a new state of the art. In the former task our best model outperforms even all previously reported ensembles.对于翻译任务，Transformer 的训练速度显著快于基于循环或卷积层的架构。在 WMT 2014 英德和英法翻译任务上，我们均达到了新的最先进水平。前者中，我们的最佳模型甚至超越了所有已报告的集成模型。

We are excited about the future of attention-based models and plan to apply them to other tasks. We plan to extend the Transformer to problems involving input and output modalities other than text and to investigate local, restricted attention mechanisms to efficiently handle large inputs and outputs such as images, audio and video. Making generation less sequential is another research goals of ours.我们对基于注意力的模型的未来充满期待，并计划将其应用于其他任务。我们将把 Transformer 扩展到文本以外的输入输出模态，并研究局部、受限注意力机制，以高效处理图像、音频和视频等大规模输入输出。让生成过程更少顺序化也是我们的研究目标之一。

The code we used to train and evaluate our models is available at https://github.com/tensorflow/tensor2tensor.我们用于训练和评估模型的代码可在 https://github.com/tensorflow/tensor2tensor 获取。

Acknowledgements致谢

We are grateful to Nal Kalchbrenner and Stephan Gouws for their fruitful comments, corrections and inspiration.我们感谢 Nal Kalchbrenner 和 Stephan Gouws 的宝贵评论、纠正和启发。

References参考文献

[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. [1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. 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. [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. [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. [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. [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. [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. [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. [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. [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. [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. 用于图像识别的深度残差学习。发表于 IEEE 计算机视觉与模式识别会议论文集, 第 770–778 页, 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动：学习长期依赖性的难题, 2001.
[13] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735–1780, 1997. [13] Sepp Hochreiter 和 Jürgen Schmidhuber. 长短期记忆。 Neural Computation, 9(8):1735–1780, 1997.
[14] Zhongqiang Huang and Mary Harper. Self-training PCFG grammars with latent annotations across languages. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 832–841. ACL, August 2009. [14] Zhongqiang Huang 和 Mary Harper. 跨语言的潜在注释自训练 PCFG 文法。发表于 2009 年自然语言处理经验方法会议论文集, 第 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. 探索语言模型的极限。 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. 主动记忆能取代注意力机制吗？发表于《神经信息处理系统进展》, (NIPS), 2016.
[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。发表于国际学习表征会议 (ICLR), 2016.
[18] Nal Kalchbrenner, Lasse Espeholt, Karen Simonyan, Aaron van den Oord, Alex Graves, and Koray Kavukcuoglu. Neural machine translation in linear time. arXiv preprint arXiv:1610.10099v2, 2017. [18] Nal Kalchbrenner, Lasse Espeholt, Karen Simonyan, Aaron van den Oord, Alex Graves, 和 Koray Kavukcuoglu. 线性时间内的神经机器翻译。 arXiv 预印本 arXiv:1610.10099v2, 2017.
[19] Yoon Kim, Carl Denton, Luong Hoang, and Alexander M. Rush. Structured attention networks. In International Conference on Learning Representations, 2017. [19] Yoon Kim, Carl Denton, Luong Hoang, 和 Alexander M. Rush. 结构化注意力网络。国际学习表征会议, 2017.
[20] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. [20] Diederik Kingma and Jimmy Ba. Adam：一种随机优化方法。 In ICLR, 2015.
[21] Oleksii Kuchaiev and Boris Ginsburg. Factorization tricks for LSTM networks. arXiv preprint arXiv:1703.10722, 2017. [21] Oleksii Kuchaiev 和 Boris Ginsburg. LSTM 网络的因式分解技巧。 arXiv 预印本 arXiv:1703.10722, 2017.
[22] Zhouhan Lin, Minwei Feng, Cicero Nogueira dos Santos, Mo Yu, Bing Xiang, Bowen Zhou, and Yoshua Bengio. A structured self-attentive sentence embedding. arXiv preprint arXiv:1703.03130, 2017. [22] Zhouhan Lin, Minwei Feng, Cicero Nogueira dos Santos, Mo Yu, Bing Xiang, Bowen Zhou, 和 Yoshua Bengio. 结构化自注意力句子嵌入。 arXiv 预印本 arXiv:1703.03130, 2017.
[23] Minh-Thang Luong, Quoc V. Le, Ilya Sutskever, Oriol Vinyals, and Lukasz Kaiser. Multi-task sequence to sequence learning. arXiv preprint arXiv:1511.06114, 2015. [23] Minh-Thang Luong, Quoc V. Le, Ilya Sutskever, Oriol Vinyals, 和 Lukasz Kaiser. 多任务序列到序列学习。 arXiv 预印本 arXiv:1511.06114, 2015.
[24] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-based neural machine translation. arXiv preprint arXiv:1508.04025, 2015. [24] Minh-Thang Luong, Hieu Pham, 和 Christopher D Manning. 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.
[25] Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. Building a large annotated corpus of english: The penn treebank. Computational linguistics, 19(2):313–330, 1993. [25] Mitchell P Marcus, Mary Ann Marcinkiewicz, 和 Beatrice Santorini. 构建大型英语标注语料库：宾州树库。计算语言学, 19(2):313–330, 1993.
[26] David McClosky, Eugene Charniak, and Mark Johnson. Effective self-training for parsing. In Proceedings of the Human Language Technology Conference of the NAACL, Main Conference, pages 152–159. ACL, June 2006. [26] David McClosky, Eugene Charniak, 和 Mark Johnson. 有效的自训练解析方法。 NAACL 人类语言技术会议主会论文集, 第 152–159 页. 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. 可分解注意力模型。自然语言处理实证方法, 2016.
[28] Romain Paulus, Caiming Xiong, and Richard Socher. A deep reinforced model for abstractive summarization. arXiv preprint arXiv:1705.04304, 2017. [28] Romain Paulus, Caiming Xiong, 和 Richard Socher. 用于抽象式摘要的深度强化模型。 arXiv 预印本 arXiv:1705.04304, 2017.
[29] Slav Petrov, Leon Barrett, Romain Thibaux, and Dan Klein. Learning accurate, compact, and interpretable tree annotation. In Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the ACL, pages 433–440. ACL, July 2006. [29] Slav Petrov, Leon Barrett, Romain Thibaux, 和 Dan Klein. 学习准确、紧凑且可解释的树标注。第 21 届国际计算语言学会议暨第 44 届 ACL 年会论文集, 第 433–440 页. ACL, 2006 年 7 月.
[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. 使用输出嵌入改进语言模型。 arXiv 预印本 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. 基于子词单元的神经机器翻译处理罕见词。 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. 超大神经网络：稀疏门控混合专家层。 arXiv 预印本 arXiv:1701.06538, ber 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：防止神经网络过拟合的简单方法。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。 C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, 和 R. Garnett 编, 神经信息处理系统进展 28, 第 2440–2448 页. 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. 基于神经网络的序列到序列学习。神经信息处理系统进展, 第 3104–3112 页, 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中的 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. 将语法作为外语。神经信息处理系统进展, 2015.
[38] Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google’s neural machine translation system: Bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144, 2016. [38] Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, 等. 谷歌神经机器翻译系统：弥合人机翻译差距。 arXiv 预印本 arXiv:1609.08144, 2016.
[39] Jie Zhou, Ying Cao, Xuguang Wang, Peng Li, and Wei Xu. Deep recurrent models with fast-forward connections for neural machine translation. CoRR, abs/1606.04199, 2016. [39] Jie Zhou, Ying Cao, Xuguang Wang, Peng Li, 和 Wei Xu. 用于神经机器翻译的带快速前向连接的深度循环模型。 CoRR, abs/1606.04199, 2016.
[40] Muhua Zhu, Yue Zhang, Wenliang Chen, Min Zhang, and Jingbo Zhu. Fast and accurate shift-reduce constituent parsing. In Proceedings of the 51st Annual Meeting of the ACL (Volume 1: Long Papers), pages 434–443. ACL, August 2013. [40] Muhua Zhu, Yue Zhang, Wenliang Chen, Min Zhang, 和 Jingbo Zhu. 快速准确的移进归约成分句法分析。第 51 届 ACL 年会论文集（第 1 卷：长文), 第 434–443 页. ACL, 2013 年 8 月.