Provided proper attribution is provided, Google hereby grants permission to reproduce the tables and figures in this paper solely for use in journalistic or scholarly works.
---
在适当归因的前提下，Google特此授权仅出于新闻报道或学术研究目的复制本文中的表格和图表。
---
# Attention Is All You Need
---
# 注意力机制即你所需要的一切
---
\textbf{Ashish Vaswani}$^{*}$
Google Brain
avaswani@google.com
\textbf{Llion Jones}$^{*}$
Google Research
llion@google.com
Noam Shazeer$^*$
Google Brain
noam@google.com
**Aidan N. Gomez**$^{*\ \dagger}$
University of Toronto
aidan@cs.toronto.edu
**Niki Parmar**$^*$
Google Research
nikip@google.com
Jakob Uszkoreit$^{*}$
Google Research
usz@google.com
Łukasz Kaiser*
Google Brain
lukaszkaiser@google.com
\textbf{Illia Polosukhin}\textsuperscript{*} \textdagger
illia.polosukhin@gmail.com
---
\textbf{伊利亚·波洛苏金}\textsuperscript{*} \textdagger
illia.polosukhin@gmail.com
---
## 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.
---
主流的序列转换模型基于复杂的循环神经网络或卷积神经网络，这些网络包含编码器和解码器。表现最佳的模型还通过注意力机制将编码器和解码器连接起来。我们提出了一种全新的简单网络架构——Transformer，它完全基于注意力机制，彻底摒弃了循环和卷积。在两个机器翻译任务上的实验表明，这些模型在质量上更优，同时具有更高的并行性，并且训练时间显著减少。我们的模型在 WMT 2014 英德翻译任务上取得了 28.4 的 BLEU 分数，相比现有最佳结果（包括集成模型）提升了超过 2 个 BLEU 点。在 WMT 2014 英法翻译任务上，我们的模型在八块 GPU 上训练 3.5 天后，建立了新的单模型最先进 BLEU 分数 41.8，这仅占文献中最佳模型训练成本的一小部分。我们通过将该模型成功应用于英语成分句法分析（包括大规模和小规模训练数据），证明了 Transformer 能够很好地泛化到其他任务。
---
arXiv:1706.03762v7 [cs.CL] 2 Aug 2023
---
arXiv:1706.03762v7 [cs.CL] 2023年8月2日
---
$^*$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.
$^\dagger$Work performed while at Google Brain.
$^\ddagger$Work performed while at Google Research.
---
$^*$同等贡献。作者排名顺序为随机排列。Jakob 提出用自注意力机制替代循环神经网络（RNN），并率先启动了评估该想法的工作。Ashish 与 Illia 共同设计并实现了首个 Transformer 模型，并在本工作的各个方面发挥了关键作用。Noam 提出了缩放点积注意力、多头注意力以及无参数位置表示方法，并几乎参与了本工作的所有细节。Niki 在我们的原始代码库和 tensor2tensor 中设计、实现、调优并评估了无数种模型变体。Llion 也尝试了新颖的模型变体，负责我们的初始代码库、高效推理及可视化工作。Lukasz 和 Aidan 花费了大量时间设计和实现 tensor2tensor，取代了我们早期的代码库，大幅提升了结果性能并极大加速了我们的研究进程。
$^\dagger$在 Google Brain 工作期间完成。
$^\ddagger$在 Google Research 工作期间完成。
---
## 1 Introduction
---
## 1 引言
---
Recurrent neural networks, long short-term memory [13] and gated recurrent [7] neural networks in particular, have been firmly established as state of the art approaches in sequence modeling and transduction problems such as language modeling and machine translation [35, 2, 5]. Numerous efforts have since continued to push the boundaries of recurrent language models and encoder-decoder architectures [38, 24, 15].
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.
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.
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.
---
循环神经网络，特别是长短期记忆网络 [13] 和门控循环单元 [7]，已在序列建模和转换问题（如语言建模和机器翻译）中牢固确立了其作为最先进方法的地位 [35, 2, 5]。此后，众多研究持续推动循环语言模型和编码器-解码器架构的边界 [38, 24, 15]。
循环模型通常沿输入和输出序列的符号位置对计算进行分解。将位置与计算时间步对齐，它们以前一隐藏状态 $h_{t-1}$ 和位置 $t$ 的输入为函数，生成一系列隐藏状态 $h_t$。这种固有的顺序性质排除了训练样本内的并行化，而在较长序列长度下这一点尤为关键，因为内存限制限制了跨样本的批处理。最近的工作通过因子化技巧 [21] 和条件计算 [32] 在计算效率方面取得了显著改进，同时在后者的情况下也提升了模型性能。然而，顺序计算的根本约束仍然存在。
注意力机制已成为各种任务中引人注目的序列建模和转换模型的组成部分，允许在不考虑输入或输出序列中距离的情况下建模依赖关系 [2, 19]。然而，除少数情况外 [27]，此类注意力机制通常与循环网络结合使用。
在这项工作中，我们提出了 Transformer，这是一种摒弃循环、完全依赖注意力机制来建立输入和输出之间全局依赖关系的模型架构。Transformer 允许更大幅度的并行化，并且在使用八块 P100 GPU 训练仅十二小时后即可达到翻译质量的新最先进水平。
---
## 2 Background
---
## 2 背景
---
The goal of reducing sequential computation also forms the foundation of the Extended Neural GPU [16], ByteNet [18] and ConvS2S [9], all of which use convolutional neural networks as basic building block, computing hidden representations in parallel for all input and output positions. In these models, the number of operations required to relate signals from two arbitrary input or output positions grows in the distance between positions, linearly for ConvS2S and logarithmically for ByteNet. This makes it more difficult to learn dependencies between distant positions [12]. In the Transformer this is reduced to a constant number of operations, albeit at the cost of reduced effective resolution due to averaging attention-weighted positions, an effect we counteract with Multi-Head Attention as described in section 3.2.
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].
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].
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].
---
减少序列计算的目标也构成了扩展神经GPU [16]、ByteNet [18] 和 ConvS2S [9] 的基础，这些模型均以卷积神经网络作为基本构建模块，并行计算所有输入和输出位置的隐藏表示。在这些模型中，关联任意两个输入或输出位置信号所需的操作数量随位置间距离的增加而增长：ConvS2S 呈线性增长，ByteNet 呈对数增长。这使得学习远距离位置之间的依赖关系变得更加困难 [12]。在 Transformer 中，这一操作数量被减少为常数级，尽管由于平均注意力加权位置导致有效分辨率降低，但我们通过第 3.2 节所述的多头注意力机制来抵消这种影响。
自注意力（有时称为内部注意力）是一种关注机制，用于关联单个序列的不同位置，以计算该序列的表示。自注意力已成功应用于多种任务，包括阅读理解、摘要生成、文本蕴含以及学习任务无关的句子表示 [4, 27, 28, 22]。
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结构，并在简单的语言问答和语言建模任务中表现出良好的性能 [34]。
然而，据我们所知，Transformer 是第一个完全依赖自注意力来计算其输入和输出表示的转换模型，未使用序列对齐的循环神经网络或卷积。在接下来的章节中，我们将描述 Transformer，阐明自注意力的动机，并讨论其相较于 [17, 18] 和 [9] 等模型的优势。
---
## 3 Model Architecture
---
## 3 模型架构
---
Most competitive neural sequence transduction models have an encoder-decoder structure [5, 2, 35]. Here, the encoder maps an input sequence of symbol representations $(x_1, ..., x_n)$ to a sequence of continuous representations $\mathbf{z} = (z_1, ..., z_n)$. Given $\mathbf{z}$, the decoder then generates an output sequence $(y_1, ..., y_m)$ of symbols one element at a time. At each step the model is auto-regressive [10], consuming the previously generated symbols as additional input when generating the next.
---
大多数具有竞争力的神经序列转导模型都采用编码器-解码器结构 [5, 2, 35]。在此结构中，编码器将符号表示的输入序列 $(x_1, ..., x_n)$ 映射为连续表示序列 $\mathbf{z} = (z_1, ..., z_n)$。给定 $\mathbf{z}$ 后，解码器则逐个元素地生成符号输出序列 $(y_1, ..., y_m)$。在每一步中，该模型是自回归的 [10]，即在生成下一个符号时，会将之前生成的符号作为额外输入进行消耗。
---
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 采用上述整体架构，编码器与解码器均使用堆叠的自注意力机制和逐点全连接层，如图1所示，分别位于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 $N = 6$ 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 $\text{LayerNorm}(x + \text{Sublayer}(x))$, where $\text{Sublayer}(x)$ 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 $d_{\text{model}} = 512$.
**Decoder:** The decoder is also composed of a stack of $N = 6$ identical layers. In addition to the two sub-layers in each encoder layer, the decoder inserts a third sub-layer, which performs multi-head attention over the output of the encoder stack. Similar to the encoder, we employ residual connections around each of the sub-layers, followed by layer normalization. We also modify the self-attention sub-layer in the decoder stack to prevent positions from attending to subsequent positions. This masking, combined with fact that the output embeddings are offset by one position, ensures that the predictions for position $i$ can depend only on the known outputs at positions less than $i$.
---
**编码器：** 编码器由 $N = 6$ 个相同层的堆叠组成。每个层包含两个子层。第一个是多头自注意力机制，第二个是简单的逐位置全连接前馈网络。我们在每个子层周围使用残差连接 [11]，随后进行层归一化 [1]。也就是说，每个子层的输出为 $\text{LayerNorm}(x + \text{Sublayer}(x))$，其中 $\text{Sublayer}(x)$ 是该子层本身实现的函数。为了便于这些残差连接，模型中的所有子层以及嵌入层均产生维度为 $d_{\text{model}} = 512$ 的输出。
**解码器：** 解码器也由 $N = 6$ 个相同层的堆叠组成。除了每个编码器层中的两个子层外，解码器还插入了第三个子层，该子层对编码器堆栈的输出执行多头注意力操作。与编码器类似，我们在每个子层周围使用残差连接，随后进行层归一化。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。这种掩码机制，结合输出嵌入偏移一个位置的事实，确保了位置 $i$ 的预测仅依赖于小于 $i$ 的位置处的已知输出。
---
## 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）缩放点积注意力。（右）多头注意力由多个并行运行的注意力层组成。
---
of the values, where the weight assigned to each value is computed by a compatibility function of the query with the corresponding key.
---
这些值的加权平均，其中每个值所分配的权重由查询与相应键之间的兼容性函数计算得出。
---
## 3.2.1 Scaled Dot-Product Attention
---
## 3.2.1 缩放点积注意力
---
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.
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:
$$
\text{Attention}(Q, K, V) = \text{softmax}(\frac{QK^T}{\sqrt{d_k}})V \tag{1}
$$
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.
While for small values of $d_k$ the two mechanisms perform similarly, additive attention outperforms dot product attention without scaling for larger values of $d_k$ [3]. We suspect that for large values of $d_k$, the dot products grow large in magnitude, pushing the softmax function into regions where it has extremely small gradients $^4$. To counteract this effect, we scale the dot products by $\frac{1}{\sqrt{d_k}}$.
---
我们将我们的特定注意力机制称为“缩放点积注意力”（图 2）。输入由维度为 $d_k$ 的查询（queries）和键（keys），以及维度为 $d_v$ 的值（values）组成。我们计算查询与所有键的点积，将每个结果除以 $\sqrt{d_k}$，然后应用 softmax 函数以获得在值上的权重。
在实践中，我们将一组查询同时计算注意力函数，并将它们打包成一个矩阵 $Q$。键和值也被分别打包成矩阵 $K$ 和 $V$。我们按以下方式计算输出矩阵：
$$
\text{Attention}(Q, K, V) = \text{softmax}(\frac{QK^T}{\sqrt{d_k}})V \tag{1}
$$
两种最常用的注意力函数是加性注意力 [2] 和点积（乘法）注意力。除了缩放因子 $\frac{1}{\sqrt{d_k}}$ 之外，点积注意力与我们的算法完全相同。加性注意力使用具有单个隐藏层的前馈网络来计算兼容性函数。尽管两者在理论复杂度上相似，但点积注意力在实践中速度更快且更节省空间，因为它可以使用高度优化的矩阵乘法代码来实现。
虽然对于较小的 $d_k$ 值，这两种机制的表现相似，但对于较大的 $d_k$ 值，未缩放的加性注意力优于点积注意力 [3]。我们推测，对于较大的 $d_k$ 值，点积的幅度会变得很大，从而将 softmax 函数推向其梯度极小的区域 $^4$。为了抵消这种影响，我们将点积乘以 $\frac{1}{\sqrt{d_k}}$ 进行缩放。
---
## 3.2.2 Multi-Head Attention
---
## 3.2.2 多头注意力
---
Instead of performing a single attention function with $d_{\text{model}}$-dimensional keys, values and queries, we found it beneficial to linearly project the queries, keys and values $h$ times with different, learned linear projections to $d_k$, $d_k$ and $d_v$ dimensions, respectively. On each of these projected versions of queries, keys and values we then perform the attention function in parallel, yielding $d_v$-dimensional
---
与使用 $d_{\text{model}}$ 维的键、值和查询执行单个注意力函数不同，我们发现将查询、键和值通过不同的、学习到的线性投影分别映射到 $d_k$、$d_k$ 和 $d_v$ 维度（共 $h$ 次）是有益的。然后，在这些投影后的查询、键和值的每个版本上并行执行注意力函数，从而得到 $d_v$ 维的输出
---
$^4$To illustrate why the dot products get large, assume that the components of $q$ and $k$ are independent random variables with mean 0 and variance 1. Then their dot product, $q \cdot k = \sum_{i=1}^{d_k} q_i k_i$, has mean 0 and variance $d_k$.
---
$^4$为了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。那么它们的点积 $q \cdot k = \sum_{i=1}^{d_k} q_i k_i$ 的均值为 0，方差为 $d_k$。
---
output values. These are concatenated and once again projected, resulting in the final values, as depicted in Figure 2.
Multi-head attention allows the model to jointly attend to information from different representation subspaces at different positions. With a single attention head, averaging inhibits this.
$$
\begin{align}
\text{MultiHead}(Q, K, V) &= \text{Concat}(\text{head}_1, ..., \text{head}_{\text{h}})W^O \tag{*}\\
\text{where head}_i &= \text{Attention}(QW_i^Q, KW_i^K, VW_i^V)
\end{align}
$$
Where the projections are parameter matrices $W_i^Q \in \mathbb{R}^{d_{\text{model}} \times d_k}$, $W_i^K \in \mathbb{R}^{d_{\text{model}} \times d_k}$, $W_i^V \in \mathbb{R}^{d_{\text{model}} \times d_v}$ and $W^O \in \mathbb{R}^{hd_v \times d_{\text{model}}}$.
In this work we employ $h = 8$ parallel attention layers, or heads. For each of these we use $d_k = d_v = d_{\text{model}}/h = 64$. Due to the reduced dimension of each head, the total computational cost is similar to that of single-head attention with full dimensionality.
---
输出值。这些值被拼接在一起并再次进行投影，从而得到最终的值，如图 2 所示。
多头注意力机制使模型能够在不同位置联合关注来自不同表示子空间的信息。如果使用单个注意力头，平均化会抑制这种能力。
$$
\begin{align}
\text{MultiHead}(Q, K, V) &= \text{Concat}(\text{head}_1, ..., \text{head}_{\text{h}})W^O \tag{*}\\
\text{where head}_i &= \text{Attention}(QW_i^Q, KW_i^K, VW_i^V)
\end{align}
$$
其中投影参数矩阵为 $W_i^Q \in \mathbb{R}^{d_{\text{model}} \times d_k}$，$W_i^K \in \mathbb{R}^{d_{\text{model}} \times d_k}$，$W_i^V \in \mathbb{R}^{d_{\text{model}} \times d_v}$ 以及 $W^O \in \mathbb{R}^{hd_v \times d_{\text{model}}}$。
在本工作中，我们采用 $h = 8$ 个并行注意力层（即头）。对于每个头，我们使用 $d_k = d_v = d_{\text{model}}/h = 64$。由于每个头的维度降低，总计算成本与具有全维度的单头注意力相似。
---
## 3.2.3 Applications of Attention in our Model
---
## 3.2.3 注意力机制在我们模型中的应用
---
The Transformer uses multi-head attention in three different ways:
* 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].
* 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.
* 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.
---
Transformer 以三种不同的方式使用多头注意力机制：
* 在“编码器-解码器注意力”层中，查询（queries）来自前一个解码器层，而键（keys）和值（values）来自编码器的输出。这使得解码器中的每个位置都能关注输入序列中的所有位置。这模拟了序列到序列模型（如 [38, 2, 9]）中典型的编码器-解码器注意力机制。
* 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，在本例中即为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中上一层的所有位置。
* 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并包括该位置在内的所有位置。我们需要防止解码器中出现向左的信息流动，以保持自回归特性。我们在缩放点积注意力内部通过掩码操作（将对应于非法连接的 softmax 输入中的所有值设置为 $-\infty$）来实现这一点。参见图 2。
---
## 3.3 Position-wise Feed-Forward Networks
---
## 3.3 逐位置的前馈网络
---
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.
$$
\text{FFN}(x) = \max(0, xW_1 + b_1)W_2 + b_2 \tag{2}
$$
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 $d_{ff} = 2048$.
---
除了注意力子层之外，我们编码器和解码器中的每一层还包含一个全连接前馈网络，该网络分别且相同地应用于每个位置。它由两个线性变换组成，中间夹着一个 ReLU 激活函数。
$$
\text{FFN}(x) = \max(0, xW_1 + b_1)W_2 + b_2 \tag{2}
$$
尽管这些线性变换在不同位置上是相同的，但它们使用不同的参数，并且从一层到另一层也会变化。另一种描述方式是将其视为两个卷积核大小为 1 的卷积操作。输入和输出的维度为 $d_{\text{model}} = 512$，而内部层的维度为 $d_{ff} = 2048$。
---
## 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 $d_{\mathrm{model}}$. 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 $\sqrt{d_{\mathrm{model}}}$.
---
与其他序列转导模型类似，我们使用学习到的嵌入层将输入token和输出token转换为维度为 $d_{\mathrm{model}}$ 的向量。我们还使用通常的学习线性变换和softmax函数，将解码器的输出转换为预测的下一个token的概率。在我们的模型中，我们在两个嵌入层与预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 $\sqrt{d_{\mathrm{model}}}$。
---
Table 1: Maximum path lengths, per-layer complexity and minimum number of sequential operations for different layer types. $n$ is the sequence length, $d$ is the representation dimension, $k$ is the kernel size of convolutions and $r$ the size of the neighborhood in restricted self-attention.
---
表1：不同层类型的最大路径长度、每层复杂度和最小顺序操作数。$n$ 是序列长度，$d$ 是表示维度，$k$ 是卷积的核大小，$r$ 是受限自注意力中的邻域大小。
---
| Layer Type | Complexity per Layer | Sequential Operations | Maximum Path Length |
| :---: | :---: | :---: | :---: |
| 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) | $O(r \cdot n \cdot d)$ | $O(1)$ | $O(n/r)$ |
---
| 层类型 | 每层复杂度 | 顺序操作数 | 最大路径长度 |
| :---: | :---: | :---: | :---: |
| 自注意力 | $O(n^2 \cdot d)$ | $O(1)$ | $O(1)$ |
| 循环 | $O(n \cdot d^2)$ | $O(n)$ | $O(n)$ |
| 卷积 | $O(k \cdot n \cdot d^2)$ | $O(1)$ | $O(log_k(n))$ |
| 自注意力（受限） | $O(r \cdot n \cdot d)$ | $O(1)$ | $O(n/r)$ |
---
## 3.5 Positional Encoding
---
## 3.5 位置编码
---
Since our model contains no recurrence and no convolution, in order for the model to make use of the order of the sequence, we must inject some information about the relative or absolute position of the tokens in the sequence. To this end, we add "positional encodings" to the input embeddings at the bottoms of the encoder and decoder stacks. The positional encodings have the same dimension $d_{\text{model}}$ as the embeddings, so that the two can be summed. There are many choices of positional encodings, learned and fixed [9].
In this work, we use sine and cosine functions of different frequencies:
$$
\begin{align}
PE_{(pos,2i)} &= sin(pos/10000^{2i/d_{\text{model}}}) \\
PE_{(pos,2i+1)} &= cos(pos/10000^{2i/d_{\text{model}}})
\end{align}
$$
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 $PE_{pos}$.
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.
---
由于我们的模型不包含循环和卷积，为了使模型能够利用序列的顺序信息，我们必须注入关于序列中 token 相对或绝对位置的一些信息。为此，我们在编码器堆栈和解码器堆栈底部的输入嵌入中添加“位置编码”。位置编码的维度 $d_{\text{model}}$ 与嵌入的维度相同，因此可以将两者相加。位置编码有许多选择，包括学习得到的和固定的 [9]。
在本工作中，我们使用不同频率的正弦和余弦函数：
$$
\begin{align}
PE_{(pos,2i)} &= sin(pos/10000^{2i/d_{\text{model}}}) \\
PE_{(pos,2i+1)} &= cos(pos/10000^{2i/d_{\text{model}}})
\end{align}
$$
其中 $pos$ 是位置，$i$ 是维度。也就是说，位置编码的每个维度对应一个正弦波。波长从 $2\pi$ 到 $10000 \cdot 2\pi$ 形成几何级数。我们选择这个函数是因为我们假设它能使模型轻松学习基于相对位置的注意力机制，因为对于任何固定的偏移量 $k$，$PE_{pos+k}$ 可以表示为 $PE_{pos}$ 的线性函数。
我们还尝试了使用学习得到的位置嵌入 [9]，发现这两种版本产生的结果几乎相同（见表 3 第 (E) 行）。我们选择了正弦版本，因为它可能使模型能够外推到比训练过程中遇到的更长的序列长度。
---
## 4 Why Self-Attention
---
## 4 为什么使用自注意力
---
In this section we compare various aspects of self-attention layers to the recurrent and convolutional layers commonly used for mapping one variable-length sequence of symbol representations $(x_1, ..., x_n)$ 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.
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.
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.
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
---
在本节中，我们将自注意力层与常用于将符号表示的变长序列 $(x_1, ..., x_n)$ 映射为等长序列 $(z_1, ..., z_n)$ 的循环层和卷积层的各个方面进行比较，其中 $x_i, z_i \in \mathbb{R}^d$，例如典型序列转导编码器或解码器中的隐藏层。为了阐明我们使用自注意力的动机，我们考虑三个期望特性。
其一是每个层的总计算复杂度。其二是可以并行化的计算量，通过所需的最小顺序操作数量来衡量。
其三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。影响学习此类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 traversed 的路径长度。输入和输出序列中任意位置组合之间的路径越短，就越容易学习长程依赖 [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。
如表 1 所示，自注意力层以恒定数量的顺序执行操作连接所有位置，而循环层需要 $O(n)$ 个顺序操作。就计算复杂度而言，当序列
---
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.
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.
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.
---
长度 $n$ 小于表示维度 $d$，这在机器翻译中最先进模型所使用的句子表示中最为常见，例如 word-piece [38] 和 byte-pair [31] 表示。为了提高涉及非常长序列的任务的计算性能，可以将自注意力限制为仅考虑以相应输出位置为中心的输入序列中大小为 $r$ 的邻域。这将使最大路径长度增加到 $O(n/r)$。我们计划在未来的工作中进一步研究这种方法。
单个卷积层，其核宽度 $k < n$，并不能连接所有输入和输出位置对。在连续核的情况下，这需要 $O(n/k)$ 个卷积层的堆叠；而在空洞卷积的情况下 [18]，则需要 $O(\log_k(n))$ 个卷积层，从而增加了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，成本高出 $k$ 倍。然而，可分离卷积 [6] 将复杂度显著降低至 $O(k \cdot n \cdot d + n \cdot d^2)$。即使当 $k = n$ 时，可分离卷积的复杂度也等于自注意力层与逐点前馈层的组合，这正是我们在模型中所采用的方法。
作为附带好处，自注意力可能会产生更具可解释性的模型。我们检查了来自我们模型的注意力分布，并在附录中呈现并讨论了相关示例。不仅各个注意力头明显学会了执行不同的任务，许多还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.
---
本节介绍了我们模型的训练方案。
---
## 5.1 Training Data and Batching
---
## 5.1 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 英德数据集上进行训练，该数据集包含约 450 万对句子。使用字节对编码（byte-pair encoding）[3] 对句子进行编码，其共享的源-目标词汇表大小约为 37,000 个词元。对于英法翻译，我们使用了规模大得多的 WMT 2014 英法数据集，该数据集包含 3600 万句句子，并将词元分割为大小为 32,000 的词片（word-piece）词汇表 [38]。句子对按照近似序列长度进行批量分组。每个训练批次包含一组句子对，其中源端词元数量和目标端词元数量均约为 25,000 个。
---
## **5.2 Hardware and Schedule**
---
## **5.2 硬件与进度安排**
---
We trained our models on one machine with 8 NVIDIA P100 GPUs. For our base models using the hyperparameters described throughout the paper, each training step took about 0.4 seconds. We trained the base models for a total of 100,000 steps or 12 hours. For our big models,(described on the bottom line of table 3), step time was 1.0 seconds. The big models were trained for 300,000 steps (3.5 days).
---
我们在配备8块NVIDIA P100 GPU的单机上训练了我们的模型。对于使用本文所述超参数的基础模型，每个训练步骤耗时约0.4秒。我们总共对基础模型进行了100,000步的训练，耗时12小时。对于大型模型（见表3最后一行），每个步骤耗时为1.0秒。大型模型共训练了300,000步（3.5天）。
---
## 5.3 Optimizer
---
## 5.3 优化器
---
We used the Adam optimizer [20] with $\beta_1 = 0.9$, $\beta_2 = 0.98$ and $\epsilon = 10^{-9}$. We varied the learning rate over the course of training, according to the formula:
$$
lrate = d_{\text{model}}^{-0.5} \cdot \min(step\_num^{-0.5}, step\_num \cdot warmup\_steps^{-1.5}) \tag{3}
$$
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 $warmup\_steps = 4000$.
---
我们使用了 Adam 优化器 [20]，其中 $\beta_1 = 0.9$，$\beta_2 = 0.98$，$\epsilon = 10^{-9}$。在训练过程中，我们根据以下公式调整学习率：
$$
lrate = d_{\text{model}}^{-0.5} \cdot \min(step\_num^{-0.5}, step\_num \cdot warmup\_steps^{-1.5}) \tag{3}
$$
这对应于在前 *warmup_steps* 个训练步骤中线性增加学习率，之后则按步数倒数的平方根成比例地降低学习率。我们使用的 $warmup\_steps = 4000$。
---
## 5.4 Regularization
---
## 5.4 正则化
---
We employ three types of regularization during training:
---
我们在训练过程中采用了三种类型的正则化方法：
---
Table 2: The Transformer achieves better BLEU scores than previous state-of-the-art models on the English-to-German and English-to-French newstest2014 tests at a fraction of the training cost.
---
表2：在英语到德语和英语到法语的newstest2014测试中，Transformer以远低于以往最先进模型的训练成本，取得了更好的BLEU分数。
---
| Model | BLEU | | Training Cost (FLOPs) | |
| :--- | :---: | :---: | :---: | :---: |
| | 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 | $\mathbf{3.3 \cdot 10^{18}}$ | |
| Transformer (big) | **28.4** | **41.8** | $2.3 \cdot 10^{19}$ | |
**Residual Dropout** We apply dropout [33] to the output of each sub-layer, before it is added to the sub-layer input and normalized. In addition, we apply dropout to the sums of the embeddings and the positional encodings in both the encoder and decoder stacks. For the base model, we use a rate of $P_{drop} = 0.1$.
---
**残差 Dropout** 我们在每个子层的输出上应用 dropout [33]，然后再将其与子层输入相加并进行归一化。此外，我们在编码器堆栈和解码器堆栈中，对嵌入（embeddings）和位置编码的和也应用 dropout。对于基础模型，我们使用的丢弃率为 $P_{drop} = 0.1$。
---
**Label Smoothing** 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.
---
**标签平滑** 在训练过程中，我们采用了值为 $\epsilon_{ls} = 0.1$ 的标签平滑 [36]。这会导致困惑度升高，因为模型学会了更加不确定，但提高了准确率和 BLEU 分数。
---
## 6 Results
---
## 6 结果
---
## 6.1 Machine Translation
---
## 6.1 机器翻译
---
On the WMT 2014 English-to-German translation task, the big transformer model (Transformer (big) in Table 2) outperforms the best previously reported models (including ensembles) by more than 2.0 BLEU, establishing a new state-of-the-art BLEU score of 28.4. The configuration of this model is listed in the bottom line of Table 3. Training took 3.5 days on 8 P100 GPUs. Even our base model surpasses all previously published models and ensembles, at a fraction of the training cost of any of the competitive models.
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.
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].
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 $^5$.
---
在 WMT 2014 英译德翻译任务上，大型 Transformer 模型（表 2 中的 Transformer (big)）比此前报道的最佳模型（包括集成模型）高出超过 2.0 BLEU，确立了 28.4 的新最先进 BLEU 分数。该模型的配置列于表 3 的最后一行。在 8 块 P100 GPU 上训练耗时 3.5 天。即使是我们的基础模型，也超越了所有此前发表的模型和集成模型，且训练成本仅为任何竞争性模型的一小部分。
在 WMT 2014 英译法翻译任务上，我们的大型模型取得了 41.0 的 BLEU 分数，优于所有此前发表的单模型，且训练成本不到之前最先进模型的 $1/4$。用于英译法训练的 Transformer (big) 模型采用了 $P_{drop} = 0.1$ 的丢弃率，而非 0.3。
对于基础模型，我们使用通过平均最后 5 个检查点得到的单个模型，这些检查点每隔 10 分钟保存一次。对于大型模型，我们平均了最后 20 个检查点。我们使用了束搜索，束宽为 4，长度惩罚系数 $\alpha = 0.6$ [38]。这些超参数是在开发集上进行实验后确定的。我们将推理过程中的最大输出长度设置为输入长度 + 50，但在可能时提前终止 [38]。
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的其他模型架构进行了比较。我们通过将训练时间、使用的 GPU 数量以及每个 GPU 的持续单精度浮点计算能力的估计值相乘，来估算训练模型所使用的浮点运算次数 $^5$。
---
## 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 中不同组件的重要性，我们以多种方式调整了基础模型，并测量了在英语到德语翻译任务上的性能变化
---
$^5$We used values of 2.8, 3.7, 6.0 and 9.5 TFLOPS for K80, K40, M40 and P100, respectively.
---
$^5$我们分别使用2.8、3.7、6.0和9.5 TFLOPS作为K80、K40、M40和P100的值。
---
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。所有指标均在英语到德语翻译的开发集newstest2013上计算。列出的困惑度是基于字节对编码（byte-pair encoding）的词元（wordpiece）级别，不应与基于单词的困惑度进行比较。
---
| | $N$ | $d_{\text{model}}$ | $d_{\text{ff}}$ | $h$ | $d_k$ | $d_v$ | $P_{drop}$ | $\epsilon_{ls}$ | train steps | PPL (dev) | BLEU (dev) | params $\times 10^6$ |
| :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
| 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 |
| | | | | 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 |
development set, newstest2013. We used beam search as described in the previous section, but no checkpoint averaging. We present these results in Table 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.
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.
---
开发集为 newstest2013。我们使用了上一节中描述的束搜索方法，但未进行检查点平均。我们将这些结果展示在表 3 中。
在表 3 的行 (A) 中，我们改变了注意力头的数量以及注意力键和值的维度，同时保持计算量不变，如第 3.2.2 节所述。虽然单头注意力的 BLEU 分数比最佳设置低 0.9，但注意力头过多也会导致质量下降。
在表 3 的行 (B) 中，我们观察到减少注意力键的大小 $d_k$ 会损害模型质量。这表明确定兼容性并不容易，且使用比点积更复杂的兼容性函数可能有益。我们在行 (C) 和 (D) 中进一步观察到，正如预期，更大的模型效果更好，而 dropout 在避免过拟合方面非常有效。在行 (E) 中，我们用学习到的位置嵌入 [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].
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.
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
---
为了评估 Transformer 是否能够泛化到其他任务，我们在英语成分句法分析（constituency parsing）上进行了实验。该任务具有特定的挑战：输出受到严格的结构性约束，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]。
我们在 Penn Treebank [25] 的华尔街日报（WSJ）部分上训练了一个 4 层 Transformer，其中 $d_{model} = 1024$，包含约 4 万个训练句子。我们还在一个半监督设置下进行了训练，使用了来自高置信度和 BerkeleyParser 语料库的更大规模数据，大约包含 1700 万个句子 [37]。对于仅使用 WSJ 的设置，我们使用了 16K 词元的词汇表；对于半监督设置，我们使用了 32K 词元的词汇表。
我们在第 22 节的开发集上仅进行了少量实验以选择 dropout（包括注意力机制和残差连接，见第 5.4 节）、学习率和束搜索大小，所有其他参数均与英德基础翻译模型保持一致。在推理过程中，我们
---
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中泛化效果良好（结果来自WSJ的第23节）
---
| Parser | Training | WSJ 23 F1 |
| :---: | :---: | :---: |
| Vinyals & Kaiser et 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 et 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 |
---
| 解析器 | 训练方式 | WSJ 23 F1 |
| :---: | :---: | :---: |
| Vinyals & Kaiser 等 (2014) [37] | 仅使用 WSJ，判别式 | 88.3 |
| Petrov 等 (2006) [29] | 仅使用 WSJ，判别式 | 90.4 |
| Zhu 等 (2013) [40] | 仅使用 WSJ，判别式 | 90.4 |
| Dyer 等 (2016) [8] | 仅使用 WSJ，判别式 | 91.7 |
| Transformer（4 层） | 仅使用 WSJ，判别式 | 91.3 |
| Zhu 等 (2013) [40] | 半监督 | 91.3 |
| Huang & Harper (2009) [14] | 半监督 | 91.3 |
| McClosky 等 (2006) [26] | 半监督 | 92.1 |
| Vinyals & Kaiser 等 (2014) [37] | 半监督 | 92.1 |
| Transformer（4 层） | 半监督 | 92.7 |
| Luong 等 (2015) [23] | 多任务 | 93.0 |
| Dyer 等 (2016) [8] | 生成式 | 93.3 |
---
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.
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].
In contrast to RNN sequence-to-sequence models [37], the Transformer outperforms the Berkeley-Parser [29] even when training only on the WSJ training set of 40K sentences.
---
将最大输出长度增加至输入长度加 300。在仅使用 WSJ 数据和半监督设置中，我们均使用了束大小为 21 且 $\alpha = 0.3$。
表 4 中的结果表明，尽管缺乏针对特定任务的调优，我们的模型表现却出乎意料地好，其结果优于除循环神经网络语法 [8] 之外的所有先前报告的模型。
与 RNN 序列到序列模型 [37] 相比，Transformer 即使在仅在包含 40K 句子的 WSJ 训练集上进行训练的情况下，也优于 Berkeley-Parser [29]。
---
## 7 Conclusion
---
## 7 结论
---
In this work, we presented the Transformer, the first sequence transduction model based entirely on attention, replacing the recurrent layers most commonly used in encoder-decoder architectures with multi-headed self-attention.
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.
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.
The code we used to train and evaluate our models is available at https://github.com/tensorflow/tensor2tensor.
**Acknowledgements** We are grateful to Nal Kalchbrenner and Stephan Gouws for their fruitful comments, corrections and inspiration.
---
在这项工作中，我们提出了 Transformer，这是一种完全基于注意力的序列转导模型，用多头自注意力机制替换了编码器-解码器架构中最常用的循环层。
在翻译任务中，Transformer 的训练速度显著快于基于循环或卷积层的架构。在 WMT 2014 英德翻译任务和 WMT 2014 英法翻译任务上，我们都取得了新的最先进水平。在前者任务中，我们的最佳模型甚至优于此前报道的所有集成模型。
我们对基于注意力的模型的未来充满期待，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输入和输出模态的问题，并研究局部受限注意力机制，以高效处理图像、音频和视频等大型输入和输出。减少生成的序列性也是我们的另一个研究目标。
用于训练和评估我们模型的代码可在 https://github.com/tensorflow/tensor2tensor 获取。
**致谢** 我们要感谢 Nal Kalchbrenner 和 Stephan Gouws 提供的富有成效的评论、修正和启发。
---
## References
---
## 参考文献
---
[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. *arXiv preprint arXiv:1607.06450*, 2016.
[2] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. *CoRR*, abs/1409.0473, 2014.
[3] Denny Britz, Anna Goldie, Minh-Thang Luong, and Quoc V. Le. Massive exploration of neural machine translation architectures. *CoRR*, abs/1703.03906, 2017.
[4] Jianpeng Cheng, Li Dong, and Mirella Lapata. Long short-term memory-networks for machine reading. *arXiv preprint arXiv:1601.06733*, 2016.
---
[1] Jimmy Lei Ba, Jamie Ryan Kiros 和 Geoffrey E Hinton。层归一化。*arXiv 预印本 arXiv:1607.06450*, 2016.
[2] Dzmitry Bahdanau, Kyunghyun Cho 和 Yoshua Bengio。通过联合学习对齐与翻译实现神经机器翻译。*CoRR*, abs/1409.0473, 2014.
[3] Denny Britz, Anna Goldie, Minh-Thang Luong 和 Quoc V. Le。对神经机器翻译架构的大规模探索。*CoRR*, abs/1703.03906, 2017.
[4] Jianpeng Cheng, Li Dong 和 Mirella Lapata。用于机器阅读的长短期记忆网络。*arXiv 预印本 arXiv:1601.06733*, 2016.
---
[5] Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. *CoRR*, abs/1406.1078, 2014.
[6] Francois Chollet. Xception: Deep learning with depthwise separable convolutions. *arXiv preprint arXiv:1610.02357*, 2016.
[7] Junyoung Chung, Çağlar Gülçehre, Kyunghyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. *CoRR*, abs/1412.3555, 2014.
[8] Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, and Noah A. Smith. Recurrent neural network grammars. In *Proc. of NAACL*, 2016.
[9] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N. Dauphin. Convolutional sequence to sequence learning. *arXiv preprint arXiv:1705.03122v2*, 2017.
[10] Alex Graves. Generating sequences with recurrent neural networks. *arXiv preprint arXiv:1308.0850*, 2013.
[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In *Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition*, pages 770–778, 2016.
[12] Sepp Hochreiter, Yoshua Bengio, Paolo Frasconi, and Jürgen Schmidhuber. Gradient flow in recurrent nets: the difficulty of learning long-term dependencies, 2001.
[13] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. *Neural computation*, 9(8):1735–1780, 1997.
[14] Zhongqiang Huang and Mary Harper. Self-training PCFG grammars with latent annotations across languages. In *Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing*, pages 832–841. ACL, August 2009.
[15] Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. Exploring the limits of language modeling. *arXiv preprint arXiv:1602.02410*, 2016.
[16] Łukasz Kaiser and Samy Bengio. Can active memory replace attention? In *Advances in Neural Information Processing Systems, (NIPS)*, 2016.
[17] Łukasz Kaiser and Ilya Sutskever. Neural GPUs learn algorithms. In *International Conference on Learning Representations (ICLR)*, 2016.
[18] Nal Kalchbrenner, Lasse Espeholt, Karen Simonyan, Aaron van den Oord, Alex Graves, and Koray Kavukcuoglu. Neural machine translation in linear time. *arXiv preprint arXiv:1610.10099v2*, 2017.
[19] Yoon Kim, Carl Denton, Luong Hoang, and Alexander M. Rush. Structured attention networks. In *International Conference on Learning Representations*, 2017.
[20] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In *ICLR*, 2015.
[21] Oleksii Kuchaiev and Boris Ginsburg. Factorization tricks for LSTM networks. *arXiv preprint arXiv:1703.10722*, 2017.
[22] Zhouhan Lin, Minwei Feng, Cicero Nogueira dos Santos, Mo Yu, Bing Xiang, Bowen Zhou, and Yoshua Bengio. A structured self-attentive sentence embedding. *arXiv preprint arXiv:1703.03130*, 2017.
[23] Minh-Thang Luong, Quoc V. Le, Ilya Sutskever, Oriol Vinyals, and Lukasz Kaiser. Multi-task sequence to sequence learning. *arXiv preprint arXiv:1511.06114*, 2015.
[24] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-based neural machine translation. *arXiv preprint arXiv:1508.04025*, 2015.
---
[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以用于统计机器翻译。*CoRR*, abs/1406.1078, 2014.
[6] Francois Chollet。Xception：基于深度可分离卷积的深度学习。*arXiv preprint arXiv:1610.02357*, 2016.
[7] Junyoung Chung, Çağlar Gülçehre, Kyunghyun Cho, 和 Yoshua Bengio。门控循环神经网络在序列建模上的实证评估。*CoRR*, abs/1412.3555, 2014.
[8] Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, 和 Noah A. Smith。递归神经语法。在 *Proc. of NAACL*, 2016.
[9] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, 和 Yann N. Dauphin。卷积序列到序列学习。*arXiv preprint arXiv:1705.03122v2*, 2017.
[10] Alex Graves。使用递归神经网络生成序列。*arXiv preprint arXiv:1308.0850*, 2013.
[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, 和 Jian Sun。用于图像识别的深度残差学习。在 *Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition*, pages 770–778, 2016.
[12] Sepp Hochreiter, Yoshua Bengio, Paolo Frasconi, 和 Jürgen Schmidhuber。递归网络中的梯度流：学习长期依赖关系的困难, 2001.
[13] Sepp Hochreiter 和 Jürgen Schmidhuber。长短期记忆。*Neural computation*, 9(8):1735–1780, 1997.
[14] Zhongqiang Huang 和 Mary Harper。跨语言的自训练 PCFG 语法与潜在标注。在 *Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing*, pages 832–841. ACL, August 2009.
[15] Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, 和 Yonghui Wu。探索语言模型的极限。*arXiv preprint arXiv:1602.02410*, 2016.
[16] Łukasz Kaiser 和 Samy Bengio。活动记忆能否取代注意力机制？在 *Advances in Neural Information Processing Systems, (NIPS)*, 2016.
[17] Łukasz Kaiser 和 Ilya Sutskever。神经 GPU 学习算法。在 *International Conference on Learning Representations (ICLR)*, 2016.
[18] Nal Kalchbrenner, Lasse Espeholt, Karen Simonyan, Aaron van den Oord, Alex Graves, 和 Koray Kavukcuoglu。线性时间的神经机器翻译。*arXiv preprint arXiv:1610.10099v2*, 2017.
[19] Yoon Kim, Carl Denton, Luong Hoang, 和 Alexander M. Rush。结构化注意力网络。在 *International Conference on Learning Representations*, 2017.
[20] Diederik Kingma 和 Jimmy Ba。Adam：一种随机优化方法。在 *ICLR*, 2015.
[21] Oleksii Kuchaiev 和 Boris Ginsburg。LSTM 网络的因式分解技巧。*arXiv preprint arXiv:1703.10722*, 2017.
[22] Zhouhan Lin, Minwei Feng, Cicero Nogueira dos Santos, Mo Yu, Bing Xiang, Bowen Zhou, 和 Yoshua Bengio。一种结构化的自注意力句子嵌入。*arXiv preprint arXiv:1703.03130*, 2017.
[23] Minh-Thang Luong, Quoc V. Le, Ilya Sutskever, Oriol Vinyals, 和 Lukasz Kaiser。多任务序列到序列学习。*arXiv preprint arXiv:1511.06114*, 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 preprint arXiv:1508.04025*, 2015.
---
[25] Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. Building a large annotated corpus of english: The penn treebank. *Computational linguistics*, 19(2):313–330, 1993.
[26] David McClosky, Eugene Charniak, and Mark Johnson. Effective self-training for parsing. In *Proceedings of the Human Language Technology Conference of the NAACL, Main Conference*, pages 152–159. ACL, June 2006.
[27] Ankur Parikh, Oscar Täckström, Dipanjan Das, and Jakob Uszkoreit. A decomposable attention model. In *Empirical Methods in Natural Language Processing*, 2016.
[28] Romain Paulus, Caiming Xiong, and Richard Socher. A deep reinforced model for abstractive summarization. *arXiv preprint arXiv:1705.04304*, 2017.
[29] Slav Petrov, Leon Barrett, Romain Thibaux, and Dan Klein. Learning accurate, compact, and interpretable tree annotation. In *Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the ACL*, pages 433–440. ACL, July 2006.
[30] Ofir Press and Lior Wolf. Using the output embedding to improve language models. *arXiv preprint arXiv:1608.05859*, 2016.
[31] Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units. *arXiv preprint arXiv:1508.07909*, 2015.
[32] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. *arXiv preprint arXiv:1701.06538*, 2017.
[33] Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. *Journal of Machine Learning Research*, 15(1):1929–1958, 2014.
[34] Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. End-to-end memory networks. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors, *Advances in Neural Information Processing Systems 28*, pages 2440–2448. Curran Associates, Inc., 2015.
[35] Ilya Sutskever, Oriol Vinyals, and Quoc VV Le. Sequence to sequence learning with neural networks. In *Advances in Neural Information Processing Systems*, pages 3104–3112, 2014.
[36] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. *CoRR*, abs/1512.00567, 2015.
[37] Vinyals & Kaiser, Koo, Petrov, Sutskever, and Hinton. Grammar as a foreign language. In *Advances in Neural Information Processing Systems*, 2015.
[38] Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google’s neural machine translation system: Bridging the gap between human and machine translation. *arXiv preprint arXiv:1609.08144*, 2016.
[39] Jie Zhou, Ying Cao, Xuguang Wang, Peng Li, and Wei Xu. Deep recurrent models with fast-forward connections for neural machine translation. *CoRR*, abs/1606.04199, 2016.
[40] Muhua Zhu, Yue Zhang, Wenliang Chen, Min Zhang, and Jingbo Zhu. Fast and accurate shift-reduce constituent parsing. In *Proceedings of the 51st Annual Meeting of the ACL (Volume 1: Long Papers)*, pages 434–443. ACL, August 2013.
---
[25] Mitchell P Marcus, Mary Ann Marcinkiewicz, 和 Beatrice Santorini. 构建大型英语标注语料库：Penn树库。*计算语言学*, 19(2):313–330, 1993.
[26] David McClosky, Eugene Charniak, 和 Mark Johnson. 用于句法分析的有效自训练方法。在 *NAACL人类语言技术会议主会议论文集* 中，页码 152–159。ACL, 2006年6月.
[27] Ankur Parikh, Oscar Täckström, Dipanjan Das, 和 Jakob Uszkoreit. 可分解注意力模型。在 *自然语言处理中的实证方法* 中, 2016.
[28] Romain Paulus, Caiming Xiong, 和 Richard Socher. 用于摘要生成的深度强化模型。*arXiv预印本 arXiv:1705.04304*, 2017.
[29] Slav Petrov, Leon Barrett, Romain Thibaux, 和 Dan Klein. 学习准确、紧凑且可解释的树标注。在 *第21届国际计算语言学与第44届ACL年会论文集* 中，页码 433–440。ACL, 2006年7月.
[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, 和 Alexandra Birch. 使用子词单元进行稀有词的神经机器翻译。*arXiv预印本 arXiv:1508.07909*, 2015.
[32] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, 和 Jeff Dean. 极其庞大的神经网络：稀疏门控混合专家层。*arXiv预印本 arXiv:1701.06538*, 2017.
[33] Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, 和 Ruslan Salakhutdinov. Dropout：防止神经网络过拟合的一种简单方法。*机器学习研究期刊*, 15(1):1929–1958, 2014.
[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 编辑的 *神经信息处理系统进展 28* 中，页码 2440–2448。Curran Associates, Inc., 2015.
[35] Ilya Sutskever, Oriol Vinyals, 和 Quoc VV Le. 使用神经网络进行序列到序列学习。在 *神经信息处理系统进展* 中，页码 3104–3112, 2014.
[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, 和 Hinton. 语法作为外语。在 *神经信息处理系统进展* 中, 2015.
[38] Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, 等人。Google神经机器翻译系统：弥合人工翻译与机器翻译之间的差距。*arXiv预印本 arXiv:1609.08144*, 2016.
[39] Jie Zhou, Ying Cao, Xuguang Wang, Peng Li, 和 Wei Xu. 具有前向连接深度循环模型的神经机器翻译。*CoRR*, abs/1606.04199, 2016.
[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。在 *ACL第51届年会论文集（第一卷：长文）* 中，页码 434–443。ACL, 2013年8月.
---
---
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：编码器第5层（共6层）自注意力机制中遵循长距离依赖的一个示例。许多注意力头关注动词“making”的远距离依赖，从而补全短语“making...more difficult”。此处仅展示单词“making”的注意力分布。不同颜色代表不同的注意力头。建议使用彩色查看效果更佳。
---
Figure 4: Two attention heads, also in layer 5 of 6, apparently involved in anaphora resolution. Top: Full attentions for head 5. Bottom: Isolated attentions from just the word 'its' for attention heads 5 and 6. Note that the attentions are very sharp for this word.
---
图4：两个注意力头，同样位于第5层（共6层），显然参与了解释回指。顶部：头5的完整注意力分布。底部：仅从单词“its”中提取的注意力头5和6的孤立注意力分布。请注意，该词的注意力分布非常尖锐。
---
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的行为似乎与句子结构相关。我们在上面给出了两个这样的例子，分别来自第 $5$ 层（共 $6$ 层）编码器自注意力的两个不同注意力头。这些注意力头显然学会了执行不同的任务。
---