TTiti的学习笔记
首页 / 专业知识 / 40-References/Papers/transformer - Transformer/01_original.md

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

专业知识 · 40-References/Papers/transformer - Transformer/01_original.md

--- title: "Provided proper attribution is provided, Google hereby grants permission to reproduce the tables and figures in this paper solely for use in journalistic or" aliases: - "Transformer" - "arXiv:1706.03762" source: "https://arxiv.org/abs/1706.03762" arxiv: "1706.03762" created: 2026-07-16 type: paper-translation status: extraction-complete_translation-pending tags: - paper - ml - deep-learning - nlp


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

原文全文

Page 1

<a id="S0001"></a> Source: p.1 S0001

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.

<a id="S0002"></a> Source: p.1 S0002

Attention Is All You Need Ashish Vaswani∗ Noam Shazeer∗ Niki Parmar∗ Jakob Uszkoreit∗ Google Brain Google Brain Google Research Google Research avaswani@google.com noam@google.com nikip@google.com usz@google.com Llion Jones∗ Aidan N.

<a id="S0003"></a> Source: p.1 S0003

Gomez∗ † Łukasz Kaiser∗ Google Research University of Toronto Google Brain llion@google.com aidan@cs.toronto.edu lukaszkaiser@google.com Illia Polosukhin∗ ‡ illia.polosukhin@gmail.com Abstract The dominant sequence transduction models are based on complex recurrent or convolutional neural networks that include an encoder and a decoder.

<a id="S0004"></a> Source: p.1 S0004

The best performing models also connect the encoder and decoder through an attention mechanism.

<a id="S0005"></a> Source: p.1 S0005

We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely.

<a id="S0006"></a> Source: p.1 S0006

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.

<a id="S0007"></a> Source: p.1 S0007

Our model achieves 28.4 BLEU on the WMT 2014 Englishto-German translation task, improving over the existing best results, including ensembles, by over 2 BLEU.

<a id="S0008"></a> Source: p.1 S0008

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.

<a id="S0009"></a> Source: p.1 S0009

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. ∗Equal contribution.

<a id="S0010"></a> Source: p.1 S0010

Jakob proposed replacing RNNs with self-attention and started the effort to evaluate this idea.

<a id="S0011"></a> Source: p.1 S0011

Ashish, with Illia, designed and implemented the first Transformer models and has been crucially involved in every aspect of this work.

<a id="S0012"></a> Source: p.1 S0012

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.

<a id="S0013"></a> Source: p.1 S0013

Niki designed, implemented, tuned and evaluated countless model variants in our original codebase and tensor2tensor.

<a id="S0014"></a> Source: p.1 S0014

Llion also experimented with novel model variants, was responsible for our initial codebase, and efficient inference and visualizations.

<a id="S0015"></a> Source: p.1 S0015

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. 31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA. 3202 guA 2 ]LC.sc[ 7v26730.6071:viXra

Page 2

<a id="S0016"></a> Source: p.2 S0016

1 Introduction 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].

<a id="S0017"></a> Source: p.2 S0017

Numerous efforts have since continued to push the boundaries of recurrent language models and encoder-decoder architectures [38, 24, 15].

<a id="S0018"></a> Source: p.2 S0018

Recurrent models typically factor computation along the symbol positions of the input and output sequences.

<a id="S0019"></a> Source: p.2 S0019

Aligning the positions to steps in computation time, they generate a sequence of hidden states h , as a function of the previous hidden state h and the input for position t.

<a id="S0020"></a> Source: p.2 S0020

This inherently t t−1 sequential nature precludes parallelization within training examples, which becomes critical at longer sequence lengths, as memory constraints limit batching across examples.

<a id="S0021"></a> Source: p.2 S0021

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.

<a id="S0022"></a> Source: p.2 S0022

The fundamental constraint of sequential computation, however, remains.

<a id="S0023"></a> Source: p.2 S0023

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].

<a id="S0024"></a> Source: p.2 S0024

In all but a few cases [27], however, such attention mechanisms are used in conjunction with a recurrent network.

<a id="S0025"></a> Source: p.2 S0025

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.

<a id="S0026"></a> Source: p.2 S0026

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. 2 Background 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.

<a id="S0027"></a> Source: p.2 S0027

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.

<a id="S0028"></a> Source: p.2 S0028

This makes it more difficult to learn dependencies between distant positions [12].

<a id="S0029"></a> Source: p.2 S0029

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.

<a id="S0030"></a> Source: p.2 S0030

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.

<a id="S0031"></a> Source: p.2 S0031

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].

<a id="S0032"></a> Source: p.2 S0032

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

<a id="S0033"></a> Source: p.2 S0033

To the best of our knowledge, however, the Transformer is the first transduction model relying entirely on self-attention to compute representations of its input and output without using sequencealigned RNNs or convolution.

<a id="S0034"></a> Source: p.2 S0034

In the following sections, we will describe the Transformer, motivate self-attention and discuss its advantages over models such as [17, 18] and [9]. 3 Model Architecture Most competitive neural sequence transduction models have an encoder-decoder structure [5, 2, 35].

<a id="S0035"></a> Source: p.2 S0035

Here, the encoder maps an input sequence of symbol representations (x , ..., x ) to a sequence 1 n of continuous representations z = (z , ..., z ).

<a id="S0036"></a> Source: p.2 S0036

Given z, the decoder then generates an output 1 n sequence (y , ..., y ) of symbols one element at a time.

<a id="S0037"></a> Source: p.2 S0037

At each step the model is auto-regressive 1 m [10], consuming the previously generated symbols as additional input when generating the next. 2

Page 3

<a id="S0038"></a> Source: p.3 S0038

Figure 1: The Transformer - model architecture.

<a id="S0039"></a> Source: p.3 S0039

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. 3.1 Encoder and Decoder Stacks Encoder: The encoder is composed of a stack of N = 6 identical layers.

<a id="S0040"></a> Source: p.3 S0040

The first is a multi-head self-attention mechanism, and the second is a simple, positionwise fully connected feed-forward network.

<a id="S0041"></a> Source: p.3 S0041

We employ a residual connection [11] around each of the two sub-layers, followed by layer normalization [1].

<a id="S0042"></a> Source: p.3 S0042

That is, the output of each sub-layer is LayerNorm(x + Sublayer(x)), where Sublayer(x) is the function implemented by the sub-layer itself.

<a id="S0043"></a> Source: p.3 S0043

To facilitate these residual connections, all sub-layers in the model, as well as the embedding layers, produce outputs of dimension d = 512. model Decoder: The decoder is also composed of a stack of N = 6 identical layers.

<a id="S0044"></a> Source: p.3 S0044

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.

<a id="S0045"></a> Source: p.3 S0045

Similar to the encoder, we employ residual connections around each of the sub-layers, followed by layer normalization.

<a id="S0046"></a> Source: p.3 S0046

We also modify the self-attention sub-layer in the decoder stack to prevent positions from attending to subsequent positions.

<a id="S0047"></a> Source: p.3 S0047

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. 3.2 Attention 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.

<a id="S0048"></a> Source: p.3 S0048

The output is computed as a weighted sum 3

Page 4

<a id="S0049"></a> Source: p.4 S0049

Scaled Dot-Product Attention Multi-Head Attention Figure 2: (left) Scaled Dot-Product Attention. (right) Multi-Head Attention consists of several attention layers running in parallel. of the values, where the weight assigned to each value is computed by a compatibility function of the query with the corresponding key. 3.2.1 Scaled Dot-Product Attention We call our particular attention "Scaled Dot-Product Attention" (Figure 2).

<a id="S0050"></a> Source: p.4 S0050

The input consists of queries and keys of dimension d k , a√nd values of dimension d v .

<a id="S0051"></a> Source: p.4 S0051

We compute the dot products of the query with all keys, divide each by d , and apply a softmax function to obtain the weights on the k values.

<a id="S0052"></a> Source: p.4 S0052

In practice, we compute the attention function on a set of queries simultaneously, packed together into a matrix Q.

<a id="S0053"></a> Source: p.4 S0053

The keys and values are also packed together into matrices K and V .

<a id="S0054"></a> Source: p.4 S0054

We compute the matrix of outputs as: QKT Attention(Q, K, V ) = softmax( √ )V (1) d k The two most commonly used attention functions are additive attention [2], and dot-product (multiplicative) attention.

<a id="S0055"></a> Source: p.4 S0055

Dot-product attention is identical to our algorithm, except for the scaling factor of √1 .

<a id="S0056"></a> Source: p.4 S0056

Additive attention computes the compatibility function using a feed-forward network with dk a single hidden layer.

<a id="S0057"></a> Source: p.4 S0057

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.

<a id="S0058"></a> Source: p.4 S0058

While for small values of d the two mechanisms perform similarly, additive attention outperforms k dot product attention without scaling for larger values of d [3].

<a id="S0059"></a> Source: p.4 S0059

We suspect that for large values of k d , the dot products grow large in magnitude, pushing the softmax function into regions where it has k extremely small gradients 4.

<a id="S0060"></a> Source: p.4 S0060

To counteract this effect, we scale the dot products by √1 . dk 3.2.2 Multi-Head Attention Instead of performing a single attention function with d -dimensional keys, values and queries, model we found it beneficial to linearly project the queries, keys and values h times with different, learned linear projections to d , d and d dimensions, respectively.

<a id="S0061"></a> Source: p.4 S0061

On each of these projected versions of k k v queries, keys and values we then perform the attention function in parallel, yielding d -dimensional v 4To 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.

<a id="S0062"></a> Source: p.4 S0062

Then their dot product, q · k = (cid:80)dk q k , has mean 0 and variance d . i=1 i i k 4

Page 5

<a id="S0063"></a> Source: p.5 S0063

These are concatenated and once again projected, resulting in the final values, as depicted in Figure 2.

<a id="S0064"></a> Source: p.5 S0064

Multi-head attention allows the model to jointly attend to information from different representation subspaces at different positions.

<a id="S0065"></a> Source: p.5 S0065

With a single attention head, averaging inhibits this.

<a id="S0066"></a> Source: p.5 S0066

MultiHead(Q, K, V ) = Concat(head , ..., head )W O 1 h where head = Attention(QW Q, KW K, V W V ) i i i i Where the projections are parameter matrices W Q ∈ Rdmodel×dk , W K ∈ Rdmodel×dk , W V ∈ Rdmodel×dv i i i and W O ∈ Rhdv×dmodel.

<a id="S0067"></a> Source: p.5 S0067

In this work we employ h = 8 parallel attention layers, or heads.

<a id="S0068"></a> Source: p.5 S0068

For each of these we use d = d = d /h = 64.

<a id="S0069"></a> Source: p.5 S0069

Due to the reduced dimension of each head, the total computational cost k v model is similar to that of single-head attention with full dimensionality. 3.2.3 Applications of Attention in our Model 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.

<a id="S0070"></a> Source: p.5 S0070

This allows every position in the decoder to attend over all positions in the input sequence.

<a id="S0071"></a> Source: p.5 S0071

This mimics the typical encoder-decoder attention mechanisms in sequence-to-sequence models such as [38, 2, 9]. • The encoder contains self-attention layers.

<a id="S0072"></a> Source: p.5 S0072

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.

<a id="S0073"></a> Source: p.5 S0073

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.

<a id="S0074"></a> Source: p.5 S0074

We need to prevent leftward information flow in the decoder to preserve the auto-regressive property.

<a id="S0075"></a> Source: p.5 S0075

We implement this inside of scaled dot-product attention by masking out (setting to −∞) all values in the input of the softmax which correspond to illegal connections.

<a id="S0076"></a> Source: p.5 S0076

See Figure 2. 3.3 Position-wise Feed-Forward Networks 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.

<a id="S0077"></a> Source: p.5 S0077

This consists of two linear transformations with a ReLU activation in between.

<a id="S0078"></a> Source: p.5 S0078

FFN(x) = max(0, xW + b )W + b (2) 1 1 2 2 While the linear transformations are the same across different positions, they use different parameters from layer to layer.

<a id="S0079"></a> Source: p.5 S0079

Another way of describing this is as two convolutions with kernel size 1.

<a id="S0080"></a> Source: p.5 S0080

The dimensionality of input and output is d = 512, and the inner-layer has dimensionality model d = 2048. ff 3.4 Embeddings and Softmax Similarly to other sequence transduction models, we use learned embeddings to convert the input tokens and output tokens to vectors of dimension d .

<a id="S0081"></a> Source: p.5 S0081

We also use the usual learned linear transformodel mation and softmax function to convert the decoder output to predicted next-token probabilities.

<a id="S0082"></a> Source: p.5 S0082

In our model, we share the same weight matrix between the two embedding layers and the pre-√softmax linear transformation, similar to [30].

<a id="S0083"></a> Source: p.5 S0083

In the embedding layers, we multiply those weights by d . model 5

Page 6

<a id="S0084"></a> Source: p.6 S0084

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.

<a id="S0085"></a> Source: p.6 S0085

Layer Type Complexity per Layer Sequential Maximum Path Length Operations Self-Attention O(n2 · d) O(1) O(1) Recurrent O(n · d2) O(n) O(n) Convolutional O(k · n · d2) O(1) O(log (n)) k Self-Attention (restricted) O(r · n · d) O(1) O(n/r) 3.5 Positional Encoding 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.

<a id="S0086"></a> Source: p.6 S0086

To this end, we add "positional encodings" to the input embeddings at the bottoms of the encoder and decoder stacks.

<a id="S0087"></a> Source: p.6 S0087

The positional encodings have the same dimension d model as the embeddings, so that the two can be summed.

<a id="S0088"></a> Source: p.6 S0088

There are many choices of positional encodings, learned and fixed [9].

<a id="S0089"></a> Source: p.6 S0089

In this work, we use sine and cosine functions of different frequencies: P E = sin(pos/100002i/dmodel) (pos,2i) P E = cos(pos/100002i/dmodel) (pos,2i+1) where pos is the position and i is the dimension.

<a id="S0090"></a> Source: p.6 S0090

That is, each dimension of the positional encoding corresponds to a sinusoid.

<a id="S0091"></a> Source: p.6 S0091

The wavelengths form a geometric progression from 2π to 10000 · 2π.

<a id="S0092"></a> Source: p.6 S0092

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, P E can be represented as a linear function of pos+k P E . 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)).

<a id="S0093"></a> Source: p.6 S0093

We chose the sinusoidal version because it may allow the model to extrapolate to sequence lengths longer than the ones encountered during training. 4 Why Self-Attention 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 , ..., x ) to another sequence of equal length (z , ..., z ), with x , z ∈ Rd, such as a hidden 1 n 1 n i i layer in a typical sequence transduction encoder or decoder.

<a id="S0094"></a> Source: p.6 S0094

Motivating our use of self-attention we consider three desiderata.

<a id="S0095"></a> Source: p.6 S0095

One is the total computational complexity per layer.

<a id="S0096"></a> Source: p.6 S0096

Another is the amount of computation that can be parallelized, as measured by the minimum number of sequential operations required.

<a id="S0097"></a> Source: p.6 S0097

The third is the path length between long-range dependencies in the network.

<a id="S0098"></a> Source: p.6 S0098

Learning long-range dependencies is a key challenge in many sequence transduction tasks.

<a id="S0099"></a> Source: p.6 S0099

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.

<a id="S0100"></a> Source: p.6 S0100

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].

<a id="S0101"></a> Source: p.6 S0101

Hence we also compare the maximum path length between any two input and output positions in networks composed of the different layer types.

<a id="S0102"></a> Source: p.6 S0102

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.

<a id="S0103"></a> Source: p.6 S0103

In terms of computational complexity, self-attention layers are faster than recurrent layers when the sequence 6

Page 7

<a id="S0104"></a> Source: p.7 S0104

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.

<a id="S0105"></a> Source: p.7 S0105

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.

<a id="S0106"></a> Source: p.7 S0106

This would increase the maximum path length to O(n/r).

<a id="S0107"></a> Source: p.7 S0107

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.

<a id="S0108"></a> Source: p.7 S0108

Doing so requires a stack of O(n/k) convolutional layers in the case of contiguous kernels, or O(log (n)) in the case of dilated convolutions [18], increasing the length of the longest paths k between any two positions in the network.

<a id="S0109"></a> Source: p.7 S0109

Convolutional layers are generally more expensive than recurrent layers, by a factor of k.

<a id="S0110"></a> Source: p.7 S0110

Separable convolutions [6], however, decrease the complexity considerably, to O(k · n · d + n · d2).

<a id="S0111"></a> Source: p.7 S0111

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.

<a id="S0112"></a> Source: p.7 S0112

As side benefit, self-attention could yield more interpretable models.

<a id="S0113"></a> Source: p.7 S0113

We inspect attention distributions from our models and present and discuss examples in the appendix.

<a id="S0114"></a> Source: p.7 S0114

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. 5 Training This section describes the training regime for our models. 5.1 Training Data and Batching We trained on the standard WMT 2014 English-German dataset consisting of about 4.5 million sentence pairs.

<a id="S0115"></a> Source: p.7 S0115

Sentences were encoded using byte-pair encoding [3], which has a shared sourcetarget vocabulary of about 37000 tokens.

<a id="S0116"></a> Source: p.7 S0116

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].

<a id="S0117"></a> Source: p.7 S0117

Sentence pairs were batched together by approximate sequence length.

<a id="S0118"></a> Source: p.7 S0118

Each training batch contained a set of sentence pairs containing approximately 25000 source tokens and 25000 target tokens. 5.2 Hardware and Schedule We trained our models on one machine with 8 NVIDIA P100 GPUs.

<a id="S0119"></a> Source: p.7 S0119

For our base models using the hyperparameters described throughout the paper, each training step took about 0.4 seconds.

<a id="S0120"></a> Source: p.7 S0120

We trained the base models for a total of 100,000 steps or 12 hours.

<a id="S0121"></a> Source: p.7 S0121

For our big models,(described on the bottom line of table 3), step time was 1.0 seconds.

<a id="S0122"></a> Source: p.7 S0122

The big models were trained for 300,000 steps (3.5 days). 5.3 Optimizer We used the Adam optimizer [20] with β = 0.9, β = 0.98 and ϵ = 10−9.

<a id="S0123"></a> Source: p.7 S0123

We varied the learning 1 2 rate over the course of training, according to the formula: lrate = d−0.5 · min(step_num−0.5, step_num · warmup_steps−1.5) (3) model 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.

<a id="S0124"></a> Source: p.7 S0124

We used warmup_steps = 4000. 5.4 Regularization We employ three types of regularization during training: 7

Page 8

<a id="S0125"></a> Source: p.8 S0125

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.

<a id="S0126"></a> Source: p.8 S0126

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 · 1020 GNMT + RL [38] 24.6 39.92 2.3 · 1019 1.4 · 1020 ConvS2S [9] 25.16 40.46 9.6 · 1018 1.5 · 1020 MoE [32] 26.03 40.56 2.0 · 1019 1.2 · 1020 Deep-Att + PosUnk Ensemble [39] 40.4 8.0 · 1020 GNMT + RL Ensemble [38] 26.30 41.16 1.8 · 1020 1.1 · 1021 ConvS2S Ensemble [9] 26.36 41.29 7.7 · 1019 1.2 · 1021 Transformer (base model) 27.3 38.1 3.3 · 1018 Transformer (big) 28.4 41.8 2.3 · 1019 Residual Dropout We apply dropout [33] to the output of each sub-layer, before it is added to the sub-layer input and normalized.

<a id="S0127"></a> Source: p.8 S0127

In addition, we apply dropout to the sums of the embeddings and the positional encodings in both the encoder and decoder stacks.

<a id="S0128"></a> Source: p.8 S0128

For the base model, we use a rate of P = 0.1. drop Label Smoothing During training, we employed label smoothing of value ϵ = 0.1 [36].

<a id="S0129"></a> Source: p.8 S0129

This ls hurts perplexity, as the model learns to be more unsure, but improves accuracy and BLEU score. 6 Results 6.1 Machine Translation 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.

<a id="S0130"></a> Source: p.8 S0130

The configuration of this model is listed in the bottom line of Table 3.

<a id="S0131"></a> Source: p.8 S0131

Even our base model surpasses all previously published models and ensembles, at a fraction of the training cost of any of the competitive models.

<a id="S0132"></a> Source: p.8 S0132

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.

<a id="S0133"></a> Source: p.8 S0133

The Transformer (big) model trained for English-to-French used dropout rate P = 0.1, instead of 0.3. drop For the base models, we used a single model obtained by averaging the last 5 checkpoints, which were written at 10-minute intervals.

<a id="S0134"></a> Source: p.8 S0134

For the big models, we averaged the last 20 checkpoints.

<a id="S0135"></a> Source: p.8 S0135

We used beam search with a beam size of 4 and length penalty α = 0.6 [38].

<a id="S0136"></a> Source: p.8 S0136

These hyperparameters were chosen after experimentation on the development set.

<a id="S0137"></a> Source: p.8 S0137

We set the maximum output length during inference to input length + 50, but terminate early when possible [38].

<a id="S0138"></a> Source: p.8 S0138

Table 2 summarizes our results and compares our translation quality and training costs to other model architectures from the literature.

<a id="S0139"></a> Source: p.8 S0139

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. 6.2 Model Variations 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 5We used values of 2.8, 3.7, 6.0 and 9.5 TFLOPS for K80, K40, M40 and P100, respectively. 8

Page 9

<a id="S0140"></a> Source: p.9 S0140

Table 3: Variations on the Transformer architecture.

<a id="S0141"></a> Source: p.9 S0141

Unlisted values are identical to those of the base model.

<a id="S0142"></a> Source: p.9 S0142

All metrics are on the English-to-German translation development set, newstest2013.

<a id="S0143"></a> Source: p.9 S0143

Listed perplexities are per-wordpiece, according to our byte-pair encoding, and should not be compared to per-word perplexities. train PPL BLEU params N d d h d d P ϵ model ff k v drop ls steps (dev) (dev) ×106 base 6 512 2048 8 64 64 0.1 0.1 100K 4.92 25.8 65 1 512 512 5.29 24.9 4 128 128 5.00 25.5 (A) 16 32 32 4.91 25.8 32 16 16 5.01 25.4 16 5.16 25.1 58 (B) 32 5.01 25.4 60 2 6.11 23.7 36 4 5.19 25.3 50 8 4.88 25.5 80 (C) 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 0.0 5.77 24.6 0.2 4.95 25.5 (D) 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.

<a id="S0144"></a> Source: p.9 S0144

We used beam search as described in the previous section, but no checkpoint averaging.

<a id="S0145"></a> Source: p.9 S0145

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.

<a id="S0146"></a> Source: p.9 S0146

While single-head attention is 0.9 BLEU worse than the best setting, quality also drops off with too many heads.

<a id="S0147"></a> Source: p.9 S0147

In Table 3 rows (B), we observe that reducing the attention key size d hurts model quality.

<a id="S0148"></a> Source: p.9 S0148

This k suggests that determining compatibility is not easy and that a more sophisticated compatibility function than dot product may be beneficial.

<a id="S0149"></a> Source: p.9 S0149

We further observe in rows (C) and (D) that, as expected, bigger models are better, and dropout is very helpful in avoiding over-fitting.

<a id="S0150"></a> Source: p.9 S0150

In row (E) we replace our sinusoidal positional encoding with learned positional embeddings [9], and observe nearly identical results to the base model. 6.3 English Constituency Parsing To evaluate if the Transformer can generalize to other tasks we performed experiments on English constituency parsing.

<a id="S0151"></a> Source: p.9 S0151

This task presents specific challenges: the output is subject to strong structural constraints and is significantly longer than the input.

<a id="S0152"></a> Source: p.9 S0152

Furthermore, RNN sequence-to-sequence models have not been able to attain state-of-the-art results in small-data regimes [37].

<a id="S0153"></a> Source: p.9 S0153

We trained a 4-layer transformer with d = 1024 on the Wall Street Journal (WSJ) portion of the model Penn Treebank [25], about 40K training sentences.

<a id="S0154"></a> Source: p.9 S0154

We also trained it in a semi-supervised setting, using the larger high-confidence and BerkleyParser corpora from with approximately 17M sentences [37].

<a id="S0155"></a> Source: p.9 S0155

We used a vocabulary of 16K tokens for the WSJ only setting and a vocabulary of 32K tokens for the semi-supervised setting.

<a id="S0156"></a> Source: p.9 S0156

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.

Page 10

<a id="S0157"></a> Source: p.10 S0157

Table 4: The Transformer generalizes well to English constituency parsing (Results are on Section 23 of WSJ) 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 increased the maximum output length to input length + 300.

<a id="S0158"></a> Source: p.10 S0158

We used a beam size of 21 and α = 0.3 for both WSJ only and the semi-supervised setting.

<a id="S0159"></a> Source: p.10 S0159

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].

<a id="S0160"></a> Source: p.10 S0160

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. 7 Conclusion 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.

<a id="S0161"></a> Source: p.10 S0161

For translation tasks, the Transformer can be trained significantly faster than architectures based on recurrent or convolutional layers.

<a id="S0162"></a> Source: p.10 S0162

On both WMT 2014 English-to-German and WMT 2014 English-to-French translation tasks, we achieve a new state of the art.

<a id="S0163"></a> Source: p.10 S0163

In the former task our best model outperforms even all previously reported ensembles.

<a id="S0164"></a> Source: p.10 S0164

We are excited about the future of attention-based models and plan to apply them to other tasks.

<a id="S0165"></a> Source: p.10 S0165

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.

<a id="S0166"></a> Source: p.10 S0166

Making generation less sequential is another research goals of ours.

<a id="S0167"></a> Source: p.10 S0167

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

<a id="S0168"></a> Source: p.10 S0168

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

<a id="S0169"></a> Source: p.10 S0169

References [1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton.

<a id="S0170"></a> Source: p.10 S0170

Layer normalization. arXiv preprint arXiv:1607.06450, 2016. [2] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio.

<a id="S0171"></a> Source: p.10 S0171

Neural machine translation by jointly learning to align and translate.

<a id="S0172"></a> Source: p.10 S0172

CoRR, abs/1409.0473, 2014. [3] Denny Britz, Anna Goldie, Minh-Thang Luong, and Quoc V.

<a id="S0173"></a> Source: p.10 S0173

Massive exploration of neural machine translation architectures.

<a id="S0174"></a> Source: p.10 S0174

CoRR, abs/1703.03906, 2017. [4] Jianpeng Cheng, Li Dong, and Mirella Lapata.

<a id="S0175"></a> Source: p.10 S0175

Long short-term memory-networks for machine reading. arXiv preprint arXiv:1601.06733, 2016. 10

Page 11

<a id="S0176"></a> Source: p.11 S0176

[5] Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Fethi Bougares, Holger Schwenk, and Yoshua Bengio.

<a id="S0177"></a> Source: p.11 S0177

Learning phrase representations using rnn encoder-decoder for statistical machine translation.

<a id="S0178"></a> Source: p.11 S0178

CoRR, abs/1406.1078, 2014. [6] Francois Chollet.

<a id="S0179"></a> Source: p.11 S0179

Xception: Deep learning with depthwise separable convolutions. arXiv preprint arXiv:1610.02357, 2016. [7] Junyoung Chung, Çaglar Gülçehre, Kyunghyun Cho, and Yoshua Bengio.

<a id="S0180"></a> Source: p.11 S0180

Empirical evaluation of gated recurrent neural networks on sequence modeling.

<a id="S0181"></a> Source: p.11 S0181

CoRR, abs/1412.3555, 2014. [8] Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, and Noah A.

<a id="S0182"></a> Source: p.11 S0182

In Proc. of NAACL, 2016. [9] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N.

<a id="S0183"></a> Source: p.11 S0183

Convolutional sequence to sequence learning. arXiv preprint arXiv:1705.03122v2, 2017. [10] Alex Graves.

<a id="S0184"></a> Source: p.11 S0184

Generating sequences with recurrent neural networks. arXiv preprint arXiv:1308.0850, 2013. [11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun.

<a id="S0185"></a> Source: p.11 S0185

Deep residual learning for image recognition.

<a id="S0186"></a> Source: p.11 S0186

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.

<a id="S0187"></a> Source: p.11 S0187

Gradient flow in recurrent nets: the difficulty of learning long-term dependencies, 2001. [13] Sepp Hochreiter and Jürgen Schmidhuber.

<a id="S0188"></a> Source: p.11 S0188

Neural computation, 9(8):1735–1780, 1997. [14] Zhongqiang Huang and Mary Harper.

<a id="S0189"></a> Source: p.11 S0189

Self-training PCFG grammars with latent annotations across languages.

<a id="S0190"></a> Source: p.11 S0190

In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 832–841.

<a id="S0191"></a> Source: p.11 S0191

ACL, August 2009. [15] Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu.

<a id="S0192"></a> Source: p.11 S0192

Exploring the limits of language modeling. arXiv preprint arXiv:1602.02410, 2016. [16] Łukasz Kaiser and Samy Bengio.

<a id="S0193"></a> Source: p.11 S0193

In Advances in Neural Information Processing Systems, (NIPS), 2016. [17] Łukasz Kaiser and Ilya Sutskever.

<a id="S0194"></a> Source: p.11 S0194

In International Conference on Learning Representations (ICLR), 2016. [18] Nal Kalchbrenner, Lasse Espeholt, Karen Simonyan, Aaron van den Oord, Alex Graves, and Koray Kavukcuoglu.

<a id="S0195"></a> Source: p.11 S0195

Neural machine translation in linear time. arXiv preprint arXiv:1610.10099v2, 2017. [19] Yoon Kim, Carl Denton, Luong Hoang, and Alexander M.

<a id="S0196"></a> Source: p.11 S0196

In International Conference on Learning Representations, 2017. [20] Diederik Kingma and Jimmy Ba.

<a id="S0197"></a> Source: p.11 S0197

Adam: A method for stochastic optimization.

<a id="S0198"></a> Source: p.11 S0198

In ICLR, 2015. [21] Oleksii Kuchaiev and Boris Ginsburg.

<a id="S0199"></a> Source: p.11 S0199

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.

<a id="S0200"></a> Source: p.11 S0200

Le, Ilya Sutskever, Oriol Vinyals, and Lukasz Kaiser.

<a id="S0201"></a> Source: p.11 S0201

Multi-task sequence to sequence learning. arXiv preprint arXiv:1511.06114, 2015. [24] Minh-Thang Luong, Hieu Pham, and Christopher D Manning.

<a id="S0202"></a> Source: p.11 S0202

Effective approaches to attentionbased neural machine translation. arXiv preprint arXiv:1508.04025, 2015. 11

Page 12

<a id="S0203"></a> Source: p.12 S0203

[25] Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini.

<a id="S0204"></a> Source: p.12 S0204

Building a large annotated corpus of english: The penn treebank.

<a id="S0205"></a> Source: p.12 S0205

Computational linguistics, 19(2):313–330, 1993. [26] David McClosky, Eugene Charniak, and Mark Johnson.

<a id="S0206"></a> Source: p.12 S0206

In Proceedings of the Human Language Technology Conference of the NAACL, Main Conference, pages 152–159.

<a id="S0207"></a> Source: p.12 S0207

ACL, June 2006. [27] Ankur Parikh, Oscar Täckström, Dipanjan Das, and Jakob Uszkoreit. A decomposable attention model.

<a id="S0208"></a> Source: p.12 S0208

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.

<a id="S0209"></a> Source: p.12 S0209

Learning accurate, compact, and interpretable tree annotation.

<a id="S0210"></a> Source: p.12 S0210

In Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the ACL, pages 433–440.

<a id="S0211"></a> Source: p.12 S0211

ACL, July 2006. [30] Ofir Press and Lior Wolf.

<a id="S0212"></a> Source: p.12 S0212

Using the output embedding to improve language models. arXiv preprint arXiv:1608.05859, 2016. [31] Rico Sennrich, Barry Haddow, and Alexandra Birch.

<a id="S0213"></a> Source: p.12 S0213

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.

<a id="S0214"></a> Source: p.12 S0214

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.

<a id="S0215"></a> Source: p.12 S0215

Dropout: a simple way to prevent neural networks from overfitting.

<a id="S0216"></a> Source: p.12 S0216

Journal of Machine Learning Research, 15(1):1929–1958, 2014. [34] Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus.

<a id="S0217"></a> Source: p.12 S0217

Garnett, editors, Advances in Neural Information Processing Systems 28, pages 2440–2448.

<a id="S0218"></a> Source: p.12 S0218

Curran Associates, Inc., 2015. [35] Ilya Sutskever, Oriol Vinyals, and Quoc VV Le.

<a id="S0219"></a> Source: p.12 S0219

Sequence to sequence learning with neural networks.

<a id="S0220"></a> Source: p.12 S0220

In Advances in Neural Information Processing Systems, pages 3104–3112, 2014. [36] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna.

<a id="S0221"></a> Source: p.12 S0221

Rethinking the inception architecture for computer vision.

<a id="S0222"></a> Source: p.12 S0222

CoRR, abs/1512.00567, 2015. [37] Vinyals & Kaiser, Koo, Petrov, Sutskever, and Hinton.

<a id="S0223"></a> Source: p.12 S0223

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.

<a id="S0224"></a> Source: p.12 S0224

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.

<a id="S0225"></a> Source: p.12 S0225

Deep recurrent models with fast-forward connections for neural machine translation.

<a id="S0226"></a> Source: p.12 S0226

CoRR, abs/1606.04199, 2016. [40] Muhua Zhu, Yue Zhang, Wenliang Chen, Min Zhang, and Jingbo Zhu.

<a id="S0227"></a> Source: p.12 S0227

Fast and accurate shift-reduce constituent parsing.

<a id="S0228"></a> Source: p.12 S0228

In Proceedings of the 51st Annual Meeting of the ACL (Volume 1: Long Papers), pages 434–443.

Page 13

<a id="S0229"></a> Source: p.13 S0229

Input-Input Layer5 Attention Visualizations tI tI si si ni ni siht siht tirips tirips taht taht a a ytirojam ytirojam fo fo naciremA naciremA stnemnrevog stnemnrevog evah evah dessap dessap wen wen swal swal ecnis ecnis 9002 9002 gnikam gnikam eht eht noitartsiger noitartsiger ro ro gnitov gnitov ssecorp ssecorp erom erom tluciffid tluciffid . . >SOE< >SOE< >dap< >dap< >dap< >dap< >dap< >dap< >dap< >dap< >dap< >dap< >dap< >dap< Figure 3: An example of the attention mechanism following long-distance dependencies in the encoder self-attention in layer 5 of 6.

<a id="S0230"></a> Source: p.13 S0230

Many of the attention heads attend to a distant dependency of the verb ‘making’, completing the phrase ‘making...more difficult’.

<a id="S0231"></a> Source: p.13 S0231

Attentions here shown only for the word ‘making’.

<a id="S0232"></a> Source: p.13 S0232

Different colors represent different heads.

Page 14

<a id="S0233"></a> Source: p.14 S0233

Input-Input Layer5 ehT ehT waL waL lliw lliw reven reven eb eb tcefrep tcefrep , , tub tub sti sti noitacilppa noitacilppa dluohs dluohs eb eb tsuj tsuj - siht siht si si tahw tahw ew ew era era gnissim gnissim , , ni ni ym ym noinipo noinipo . . >SOE< >SOE< >dap< >dap< Input-Input Layer5 ehT ehT waL waL lliw lliw reven reven eb eb tcefrep tcefrep , , tub tub sti sti noitacilppa noitacilppa dluohs dluohs eb eb tsuj tsuj - siht siht si si tahw tahw ew ew era era gnissim gnissim , , ni ni ym ym noinipo noinipo . . >SOE< >SOE< >dap< >dap< Figure 4: Two attention heads, also in layer 5 of 6, apparently involved in anaphora resolution.

<a id="S0234"></a> Source: p.14 S0234

Bottom: Isolated attentions from just the word ‘its’ for attention heads 5 and 6.

<a id="S0235"></a> Source: p.14 S0235

Note that the attentions are very sharp for this word. 14

Page 15

<a id="S0236"></a> Source: p.15 S0236

Input-Input Layer5 ehT ehT waL waL lliw lliw reven reven eb eb tcefrep tcefrep , , tub tub sti sti noitacilppa noitacilppa dluohs dluohs eb eb tsuj tsuj - siht siht si si tahw tahw ew ew era era gnissim gnissim , , ni ni ym ym noinipo noinipo . . >SOE< >SOE< >dap< >dap< Input-Input Layer5 ehT ehT waL waL lliw lliw reven reven eb eb tcefrep tcefrep , , tub tub sti sti noitacilppa noitacilppa dluohs dluohs eb eb tsuj tsuj - siht siht si si tahw tahw ew ew era era gnissim gnissim , , ni ni ym ym noinipo noinipo . . >SOE< >SOE< >dap< >dap< Figure 5: Many of the attention heads exhibit behaviour that seems related to the structure of the sentence.

<a id="S0237"></a> Source: p.15 S0237

We give two such examples above, from two different heads from the encoder self-attention at layer 5 of 6.

<a id="S0238"></a> Source: p.15 S0238

The heads clearly learned to perform different tasks. 15