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