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.

---

Desde que seja dada a devida atribuição, o Google concede aqui permissão para reproduzir as tabelas e figuras deste artigo exclusivamente para uso em trabalhos jornalísticos ou acadêmicos.

---

# Attention Is All You Need

---

# A Atenção é Tudo o que Você Precisa

---

**Ashish Vaswani**$^*$
Google Brain
avaswani@google.com

**Llion Jones**$^*$
Google Research
llion@google.com

Noam Shazeer$^*$
Google Brain
noam@google.com

**Aidan N. Gomez**$^{* \; \dagger}$
University of Toronto
aidan@cs.toronto.edu

\textbf{Niki Parmar}$^*$

Google Research

nikip@google.com

Illia Polosukhin$^{* \; \ddagger}$

illia.polosukhin@gmail.com

Jakob Uszkoreit$^{*}$
Google Research
usz@google.com

Łukasz Kaiser*
Google Brain
lukaszkaiser@google.com

## Abstract

---

## Resumo

---

The dominant sequence transduction models are based on complex recurrent or convolutional neural networks that include an encoder and a decoder. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles, by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data.

---

Os modelos dominantes de transdução de sequências são baseados em redes neurais recorrentes ou convolucionais complexas que incluem um codificador e um decodificador. Os modelos com melhor desempenho também conectam o codificador e o decodificador por meio de um mecanismo de atenção. Propomos uma nova arquitetura de rede simples, o Transformer, baseada exclusivamente em mecanismos de atenção, dispensando completamente a recorrência e as convoluções. Experimentos em duas tarefas de tradução automática mostram que esses modelos são superiores em qualidade, além de serem mais paralelizáveis e exigirem significativamente menos tempo para treinamento. Nosso modelo alcançou 28,4 BLEU na tarefa de tradução inglês-alemão do WMT 2014, superando os melhores resultados existentes, incluindo conjuntos de modelos, em mais de 2 BLEU. Na tarefa de tradução inglês-francês do WMT 2014, nosso modelo estabeleceu um novo estado da arte para modelos únicos com pontuação BLEU de 41,8 após três dias e meio de treinamento em oito GPUs, uma pequena fração dos custos de treinamento dos melhores modelos da literatura. Mostramos que o Transformer generaliza bem para outras tarefas ao aplicá-lo com sucesso à análise sintática constituinte do inglês, tanto com grandes quantidades quanto com dados limitados de treinamento.

---

arXiv:1706.03762v7 [cs.CL] 2 Aug 2023

---

arXiv:1706.03762v7 [cs.CL] 2 de agosto de 2023

---

*Equal contribution. Listing order is random. Jakob proposed replacing RNNs with self-attention and started the effort to evaluate this idea. Ashish, with Illia, designed and implemented the first Transformer models and has been crucially involved in every aspect of this work. Noam proposed scaled dot-product attention, multi-head attention and the parameter-free position representation and became the other person involved in nearly every detail. Niki designed, implemented, tuned and evaluated countless model variants in our original codebase and tensor2tensor. Llion also experimented with novel model variants, was responsible for our initial codebase, and efficient inference and visualizations. Lukasz and Aidan spent countless long days designing various parts of and implementing tensor2tensor, replacing our earlier codebase, greatly improving results and massively accelerating our research.

$^{\dagger}$Work performed while at Google Brain.

$^{\ddagger}$Work performed while at Google Research.

---

*Contribuição igual. A ordem de listagem é aleatória. Jakob propôs substituir as RNNs por autoatenção e iniciou o esforço para avaliar essa ideia. Ashish, com Illia, projetou e implementou os primeiros modelos Transformer e esteve crucialmente envolvido em todos os aspectos deste trabalho. Noam propôs a atenção de produto escalar escalonado, a atenção multi-cabeça e a representação de posição sem parâmetros, tornando-se a outra pessoa envolvida em quase todos os detalhes. Niki projetou, implementou, ajustou e avaliou inúmeras variantes de modelos em nosso código-fonte original e no tensor2tensor. Llion também experimentou com variantes de modelos inovadoras, foi responsável pelo nosso código-fonte inicial, bem como pela inferência eficiente e visualizações. Lukasz e Aidan passaram incontáveis longos dias projetando várias partes do tensor2tensor e implementando-o, substituindo nosso código-fonte anterior, melhorando significativamente os resultados e acelerando enormemente nossa pesquisa.

$^{\dagger}$Trabalho realizado durante o período no Google Brain.

$^{\ddagger}$Trabalho realizado durante o período no Google Research.

---

## 1 Introduction

---

## 1 Introdução

---

Recurrent neural networks, long short-term memory [13] and gated recurrent [7] neural networks in particular, have been firmly established as state of the art approaches in sequence modeling and transduction problems such as language modeling and machine translation [35, 2, 5]. Numerous efforts have since continued to push the boundaries of recurrent language models and encoder-decoder architectures [38, 24, 15].

Recurrent models typically factor computation along the symbol positions of the input and output sequences. Aligning the positions to steps in computation time, they generate a sequence of hidden states $h_t$, as a function of the previous hidden state $h_{t-1}$ and the input for position $t$. This inherently sequential nature precludes parallelization within training examples, which becomes critical at longer sequence lengths, as memory constraints limit batching across examples. Recent work has achieved significant improvements in computational efficiency through factorization tricks [21] and conditional computation [32], while also improving model performance in case of the latter. The fundamental constraint of sequential computation, however, remains.

Attention mechanisms have become an integral part of compelling sequence modeling and transduction models in various tasks, allowing modeling of dependencies without regard to their distance in the input or output sequences [2, 19]. In all but a few cases [27], however, such attention mechanisms are used in conjunction with a recurrent network.

In this work we propose the Transformer, a model architecture eschewing recurrence and instead relying entirely on an attention mechanism to draw global dependencies between input and output. The Transformer allows for significantly more parallelization and can reach a new state of the art in translation quality after being trained for as little as twelve hours on eight P100 GPUs.

---

As redes neurais recorrentes, as de memória de longo e curto prazo [13] e as redes neurais recorrentes com portas [7], em particular, consolidaram-se como abordagens de ponta em problemas de modelagem e transdução de sequências, como modelagem de linguagem e tradução automática [35, 2, 5]. Desde então, inúmeros esforços continuaram a ampliar os limites dos modelos de linguagem recorrentes e das arquiteturas codificador-decodificador [38, 24, 15].

Os modelos recorrentes tipicamente fatoram o cálculo ao longo das posições dos símbolos nas sequências de entrada e saída. Ao alinhar essas posições aos passos do tempo de computação, eles geram uma sequência de estados ocultos $h_t$, como função do estado oculto anterior $h_{t-1}$ e da entrada para a posição $t$. Essa natureza inerentemente sequencial impede a paralelização dentro dos exemplos de treinamento, o que se torna crítico em comprimentos de sequência maiores, pois as restrições de memória limitam o agrupamento (batching) entre exemplos. Trabalhos recentes alcançaram melhorias significativas na eficiência computacional por meio de truques de fatoração [21] e computação condicional [32], além de também melhorar o desempenho do modelo no caso desta última. A restrição fundamental do cálculo sequencial, no entanto, permanece.

Os mecanismos de atenção tornaram-se parte integrante de modelos promissores de modelagem e transdução de sequências em várias tarefas, permitindo a modelagem de dependências sem considerar sua distância nas sequências de entrada ou saída [2, 19]. Em todos, exceto alguns casos [27], esses mecanismos de atenção são usados em conjunto com uma rede recorrente.

Neste trabalho, propomos o Transformer, uma arquitetura de modelo que dispensa a recorrência e depende inteiramente de um mecanismo de atenção para estabelecer dependências globais entre a entrada e a saída. O Transformer permite uma paralelização significativamente maior e pode alcançar um novo estado da arte em qualidade de tradução após ser treinado por tão pouco quanto doze horas em oito GPUs P100.

---

## 2 Background

---

## 2 Contexto

---

The goal of reducing sequential computation also forms the foundation of the Extended Neural GPU [16], ByteNet [18] and ConvS2S [9], all of which use convolutional neural networks as basic building block, computing hidden representations in parallel for all input and output positions. In these models, the number of operations required to relate signals from two arbitrary input or output positions grows in the distance between positions, linearly for ConvS2S and logarithmically for ByteNet. This makes it more difficult to learn dependencies between distant positions [12]. In the Transformer this is reduced to a constant number of operations, albeit at the cost of reduced effective resolution due to averaging attention-weighted positions, an effect we counteract with Multi-Head Attention as described in section 3.2.

Self-attention, sometimes called intra-attention is an attention mechanism relating different positions of a single sequence in order to compute a representation of the sequence. Self-attention has been used successfully in a variety of tasks including reading comprehension, abstractive summarization, textual entailment and learning task-independent sentence representations [4, 27, 28, 22].

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

To the best of our knowledge, however, the Transformer is the first transduction model relying entirely on self-attention to compute representations of its input and output without using sequence-aligned RNNs or convolution. In the following sections, we will describe the Transformer, motivate self-attention and discuss its advantages over models such as [17, 18] and [9].

---

O objetivo de reduzir a computação sequencial também forma a base da Extended Neural GPU [16], ByteNet [18] e ConvS2S [9], todas as quais utilizam redes neurais convolucionais como bloco de construção básico, calculando representações ocultas em paralelo para todas as posições de entrada e saída. Nesses modelos, o número de operações necessárias para relacionar sinais de duas posições de entrada ou saída arbitrárias cresce com a distância entre as posições, linearmente para o ConvS2S e logaritmicamente para o ByteNet. Isso torna mais difícil aprender dependências entre posições distantes [12]. No Transformer, isso é reduzido a um número constante de operações, ainda que ao custo de uma resolução efetiva reduzida devido à média das posições ponderadas pela atenção, um efeito que contrabalançamos com Multi-Head Attention, conforme descrito na seção 3.2.

A autoatendimento, às vezes chamada de intra-atenção, é um mecanismo de atenção que relaciona diferentes posições de uma única sequência para calcular uma representação dessa sequência. A autoatendimento tem sido usada com sucesso em uma variedade de tarefas, incluindo compreensão de leitura, resumo abstrativo, implicação textual e aprendizado de representações de sentenças independentes de tarefa [4, 27, 28, 22].

As redes de memória end-to-end são baseadas em um mecanismo de atenção recorrente, em vez de recorrência alinhada à sequência, e demonstraram ter bom desempenho em tarefas simples de perguntas e respostas em linguagem natural e modelagem de linguagem [34].

Até onde sabemos, no entanto, o Transformer é o primeiro modelo de transdução que depende inteiramente de autoatendimento para calcular representações de sua entrada e saída, sem usar RNNs alinhados à sequência ou convolução. Nas seções seguintes, descreveremos o Transformer, justificaremos a autoatendimento e discutiremos suas vantagens em relação a modelos como [17, 18] e [9].

---

## 3 Model Architecture

---

## 3 Arquitetura do Modelo

---

Most competitive neural sequence transduction models have an encoder-decoder structure [5, 2, 35]. Here, the encoder maps an input sequence of symbol representations $(x_1, ..., x_n)$ to a sequence of continuous representations $\mathbf{z} = (z_1, ..., z_n)$. Given $\mathbf{z}$, the decoder then generates an output sequence $(y_1, ..., y_m)$ of symbols one element at a time. At each step the model is auto-regressive [10], consuming the previously generated symbols as additional input when generating the next.

---

A maioria dos modelos competitivos de transdução sequencial neural possui uma estrutura codificador-decodificador [5, 2, 35]. Aqui, o codificador mapeia uma sequência de entrada de representações de símbolos $(x_1, ..., x_n)$ para uma sequência de representações contínuas $\mathbf{z} = (z_1, ..., z_n)$. Dado $\mathbf{z}$, o decodificador gera então uma sequência de saída $(y_1, ..., y_m)$ de símbolos um elemento por vez. Em cada etapa, o modelo é autorregressivo [10], consumindo os símbolos previamente gerados como entrada adicional ao gerar o próximo.

---

<img src="images/3-1.png" style="zoom:70%; display: block; margin: 0 auto;" />

Figure 1: The Transformer - model architecture.

---

Figura 1: A arquitetura do modelo Transformer.

---

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.

---

O Transformer segue essa arquitetura geral, utilizando camadas empilhadas de autoatenção e pontos de conexão completa tanto para o codificador quanto para o decodificador, mostrados nas metades esquerda e direita da Figura 1, respectivamente.

---

### 3.1 Encoder and Decoder Stacks

---

### 3.1 Pilhas de Codificador e Decodificador

---

**Encoder:** The encoder is composed of a stack of $N = 6$ identical layers. Each layer has two sub-layers. The first is a multi-head self-attention mechanism, and the second is a simple, position-wise fully connected feed-forward network. We employ a residual connection [11] around each of the two sub-layers, followed by layer normalization [1]. That is, the output of each sub-layer is $\text{LayerNorm}(x + \text{Sublayer}(x))$, where $\text{Sublayer}(x)$ is the function implemented by the sub-layer itself. To facilitate these residual connections, all sub-layers in the model, as well as the embedding layers, produce outputs of dimension $d_{\text{model}} = 512$.

**Decoder:** The decoder is also composed of a stack of $N = 6$ identical layers. In addition to the two sub-layers in each encoder layer, the decoder inserts a third sub-layer, which performs multi-head attention over the output of the encoder stack. Similar to the encoder, we employ residual connections around each of the sub-layers, followed by layer normalization. We also modify the self-attention sub-layer in the decoder stack to prevent positions from attending to subsequent positions. This masking, combined with fact that the output embeddings are offset by one position, ensures that the predictions for position $i$ can depend only on the known outputs at positions less than $i$.

---

**Codificador:** O codificador é composto por uma pilha de $N = 6$ camadas idênticas. Cada camada possui duas subcamadas. A primeira é um mecanismo de autoatenção multi-cabeça, e a segunda é uma rede feed-forward totalmente conectada simples e posicional. Empregamos uma conexão residual [11] ao redor de cada uma das duas subcamadas, seguida por normalização em camada [1]. Ou seja, a saída de cada subcamada é $\text{LayerNorm}(x + \text{Sublayer}(x))$, onde $\text{Sublayer}(x)$ é a função implementada pela própria subcamada. Para facilitar essas conexões residuais, todas as subcamadas no modelo, bem como as camadas de incorporação, produzem saídas de dimensão $d_{\text{model}} = 512$.

**Decodificador:** O decodificador também é composto por uma pilha de $N = 6$ camadas idênticas. Além das duas subcamadas em cada camada do codificador, o decodificador insere uma terceira subcamada, que realiza atenção multi-cabeça sobre a saída da pilha do codificador. Semelhante ao codificador, empregamos conexões residuais ao redor de cada uma das subcamadas, seguidas por normalização em camada. Também modificamos a subcamada de autoatenção na pilha do decodificador para impedir que posições atendam a posições subsequentes. Essa máscara, combinada com o fato de que as incorporações de saída são deslocadas em uma posição, garante que as previsões para a posição $i$ possam depender apenas das saídas conhecidas nas posições menores que $i$.

---

### 3.2 Attention

---

### 3.2 Atenção

---

An attention function can be described as mapping a query and a set of key-value pairs to an output, where the query, keys, values, and output are all vectors. The output is computed as a weighted sum

---

Uma função de atenção pode ser descrita como mapeando uma consulta e um conjunto de pares chave-valor para uma saída, onde a consulta, as chaves, os valores e a saída são todos vetores. A saída é calculada como uma soma ponderada

---

<img src="images/4-3.png" style="zoom:70%; display: block; margin: 0 auto;" />

Figure 2: (left) Scaled Dot-Product Attention. (right) Multi-Head Attention consists of several attention layers running in parallel.

---

Figura 2: (esquerda) Atenção de Produto Escalar Pontual. (direita) A Atenção Multi-Cabeça consiste em várias camadas de atenção executadas em paralelo.

---

of the values, where the weight assigned to each value is computed by a compatibility function of the query with the corresponding key.

---

dos valores, onde o peso atribuído a cada valor é calculado por uma função de compatibilidade da consulta com a chave correspondente.

---

### 3.2.1 Scaled Dot-Product Attention

---

### 3.2.1 Atenção de Produto Escalar

---

We call our particular attention "Scaled Dot-Product Attention" (Figure 2). The input consists of queries and keys of dimension $d_k$, and values of dimension $d_v$. We compute the dot products of the query with all keys, divide each by $\sqrt{d_k}$, and apply a softmax function to obtain the weights on the values.

In practice, we compute the attention function on a set of queries simultaneously, packed together into a matrix $Q$. The keys and values are also packed together into matrices $K$ and $V$. We compute the matrix of outputs as:

$$
\text{Attention}(Q, K, V) = \text{softmax}(\frac{QK^T}{\sqrt{d_k}})V \tag{1}
$$

The two most commonly used attention functions are additive attention [2], and dot-product (multiplicative) attention. Dot-product attention is identical to our algorithm, except for the scaling factor of $\frac{1}{\sqrt{d_k}}$. Additive attention computes the compatibility function using a feed-forward network with a single hidden layer. While the two are similar in theoretical complexity, dot-product attention is much faster and more space-efficient in practice, since it can be implemented using highly optimized matrix multiplication code.

While for small values of $d_k$ the two mechanisms perform similarly, additive attention outperforms dot product attention without scaling for larger values of $d_k$ [3]. We suspect that for large values of $d_k$, the dot products grow large in magnitude, pushing the softmax function into regions where it has extremely small gradients $^{4}$. To counteract this effect, we scale the dot products by $\frac{1}{\sqrt{d_k}}$.

---

Chamamos nossa atenção especial para a “Atenção de Produto Escalar Pontual” (Figura 2). A entrada consiste em consultas e chaves de dimensão $d_k$, e valores de dimensão $d_v$. Calculamos os produtos escalares da consulta com todas as chaves, dividimos cada um por $\sqrt{d_k}$ e aplicamos uma função softmax para obter os pesos sobre os valores.

Na prática, calculamos a função de atenção em um conjunto de consultas simultaneamente, agrupadas em uma matriz $Q$. As chaves e os valores também são agrupados nas matrizes $K$ e $V$. Calculamos a matriz de saídas como:

$$
\text{Attention}(Q, K, V) = \text{softmax}(\frac{QK^T}{\sqrt{d_k}})V \tag{1}
$$

As duas funções de atenção mais comumente usadas são a atenção aditiva [2] e a atenção de produto pontual (multiplicativa). A atenção de produto pontual é idêntica ao nosso algoritmo, exceto pelo fator de escala de $\frac{1}{\sqrt{d_k}}$. A atenção aditiva calcula a função de compatibilidade usando uma rede feed-forward com uma única camada oculta. Embora as duas sejam semelhantes em complexidade teórica, a atenção de produto pontual é muito mais rápida e eficiente em termos de espaço na prática, pois pode ser implementada usando código de multiplicação de matrizes altamente otimizado.

Embora, para valores pequenos de $d_k$, os dois mecanismos tenham desempenho semelhante, a atenção aditiva supera a atenção de produto pontual sem escala para valores maiores de $d_k$ [3]. Suspeitamos que, para valores grandes de $d_k$, os produtos escalares crescem em magnitude, empurrando a função softmax para regiões onde ela possui gradientes extremamente pequenos $^{4}$. Para contrariar esse efeito, escalamos os produtos escalares por $\frac{1}{\sqrt{d_k}}$.

---

### 3.2.2 Multi-Head Attention

---

### 3.2.2 Atenção Multi-Cabeça

---

Instead of performing a single attention function with $d_{\text{model}}$-dimensional keys, values and queries, we found it beneficial to linearly project the queries, keys and values $h$ times with different, learned linear projections to $d_k$, $d_k$ and $d_v$ dimensions, respectively. On each of these projected versions of queries, keys and values we then perform the attention function in parallel, yielding $d_v$-dimensional

---

Em vez de realizar uma única função de atenção com chaves, valores e consultas de dimensão $d_{\text{model}}$, constatamos ser benéfico projetar linearmente as consultas, chaves e valores $h$ vezes usando projeções lineares aprendidas diferentes para as dimensões $d_k$, $d_k$ e $d_v$, respectivamente. Em cada uma dessas versões projetadas das consultas, chaves e valores, realizamos então a função de atenção em paralelo, produzindo saídas de dimensão $d_v$.

---

$^4$To illustrate why the dot products get large, assume that the components of $q$ and $k$ are independent random variables with mean 0 and variance 1. Then their dot product, $q \cdot k = \sum_{i=1}^{d_k} q_i k_i$, has mean 0 and variance $d_k$.

---

$^4$Para ilustrar por que os produtos escalares ficam grandes, assuma que as componentes de $q$ e $k$ são variáveis aleatórias independentes com média 0 e variância 1. Então seu produto escalar, $q \cdot k = \sum_{i=1}^{d_k} q_i k_i$, tem média 0 e variância $d_k$.

---

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

Multi-head attention allows the model to jointly attend to information from different representation subspaces at different positions. With a single attention head, averaging inhibits this.

$$
\begin{align}
&\text{MultiHead}(Q, K, V) = \text{Concat}(\text{head}_1, ..., \text{head}_{\text{h}})W^O \tag{} \\
&\text{where } \text{head}_{\text{i}} = \text{Attention}(QW_i^Q, KW_i^K, VW_i^V) \tag{}
\end{align}
$$

Where the projections are parameter matrices $W_i^Q \in \mathbb{R}^{d_{\text{model}} \times d_k}$, $W_i^K \in \mathbb{R}^{d_{\text{model}} \times d_k}$, $W_i^V \in \mathbb{R}^{d_{\text{model}} \times d_v}$ and $W^O \in \mathbb{R}^{hd_v \times d_{\text{model}}}$.

In this work we employ $h = 8$ parallel attention layers, or heads. For each of these we use $d_k = d_v = d_{\text{model}}/h = 64$. Due to the reduced dimension of each head, the total computational cost is similar to that of single-head attention with full dimensionality.

---

valores de saída. Estes são concatenados e novamente projetados, resultando nos valores finais, conforme ilustrado na Figura 2.

A atenção multi-cabeça permite que o modelo atenda simultaneamente a informações provenientes de diferentes subespaços de representação em diferentes posições. Com uma única cabeça de atenção, a média inibe isso.

$$
\begin{align}
&\text{MultiHead}(Q, K, V) = \text{Concat}(\text{head}_1, ..., \text{head}_{\text{h}})W^O \tag{} \\
&\text{onde } \text{head}_{\text{i}} = \text{Attention}(QW_i^Q, KW_i^K, VW_i^V) \tag{}
\end{align}
$$

Onde as projeções são matrizes de parâmetros $W_i^Q \in \mathbb{R}^{d_{\text{model}} \times d_k}$, $W_i^K \in \mathbb{R}^{d_{\text{model}} \times d_k}$, $W_i^V \in \mathbb{R}^{d_{\text{model}} \times d_v}$ e $W^O \in \mathbb{R}^{hd_v \times d_{\text{model}}}$.

Neste trabalho, empregamos $h = 8$ camadas de atenção paralelas, ou cabeças. Para cada uma delas, utilizamos $d_k = d_v = d_{\text{model}}/h = 64$. Devido à redução da dimensão de cada cabeça, o custo computacional total é semelhante ao da atenção de cabeça única com dimensionalidade completa.

---

### 3.2.3 Applications of Attention in our Model

---

### 3.2.3 Aplicações da Atenção em nosso Modelo

---

The Transformer uses multi-head attention in three different ways:

*   In "encoder-decoder attention" layers, the queries come from the previous decoder layer, and the memory keys and values come from the output of the encoder. This allows every position in the decoder to attend over all positions in the input sequence. This mimics the typical encoder-decoder attention mechanisms in sequence-to-sequence models such as [38, 2, 9].
*   The encoder contains self-attention layers. In a self-attention layer all of the keys, values and queries come from the same place, in this case, the output of the previous layer in the encoder. Each position in the encoder can attend to all positions in the previous layer of the encoder.
*   Similarly, self-attention layers in the decoder allow each position in the decoder to attend to all positions in the decoder up to and including that position. We need to prevent leftward information flow in the decoder to preserve the auto-regressive property. We implement this inside of scaled dot-product attention by masking out (setting to $-\infty$) all values in the input of the softmax which correspond to illegal connections. See Figure 2.

---

O Transformer utiliza atenção multi-cabeça de três maneiras diferentes:

*   Nas camadas de "atenção codificador-decodificador", as consultas vêm da camada anterior do decodificador, e as chaves e valores de memória vêm da saída do codificador. Isso permite que cada posição no decodificador atenda a todas as posições na sequência de entrada. Isso imita os mecanismos típicos de atenção codificador-decodificador em modelos sequência-para-sequência, como [38, 2, 9].
*   O codificador contém camadas de autoatendimento. Em uma camada de autoatendimento, todas as chaves, valores e consultas vêm do mesmo lugar, neste caso, da saída da camada anterior no codificador. Cada posição no codificador pode atender a todas as posições na camada anterior do codificador.
*   Da mesma forma, as camadas de autoatendimento no decodificador permitem que cada posição no decodificador atenda a todas as posições no decodificador até e incluindo essa posição. Precisamos impedir o fluxo de informação para a esquerda no decodificador para preservar a propriedade autoregressiva. Implementamos isso dentro da atenção de produto escalar escalonado mascarando (definindo como $-\infty$) todos os valores na entrada do softmax que correspondem a conexões ilegais. Veja a Figura 2.

---

### 3.3 Position-wise Feed-Forward Networks

---

### 3.3 Redes Feed-Forward Posicionais

---

In addition to attention sub-layers, each of the layers in our encoder and decoder contains a fully connected feed-forward network, which is applied to each position separately and identically. This consists of two linear transformations with a ReLU activation in between.

$$
\text{FFN}(x) = \max(0, xW_1 + b_1)W_2 + b_2 \tag{2}
$$

While the linear transformations are the same across different positions, they use different parameters from layer to layer. Another way of describing this is as two convolutions with kernel size 1. The dimensionality of input and output is $d_{\text{model}} = 512$, and the inner-layer has dimensionality $d_{ff} = 2048$.

---

Além dos subcampos de atenção, cada uma das camadas em nosso codificador e decodificador contém uma rede feed-forward totalmente conectada, que é aplicada separadamente e de forma idêntica a cada posição. Isso consiste em duas transformações lineares com uma ativação ReLU no meio.

$$
\text{FFN}(x) = \max(0, xW_1 + b_1)W_2 + b_2 \tag{2}
$$

Embora as transformações lineares sejam as mesmas em diferentes posições, elas usam parâmetros diferentes de camada para camada. Outra maneira de descrever isso é como duas convoluções com tamanho de kernel 1. A dimensionalidade da entrada e da saída é $d_{\text{model}} = 512$, e a camada interna tem dimensionalidade $d_{ff} = 2048$.

---

### 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 $d_{\text{model}}$. We also use the usual learned linear transformation and softmax function to convert the decoder output to predicted next-token probabilities. In our model, we share the same weight matrix between the two embedding layers and the pre-softmax linear transformation, similar to [30]. In the embedding layers, we multiply those weights by $\sqrt{d_{\text{model}}}$.

---

Da mesma forma que outros modelos de transdução de sequências, utilizamos embeddings aprendidos para converter os tokens de entrada e os tokens de saída em vetores de dimensão $d_{\text{model}}$. Também empregamos a transformação linear aprendida usual e a função softmax para converter a saída do decodificador em probabilidades previstas para o próximo token. Em nosso modelo, compartilhamos a mesma matriz de pesos entre as duas camadas de embedding e a transformação linear pré-softmax, semelhante ao proposto em [30]. Nas camadas de embedding, multiplicamos esses pesos por $\sqrt{d_{\text{model}}}$.

---

Table 1: Maximum path lengths, per-layer complexity and minimum number of sequential operations for different layer types. $n$ is the sequence length, $d$ is the representation dimension, $k$ is the kernel size of convolutions and $r$ the size of the neighborhood in restricted self-attention.

---

Tabela 1: Comprimentos máximos dos caminhos, complexidade por camada e número mínimo de operações sequenciais para diferentes tipos de camadas. $n$ é o comprimento da sequência, $d$ é a dimensão da representação, $k$ é o tamanho do núcleo das convoluções e $r$ é o tamanho da vizinhança na autoatenção restrita.

---

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

---

| Tipo de Camada | Complexidade por Camada | Operações Sequenciais | Comprimento Máximo do Caminho |
| :--- | :---: | :---: | :---: |
| Autoatenção | $O(n^2 \cdot d)$ | $O(1)$ | $O(1)$ |
| Recorrente | $O(n \cdot d^2)$ | $O(n)$ | $O(n)$ |
| Convolucional | $O(k \cdot n \cdot d^2)$ | $O(1)$ | $O(log_k(n))$ |
| Autoatenção (restrita) | $O(r \cdot n \cdot d)$ | $O(1)$ | $O(n/r)$ |

---

### 3.5 Positional Encoding

---

### 3.5 Codificação Posicional

---

Since our model contains no recurrence and no convolution, in order for the model to make use of the order of the sequence, we must inject some information about the relative or absolute position of the tokens in the sequence. To this end, we add "positional encodings" to the input embeddings at the bottoms of the encoder and decoder stacks. The positional encodings have the same dimension $d_{\text{model}}$ as the embeddings, so that the two can be summed. There are many choices of positional encodings, learned and fixed [9].

In this work, we use sine and cosine functions of different frequencies:

$$
\begin{align}
PE_{(pos,2i)} &= sin(pos/10000^{2i/d_{\text{model}}}) \\
PE_{(pos,2i+1)} &= cos(pos/10000^{2i/d_{\text{model}}})
\end{align}
$$

where $pos$ is the position and $i$ is the dimension. That is, each dimension of the positional encoding corresponds to a sinusoid. The wavelengths form a geometric progression from $2\pi$ to $10000 \cdot 2\pi$. We chose this function because we hypothesized it would allow the model to easily learn to attend by relative positions, since for any fixed offset $k$, $PE_{pos+k}$ can be represented as a linear function of $PE_{pos}$.

We also experimented with using learned positional embeddings [9] instead, and found that the two versions produced nearly identical results (see Table 3 row (E)). We chose the sinusoidal version because it may allow the model to extrapolate to sequence lengths longer than the ones encountered during training.

---

Como nosso modelo não contém recorrência nem convolução, para que o modelo possa utilizar a ordem da sequência, devemos injetar alguma informação sobre a posição relativa ou absoluta dos tokens na sequência. Para esse fim, adicionamos "codificações posicionais" aos embeddings de entrada na base das pilhas do codificador e do decodificador. As codificações posicionais têm a mesma dimensão $d_{\text{model}}$ dos embeddings, de modo que os dois possam ser somados. Existem muitas escolhas de codificações posicionais, aprendidas e fixas [9].

Neste trabalho, utilizamos funções seno e cosseno de diferentes frequências:

$$
\begin{align}
PE_{(pos,2i)} &= sin(pos/10000^{2i/d_{\text{model}}}) \\
PE_{(pos,2i+1)} &= cos(pos/10000^{2i/d_{\text{model}}})
\end{align}
$$

onde $pos$ é a posição e $i$ é a dimensão. Ou seja, cada dimensão da codificação posicional corresponde a uma senóide. Os comprimentos de onda formam uma progressão geométrica de $2\pi$ a $10000 \cdot 2\pi$. Escolhemos essa função porque hipotetizamos que ela permitiria ao modelo aprender facilmente a atender por posições relativas, já que, para qualquer deslocamento fixo $k$, $PE_{pos+k}$ pode ser representado como uma função linear de $PE_{pos}$.

Também experimentamos usar embeddings posicionais aprendidos [9] em vez disso e descobrimos que as duas versões produziram resultados quase idênticos (veja a linha (E) da Tabela 3). Escolhemos a versão sinusoidal porque ela pode permitir que o modelo extrapole para comprimentos de sequência maiores do que aqueles encontrados durante o treinamento.

---

## 4 Why Self-Attention

---

## 4 Por que Autoatenção

---

In this section we compare various aspects of self-attention layers to the recurrent and convolutional layers commonly used for mapping one variable-length sequence of symbol representations $(x_1, ..., x_n)$ to another sequence of equal length $(z_1, ..., z_n)$, with $x_i, z_i \in \mathbb{R}^d$, such as a hidden layer in a typical sequence transduction encoder or decoder. Motivating our use of self-attention we consider three desiderata.

One is the total computational complexity per layer. Another is the amount of computation that can be parallelized, as measured by the minimum number of sequential operations required.

The third is the path length between long-range dependencies in the network. Learning long-range dependencies is a key challenge in many sequence transduction tasks. One key factor affecting the ability to learn such dependencies is the length of the paths forward and backward signals have to traverse in the network. The shorter these paths between any combination of positions in the input and output sequences, the easier it is to learn long-range dependencies [12]. Hence we also compare the maximum path length between any two input and output positions in networks composed of the different layer types.

As noted in Table 1, a self-attention layer connects all positions with a constant number of sequentially executed operations, whereas a recurrent layer requires $O(n)$ sequential operations. In terms of computational complexity, self-attention layers are faster than recurrent layers when the sequence

---

Nesta seção, comparamos vários aspectos das camadas de autoatenção às camadas recorrentes e convolucionais comumente usadas para mapear uma sequência de representações de símbolos de comprimento variável $(x_1, ..., x_n)$ para outra sequência de comprimento igual $(z_1, ..., z_n)$, com $x_i, z_i \in \mathbb{R}^d$, como uma camada oculta em um codificador ou decodificador típico de transdução de sequências. Para motivar o uso da autoatenção, consideramos três requisitos.

Um deles é a complexidade computacional total por camada. Outro é a quantidade de computação que pode ser paralelizada, medida pelo número mínimo de operações sequenciais necessárias.

O terceiro é o comprimento do caminho entre dependências de longo alcance na rede. Aprender dependências de longo alcance é um desafio central em muitas tarefas de transdução de sequências. Um fator-chave que afeta a capacidade de aprender tais dependências é o comprimento dos caminhos que os sinais devem percorrer para frente e para trás na rede. Quanto mais curtos forem esses caminhos entre qualquer combinação de posições nas sequências de entrada e saída, mais fácil será aprender dependências de longo alcance [12]. Portanto, também comparamos o comprimento máximo do caminho entre quaisquer duas posições de entrada e saída em redes compostas pelos diferentes tipos de camadas.

Como observado na Tabela 1, uma camada de autoatenção conecta todas as posições com um número constante de operações executadas sequencialmente, enquanto uma camada recorrente requer $O(n)$ operações sequenciais. Em termos de complexidade computacional, as camadas de autoatenção são mais rápidas do que as camadas recorrentes quando a sequência

---

length $n$ is smaller than the representation dimensionality $d$, which is most often the case with sentence representations used by state-of-the-art models in machine translations, such as word-piece [38] and byte-pair [31] representations. To improve computational performance for tasks involving very long sequences, self-attention could be restricted to considering only a neighborhood of size $r$ in the input sequence centered around the respective output position. This would increase the maximum path length to $O(n/r)$. We plan to investigate this approach further in future work.

A single convolutional layer with kernel width $k < n$ does not connect all pairs of input and output positions. Doing so requires a stack of $O(n/k)$ convolutional layers in the case of contiguous kernels, or $O(\log_k(n))$ in the case of dilated convolutions [18], increasing the length of the longest paths between any two positions in the network. Convolutional layers are generally more expensive than recurrent layers, by a factor of $k$. Separable convolutions [6], however, decrease the complexity considerably, to $O(k \cdot n \cdot d + n \cdot d^2)$. Even with $k = n$, however, the complexity of a separable convolution is equal to the combination of a self-attention layer and a point-wise feed-forward layer, the approach we take in our model.

As side benefit, self-attention could yield more interpretable models. We inspect attention distributions from our models and present and discuss examples in the appendix. Not only do individual attention heads clearly learn to perform different tasks, many appear to exhibit behavior related to the syntactic and semantic structure of the sentences.

---

comprimento $n$ é menor que a dimensionalidade da representação $d$, o que ocorre na maioria dos casos com as representações de sentenças utilizadas pelos modelos mais avançados em tradução automática, como as representações word-piece [38] e byte-pair [31]. Para melhorar o desempenho computacional em tarefas que envolvem sequências muito longas, o autoatendimento poderia ser restrito a considerar apenas uma vizinhança de tamanho $r$ na sequência de entrada, centralizada em torno da respectiva posição de saída. Isso aumentaria o comprimento máximo do caminho para $O(n/r)$. Planejamos investigar essa abordagem mais detalhadamente em trabalhos futuros.

Uma única camada convolucional com largura de kernel $k < n$ não conecta todos os pares de posições de entrada e saída. Conectar todos esses pares requer uma pilha de $O(n/k)$ camadas convolucionais no caso de kernels contíguos, ou $O(\log_k(n))$ no caso de convoluções dilatadas [18], aumentando o comprimento dos caminhos mais longos entre quaisquer duas posições na rede. Camadas convolucionais são geralmente mais custosas do que camadas recorrentes por um fator de $k$. Convoluções separáveis [6], no entanto, reduzem consideravelmente a complexidade para $O(k \cdot n \cdot d + n \cdot d^2)$. Mesmo com $k = n$, a complexidade de uma convolução separável é igual à combinação de uma camada de autoatendimento e uma camada feed-forward pontual, a abordagem adotada em nosso modelo.

Como benefício adicional, o autoatendimento pode produzir modelos mais interpretáveis. Inspecionamos as distribuições de atenção dos nossos modelos e apresentamos e discutimos exemplos no apêndice. Não apenas cabeças individuais de atenção aprendem claramente a realizar tarefas diferentes, mas muitas parecem exibir comportamento relacionado à estrutura sintática e semântica das sentenças.

---

## 5 Training

---

## 5 Treinamento

---

This section describes the training regime for our models.

---

Esta seção descreve o regime de treinamento para nossos modelos.

---

### **5.1** Training Data and Batching

---

### **5.1** Dados de Treinamento e Lotes

---

We trained on the standard WMT 2014 English-German dataset consisting of about 4.5 million sentence pairs. Sentences were encoded using byte-pair encoding [3], which has a shared source-target vocabulary of about 37000 tokens. For English-French, we used the significantly larger WMT 2014 English-French dataset consisting of 36M sentences and split tokens into a 32000 word-piece vocabulary [38]. Sentence pairs were batched together by approximate sequence length. Each training batch contained a set of sentence pairs containing approximately 25000 source tokens and 25000 target tokens.

---

Treinamos no conjunto de dados padrão WMT 2014 inglês-alemão, composto por cerca de 4,5 milhões de pares de frases. As frases foram codificadas usando codificação byte-pair [3], que possui um vocabulário compartilhado entre origem e destino de aproximadamente 37.000 tokens. Para o inglês-francês, utilizamos o conjunto de dados WMT 2014 inglês-francês, significativamente maior, composto por 36 milhões de frases, e dividimos os tokens em um vocabulário word-piece de 32.000 palavras [38]. Os pares de frases foram agrupados em lotes com base no comprimento aproximado da sequência. Cada lote de treinamento continha um conjunto de pares de frases contendo aproximadamente 25.000 tokens de origem e 25.000 tokens de destino.

---

### **5.2 Hardware and Schedule**

---

### **5.2 Hardware e Cronograma**

---

We trained our models on one machine with 8 NVIDIA P100 GPUs. For our base models using the hyperparameters described throughout the paper, each training step took about 0.4 seconds. We trained the base models for a total of 100,000 steps or 12 hours. For our big models,(described on the bottom line of table 3), step time was 1.0 seconds. The big models were trained for 300,000 steps (3.5 days).

---

Treinamos nossos modelos em uma máquina com 8 GPUs NVIDIA P100. Para nossos modelos base, utilizando os hiperparâmetros descritos ao longo do artigo, cada etapa de treinamento levou cerca de 0,4 segundos. Treinamos os modelos base por um total de 100.000 etapas ou 12 horas. Para nossos modelos grandes (descritos na última linha da tabela 3), o tempo por etapa foi de 1,0 segundo. Os modelos grandes foram treinados por 300.000 etapas (3,5 dias).

---

### 5.3 Optimizer

---

### 5.3 Otimizador

---

We used the Adam optimizer [20] with $\beta_1 = 0.9$, $\beta_2 = 0.98$ and $\epsilon = 10^{-9}$. We varied the learning rate over the course of training, according to the formula:

$$
lrate = d_{\text{model}}^{-0.5} \cdot \min(step\_num^{-0.5}, step\_num \cdot warmup\_steps^{-1.5}) \tag{3}
$$

This corresponds to increasing the learning rate linearly for the first $warmup\_steps$ training steps, and decreasing it thereafter proportionally to the inverse square root of the step number. We used $warmup\_steps = 4000$.

---

Utilizamos o otimizador Adam [20] com $\beta_1 = 0.9$, $\beta_2 = 0.98$ e $\epsilon = 10^{-9}$. Variamos a taxa de aprendizado ao longo do treinamento, de acordo com a fórmula:

$$
lrate = d_{\text{model}}^{-0.5} \cdot \min(step\_num^{-0.5}, step\_num \cdot warmup\_steps^{-1.5}) \tag{3}
$$

Isso corresponde a aumentar a taxa de aprendizado linearmente durante os primeiros $warmup\_steps$ passos de treinamento e diminuí-la posteriormente proporcionalmente à raiz quadrada inversa do número do passo. Utilizamos $warmup\_steps = 4000$.

---

### 5.4 Regularization

---

### 5.4 Regularização

---

We employ three types of regularization during training:

---

Empregamos três tipos de regularização durante o treinamento:

---

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.

---

Tabela 2: O Transformer alcança melhores pontuações BLEU do que os modelos anteriores de ponta nos testes newstest2014 de inglês para alemão e inglês para francês, a uma fração do custo de treinamento.

---

| Model | BLEU | | Training Cost (FLOPs) | |
| :--- | :---: | :---: | :---: | :---: |
| | EN-DE | EN-FR | EN-DE | EN-FR |
| ByteNet [18] | 23.75 | | | |
| Deep-Att + PosUnk [39] | | 39.2 | | $1.0 \cdot 10^{20}$ |
| GNMT + RL [38] | 24.6 | 39.92 | $2.3 \cdot 10^{19}$ | $1.4 \cdot 10^{20}$ |
| ConvS2S [9] | 25.16 | 40.46 | $9.6 \cdot 10^{18}$ | $1.5 \cdot 10^{20}$ |
| MoE [32] | 26.03 | 40.56 | $2.0 \cdot 10^{19}$ | $1.2 \cdot 10^{20}$ |
| Deep-Att + PosUnk Ensemble [39] | | 40.4 | | $8.0 \cdot 10^{20}$ |
| GNMT + RL Ensemble [38] | 26.30 | 41.16 | $1.8 \cdot 10^{20}$ | $1.1 \cdot 10^{21}$ |
| ConvS2S Ensemble [9] | 26.36 | **41.29** | $7.7 \cdot 10^{19}$ | $1.2 \cdot 10^{21}$ |
| Transformer (base model) | 27.3 | 38.1 | $\mathbf{3.3 \cdot 10^{18}}$ | |
| Transformer (big) | **28.4** | **41.8** | $2.3 \cdot 10^{19}$ | |

**Residual Dropout** We apply dropout [33] to the output of each sub-layer, before it is added to the sub-layer input and normalized. In addition, we apply dropout to the sums of the embeddings and the positional encodings in both the encoder and decoder stacks. For the base model, we use a rate of $P_{drop} = 0.1$.

---

**Dropout Residual** Aplicamos dropout [33] à saída de cada subcamada, antes que ela seja adicionada à entrada da subcamada e normalizada. Além disso, aplicamos dropout às somas dos embeddings e dos codigos posicionais em ambas as pilhas do codificador e do decodificador. Para o modelo básico, utilizamos uma taxa de $P_{drop} = 0.1$.

---

**Label Smoothing** During training, we employed label smoothing of value $\epsilon_{ls} = 0.1$ [36]. This hurts perplexity, as the model learns to be more unsure, but improves accuracy and BLEU score.

---

**Suavização de Rótulos** Durante o treinamento, utilizamos suavização de rótulos com valor $\epsilon_{ls} = 0.1$ [36]. Isso prejudica a perplexidade, pois o modelo aprende a ser mais incerto, mas melhora a precisão e a pontuação BLEU.

---

## 6 Results

---

## 6 Resultados

---

### 6.1 Machine Translation

---

### 6.1 Tradução Automática

---

On the WMT 2014 English-to-German translation task, the big transformer model (Transformer (big) in Table 2) outperforms the best previously reported models (including ensembles) by more than 2.0 BLEU, establishing a new state-of-the-art BLEU score of 28.4. The configuration of this model is listed in the bottom line of Table 3. Training took 3.5 days on 8 P100 GPUs. Even our base model surpasses all previously published models and ensembles, at a fraction of the training cost of any of the competitive models.

On the WMT 2014 English-to-French translation task, our big model achieves a BLEU score of 41.0, outperforming all of the previously published single models, at less than $1/4$ the training cost of the previous state-of-the-art model. The Transformer (big) model trained for English-to-French used dropout rate $P_{drop} = 0.1$, instead of 0.3.

For the base models, we used a single model obtained by averaging the last 5 checkpoints, which were written at 10-minute intervals. For the big models, we averaged the last 20 checkpoints. We used beam search with a beam size of 4 and length penalty $\alpha = 0.6$ [38]. These hyperparameters were chosen after experimentation on the development set. We set the maximum output length during inference to input length + 50, but terminate early when possible [38].

Table 2 summarizes our results and compares our translation quality and training costs to other model architectures from the literature. We estimate the number of floating point operations used to train a model by multiplying the training time, the number of GPUs used, and an estimate of the sustained single-precision floating-point capacity of each GPU $^5$.

---

Na tarefa de tradução do inglês para o alemão do WMT 2014, o modelo grande do Transformer (Transformer (big) na Tabela 2) supera os melhores modelos relatados anteriormente (incluindo conjuntos de modelos) em mais de 2,0 BLEU, estabelecendo uma nova pontuação BLEU de estado da arte de 28,4. A configuração deste modelo está listada na última linha da Tabela 3. O treinamento levou 3,5 dias em 8 GPUs P100. Mesmo nosso modelo base supera todos os modelos e conjuntos de modelos publicados anteriormente, a uma fração do custo de treinamento de qualquer um dos modelos competitivos.

Na tarefa de tradução do inglês para o francês do WMT 2014, nosso modelo grande alcança uma pontuação BLEU de 41,0, superando todos os modelos únicos publicados anteriormente, por menos de $1/4$ do custo de treinamento do modelo anterior de estado da arte. O modelo Transformer (big) treinado para inglês para francês utilizou taxa de dropout $P_{drop} = 0,1$, em vez de 0,3.

Para os modelos base, utilizamos um único modelo obtido pela média das últimas 5 verificações de ponto (checkpoints), que foram salvas em intervalos de 10 minutos. Para os modelos grandes, fizemos a média das últimas 20 verificações de ponto. Utilizamos busca em feixe com tamanho de feixe de 4 e penalidade de comprimento $\alpha = 0,6$ [38]. Esses hiperparâmetros foram escolhidos após experimentação no conjunto de desenvolvimento. Definimos o comprimento máximo de saída durante a inferência como comprimento de entrada + 50, mas encerramos antecipadamente quando possível [38].

A Tabela 2 resume nossos resultados e compara nossa qualidade de tradução e custos de treinamento com outras arquiteturas de modelos da literatura. Estimamos o número de operações de ponto flutuante usadas para treinar um modelo multiplicando o tempo de treinamento, o número de GPUs utilizadas e uma estimativa da capacidade sustentada de ponto flutuante de precisão simples de cada GPU $^5$.

---

### 6.2 Model Variations

---

### 6.2 Variações do Modelo

---

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

---

Para avaliar a importância dos diferentes componentes do Transformer, variamos nosso modelo base de diversas maneiras, medindo a mudança no desempenho na tradução de inglês para alemão no

---

$^5$We used values of 2.8, 3.7, 6.0 and 9.5 TFLOPS for K80, K40, M40 and P100, respectively.

---

$^5$Utilizamos valores de 2,8, 3,7, 6,0 e 9,5 TFLOPS para K80, K40, M40 e P100, respectivamente.

---

Table 3: Variations on the Transformer architecture. Unlisted values are identical to those of the base model. All metrics are on the English-to-German translation development set, newstest2013. Listed perplexities are per-wordpiece, according to our byte-pair encoding, and should not be compared to per-word perplexities.

---

Tabela 3: Variações na arquitetura do Transformer. Os valores não listados são idênticos aos do modelo base. Todas as métricas estão no conjunto de desenvolvimento de tradução do inglês para o alemão, newstest2013. As perplexidades listadas são por peça de byte (byte-piece), de acordo com nossa codificação por pares de bytes, e não devem ser comparadas às perplexidades por palavra.

---

| | $N$ | $d_{\text{model}}$ | $d_{\text{ff}}$ | $h$ | $d_k$ | $d_v$ | $P_{drop}$ | $\epsilon_{ls}$ | train steps | PPL (dev) | BLEU (dev) | params $\times 10^6$ |
| :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
| base | 6 | 512 | 2048 | 8 | 64 | 64 | 0.1 | 0.1 | 100K | 4.92 | 25.8 | 65 |
| (A) | | | | 1 | 512 | 512 | | | | 5.29 | 24.9 | |
| | | | | 4 | 128 | 128 | | | | 5.00 | 25.5 | |
| | | | | 16 | 32 | 32 | | | | 4.91 | 25.8 | |
| | | | | 32 | 16 | 16 | | | | 5.01 | 25.4 | |
| (B) | | | | 16 | | | | | | 5.16 | 25.1 | 58 |
| | | | | 32 | | | | | | 5.01 | 25.4 | 60 |
| (C) | 2 | | | | | | | | | 6.11 | 23.7 | 36 |
| | 4 | | | | | | | | | 5.19 | 25.3 | 50 |
| | 8 | | | | | | | | | 4.88 | 25.5 | 80 |
| | | 256 | | | 32 | 32 | | | | 5.75 | 24.5 | 28 |
| | | 1024 | | | 128 | 128 | | | | 4.66 | 26.0 | 168 |
| | | | 1024 | | | | | | | 5.12 | 25.4 | 53 |
| | | | 4096 | | | | | | | 4.75 | 26.2 | 90 |
| (D) | | | | | | | 0.0 | | | 5.77 | 24.6 | |
| | | | | | | | 0.2 | | | 4.95 | 25.5 | |
| | | | | | | | | 0.0 | | 4.67 | 25.3 | |
| | | | | | | | | 0.2 | | 5.47 | 25.7 | |
| (E) | | positional embedding instead of sinusoids | | | | | | | | 4.92 | 25.7 | |
| big | 6 | 1024 | 4096 | 16 | | | 0.3 | | 300K | **4.33** | **26.4** | 213 |

---

| | $N$ | $d_{\text{model}}$ | $d_{\text{ff}}$ | $h$ | $d_k$ | $d_v$ | $P_{drop}$ | $\epsilon_{ls}$ | etapas de treinamento | PPL (dev) | BLEU (dev) | parâmetros $\times 10^6$ |
| :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
| base | 6 | 512 | 2048 | 8 | 64 | 64 | 0.1 | 0.1 | 100K | 4,92 | 25,8 | 65 |
| (A) | | | | 1 | 512 | 512 | | | | 5,29 | 24,9 | |
| | | | | 4 | 128 | 128 | | | | 5,00 | 25,5 | |
| | | | | 16 | 32 | 32 | | | | 4,91 | 25,8 | |
| | | | | 32 | 16 | 16 | | | | 5,01 | 25,4 | |
| (B) | | | | 16 | | | | | | 5,16 | 25,1 | 58 |
| | | | | 32 | | | | | | 5,01 | 25,4 | 60 |
| (C) | 2 | | | | | | | | | 6,11 | 23,7 | 36 |
| | 4 | | | | | | | | | 5,19 | 25,3 | 50 |
| | 8 | | | | | | | | | 4,88 | 25,5 | 80 |
| | | 256 | | | 32 | 32 | | | | 5,75 | 24,5 | 28 |
| | | 1024 | | | 128 | 128 | | | | 4,66 | 26,0 | 168 |
| | | | 1024 | | | | | | | 5,12 | 25,4 | 53 |
| | | | 4096 | | | | | | | 4,75 | 26,2 | 90 |
| (D) | | | | | | | 0,0 | | | 5,77 | 24,6 | |
| | | | | | | | 0,2 | | | 4,95 | 25,5 | |
| | | | | | | | | 0,0 | | 4,67 | 25,3 | |
| | | | | | | | | 0,2 | | 5,47 | 25,7 | |
| (E) | | incorporação posicional em vez de senoides | | | | | | | | 4,92 | 25,7 | |
| grande | 6 | 1024 | 4096 | 16 | | | 0,3 | | 300K | **4,33** | **26,4** | 213 |

---

development set, newstest2013. We used beam search as described in the previous section, but no checkpoint averaging. We present these results in Table 3.

In Table 3 rows (A), we vary the number of attention heads and the attention key and value dimensions, keeping the amount of computation constant, as described in Section 3.2.2. While single-head attention is 0.9 BLEU worse than the best setting, quality also drops off with too many heads.

In Table 3 rows (B), we observe that reducing the attention key size $d_k$ hurts model quality. This suggests that determining compatibility is not easy and that a more sophisticated compatibility function than dot product may be beneficial. We further observe in rows (C) and (D) that, as expected, bigger models are better, and dropout is very helpful in avoiding over-fitting. In row (E) we replace our sinusoidal positional encoding with learned positional embeddings [9], and observe nearly identical results to the base model.

---

conjunto de desenvolvimento, newstest2013. Utilizamos busca em feixe conforme descrito na seção anterior, mas sem média de checkpoints. Apresentamos esses resultados na Tabela 3.

Na Tabela 3, nas linhas (A), variamos o número de cabeças de atenção e as dimensões das chaves e valores de atenção, mantendo a quantidade de computação constante, conforme descrito na Seção 3.2.2. Embora a atenção com uma única cabeça seja 0,9 BLEU pior do que a melhor configuração, a qualidade também diminui quando há muitas cabeças.

Na Tabela 3, nas linhas (B), observamos que reduzir o tamanho da chave de atenção $d_k$ prejudica a qualidade do modelo. Isso sugere que determinar a compatibilidade não é fácil e que uma função de compatibilidade mais sofisticada do que o produto escalar pode ser benéfica. Observamos ainda, nas linhas (C) e (D), que, como esperado, modelos maiores são melhores, e o dropout é muito útil para evitar overfitting. Na linha (E), substituímos nossa codificação posicional senoidal por embeddings posicionais aprendidos [9] e observamos resultados quase idênticos aos do modelo base.

---

### 6.3 English Constituency Parsing

---

### 6.3 Análise Sintática do Constituente em Inglês

---

To evaluate if the Transformer can generalize to other tasks we performed experiments on English constituency parsing. This task presents specific challenges: the output is subject to strong structural constraints and is significantly longer than the input. Furthermore, RNN sequence-to-sequence models have not been able to attain state-of-the-art results in small-data regimes [37].

We trained a 4-layer transformer with $d_{model} = 1024$ on the Wall Street Journal (WSJ) portion of the Penn Treebank [25], about 40K training sentences. We also trained it in a semi-supervised setting, using the larger high-confidence and BerkleyParser corpora from with approximately 17M sentences [37]. We used a vocabulary of 16K tokens for the WSJ only setting and a vocabulary of 32K tokens for the semi-supervised setting.

We performed only a small number of experiments to select the dropout, both attention and residual (section 5.4), learning rates and beam size on the Section 22 development set, all other parameters remained unchanged from the English-to-German base translation model. During inference, we

---

Para avaliar se o Transformer pode generalizar para outras tarefas, realizamos experimentos de análise sintática constituinte em inglês. Essa tarefa apresenta desafios específicos: a saída está sujeita a fortes restrições estruturais e é significativamente mais longa do que a entrada. Além disso, modelos RNN sequência-para-sequência não conseguiram atingir resultados de ponta em regimes com poucos dados [37].

Treinamos um Transformer de 4 camadas com $d_{model} = 1024$ na porção do Wall Street Journal (WSJ) do Penn Treebank [25], composta por cerca de 40 mil frases de treinamento. Também o treinamos em um cenário semi-supervisionado, utilizando os corpora maiores de alta confiança e BerkleyParser, com aproximadamente 17 milhões de frases [37]. Utilizamos um vocabulário de 16 mil tokens para a configuração apenas WSJ e um vocabulário de 32 mil tokens para a configuração semi-supervisionada.

Realizamos apenas um pequeno número de experimentos para selecionar o dropout, tanto no mecanismo de atenção quanto nos resíduos (seção 5.4), as taxas de aprendizado e o tamanho da feixe no conjunto de desenvolvimento da Seção 22; todos os outros parâmetros permaneceram inalterados em relação ao modelo base de tradução inglês-alemão. Durante a inferência, nós

---

Table 4: The Transformer generalizes well to English constituency parsing (Results are on Section 23 of WSJ)

---

Tabela 4: O Transformer generaliza bem para a análise de constituintes do inglês (Os resultados estão na Seção 23 do 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 |

---

| Analisador | Treinamento | WSJ 23 F1 |
| :---: | :---: | :---: |
| Vinyals & Kaiser et al. (2014) [37] | Apenas WSJ, discriminativo | 88,3 |
| Petrov et al. (2006) [29] | Apenas WSJ, discriminativo | 90,4 |
| Zhu et al. (2013) [40] | Apenas WSJ, discriminativo | 90,4 |
| Dyer et al. (2016) [8] | Apenas WSJ, discriminativo | 91,7 |
| Transformer (4 camadas) | Apenas WSJ, discriminativo | 91,3 |
| Zhu et al. (2013) [40] | semissupervisionado | 91,3 |
| Huang & Harper (2009) [14] | semissupervisionado | 91,3 |
| McClosky et al. (2006) [26] | semissupervisionado | 92,1 |
| Vinyals & Kaiser et al. (2014) [37] | semissupervisionado | 92,1 |
| Transformer (4 camadas) | semissupervisionado | 92,7 |
| Luong et al. (2015) [23] | multitarefa | 93,0 |
| Dyer et al. (2016) [8] | generativo | 93,3 |

---

increased the maximum output length to input length + 300. We used a beam size of 21 and $\alpha = 0.3$ for both WSJ only and the semi-supervised setting.

Our results in Table 4 show that despite the lack of task-specific tuning our model performs surprisingly well, yielding better results than all previously reported models with the exception of the Recurrent Neural Network Grammar [8].

In contrast to RNN sequence-to-sequence models [37], the Transformer outperforms the Berkeley-Parser [29] even when training only on the WSJ training set of 40K sentences.

---

aumentou o comprimento máximo de saída para comprimento de entrada + 300. Utilizamos um tamanho de feixe de 21 e $\alpha = 0.3$ tanto para o cenário apenas com WSJ quanto para o semi-supervisionado.

Nossos resultados na Tabela 4 mostram que, apesar da falta de ajuste específico para a tarefa, nosso modelo apresenta desempenho surpreendentemente bom, obtendo resultados superiores aos de todos os modelos relatados anteriormente, exceto a Gramática de Redes Neurais Recorrentes [8].

Em contraste com os modelos sequência-para-sequência baseados em RNNs [37], o Transformer supera o Berkeley-Parser [29] mesmo quando treinado apenas no conjunto de treinamento do WSJ, composto por 40K sentenças.

---

## 7 Conclusion

---

## 7 Conclusão

---

In this work, we presented the Transformer, the first sequence transduction model based entirely on attention, replacing the recurrent layers most commonly used in encoder-decoder architectures with multi-headed self-attention.

For translation tasks, the Transformer can be trained significantly faster than architectures based on recurrent or convolutional layers. On both WMT 2014 English-to-German and WMT 2014 English-to-French translation tasks, we achieve a new state of the art. In the former task our best model outperforms even all previously reported ensembles.

We are excited about the future of attention-based models and plan to apply them to other tasks. We plan to extend the Transformer to problems involving input and output modalities other than text and to investigate local, restricted attention mechanisms to efficiently handle large inputs and outputs such as images, audio and video. Making generation less sequential is another research goals of ours.

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

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

---

Neste trabalho, apresentamos o Transformer, o primeiro modelo de transdução de sequências baseado inteiramente em atenção, substituindo as camadas recorrentes mais comumente usadas em arquiteturas codificador-decodificador por autoatenção multi-cabeça.

Para tarefas de tradução, o Transformer pode ser treinado significativamente mais rápido do que arquiteturas baseadas em camadas recorrentes ou convolucionais. Em ambas as tarefas de tradução do WMT 2014 inglês-para-alemão e WMT 2014 inglês-para-francês, alcançamos um novo estado da arte. Na primeira tarefa, nosso melhor modelo supera até mesmo todos os conjuntos previamente relatados.

Estamos entusiasmados com o futuro dos modelos baseados em atenção e planejamos aplicá-los a outras tarefas. Planejamos estender o Transformer para problemas que envolvam modalidades de entrada e saída diferentes do texto, bem como investigar mecanismos de atenção local e restrita para lidar eficientemente com entradas e saídas grandes, como imagens, áudio e vídeo. Tornar a geração menos sequencial é outro objetivo de nossa pesquisa.

O código usado para treinar e avaliar nossos modelos está disponível em https://github.com/tensorflow/tensor2tensor.

**Agradecimentos** Somos gratos a Nal Kalchbrenner e Stephan Gouws por seus comentários frutíferos, correções e inspiração.

---

## References

---

## Referências

---

[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. *arXiv preprint arXiv:1607.06450*, 2016.

[2] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. *CoRR*, abs/1409.0473, 2014.

[3] Denny Britz, Anna Goldie, Minh-Thang Luong, and Quoc V. Le. Massive exploration of neural machine translation architectures. *CoRR*, abs/1703.03906, 2017.

[4] Jianpeng Cheng, Li Dong, and Mirella Lapata. Long short-term memory-networks for machine reading. *arXiv preprint arXiv:1601.06733*, 2016.

---

[1] Jimmy Lei Ba, Jamie Ryan Kiros e Geoffrey E Hinton. Normalização por camada. *Pré-impressão arXiv arXiv:1607.06450*, 2016.

[2] Dzmitry Bahdanau, Kyunghyun Cho e Yoshua Bengio. Tradução automática neural aprendendo simultaneamente o alinhamento e a tradução. *CoRR*, abs/1409.0473, 2014.

[3] Denny Britz, Anna Goldie, Minh-Thang Luong e Quoc V. Le. Exploração massiva de arquiteturas de tradução automática neural. *CoRR*, abs/1703.03906, 2017.

[4] Jianpeng Cheng, Li Dong e Mirella Lapata. Redes de memória de longo e curto prazo para leitura de máquinas. *Pré-impressão arXiv arXiv:1601.06733*, 2016.

---

[5] Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. *CoRR*, abs/1406.1078, 2014.

[6] Francois Chollet. Xception: Deep learning with depthwise separable convolutions. *arXiv preprint arXiv:1610.02357*, 2016.

[7] Junyoung Chung, Çağlar Gülçehre, Kyunghyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. *CoRR*, abs/1412.3555, 2014.

[8] Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, and Noah A. Smith. Recurrent neural network grammars. In *Proc. of NAACL*, 2016.

[9] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N. Dauphin. Convolutional sequence to sequence learning. *arXiv preprint arXiv:1705.03122v2*, 2017.

[10] Alex Graves. Generating sequences with recurrent neural networks. *arXiv preprint arXiv:1308.0850*, 2013.

[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In *Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition*, pages 770–778, 2016.

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

[13] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. *Neural computation*, 9(8):1735–1780, 1997.

[14] Zhongqiang Huang and Mary Harper. Self-training PCFG grammars with latent annotations across languages. In *Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing*, pages 832–841. ACL, August 2009.

[15] Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. Exploring the limits of language modeling. *arXiv preprint arXiv:1602.02410*, 2016.

[16] Łukasz Kaiser and Samy Bengio. Can active memory replace attention? In *Advances in Neural Information Processing Systems, (NIPS)*, 2016.

[17] Łukasz Kaiser and Ilya Sutskever. Neural GPUs learn algorithms. In *International Conference on Learning Representations (ICLR)*, 2016.

[18] Nal Kalchbrenner, Lasse Espeholt, Karen Simonyan, Aaron van den Oord, Alex Graves, and Koray Kavukcuoglu. Neural machine translation in linear time. *arXiv preprint arXiv:1610.10099v2*, 2017.

[19] Yoon Kim, Carl Denton, Luong Hoang, and Alexander M. Rush. Structured attention networks. In *International Conference on Learning Representations*, 2017.

[20] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In *ICLR*, 2015.

[21] Oleksii Kuchaiev and Boris Ginsburg. Factorization tricks for LSTM networks. *arXiv preprint arXiv:1703.10722*, 2017.

[22] Zhouhan Lin, Minwei Feng, Cicero Nogueira dos Santos, Mo Yu, Bing Xiang, Bowen Zhou, and Yoshua Bengio. A structured self-attentive sentence embedding. *arXiv preprint arXiv:1703.03130*, 2017.

[23] Minh-Thang Luong, Quoc V. Le, Ilya Sutskever, Oriol Vinyals, and Lukasz Kaiser. Multi-task sequence to sequence learning. *arXiv preprint arXiv:1511.06114*, 2015.

[24] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-based neural machine translation. *arXiv preprint arXiv:1508.04025*, 2015.

---

[5] Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Fethi Bougares, Holger Schwenk e Yoshua Bengio. Aprendizado de representações de frases usando codificador-decodificador RNN para tradução automática estatística. *CoRR*, abs/1406.1078, 2014.

[6] Francois Chollet. Xception: Aprendizado profundo com convoluções separáveis em profundidade. *arXiv preprint arXiv:1610.02357*, 2016.

[7] Junyoung Chung, Çağlar Gülçehre, Kyunghyun Cho e Yoshua Bengio. Avaliação empírica de redes neurais recorrentes com portas no modelamento de sequências. *CoRR*, abs/1412.3555, 2014.

[8] Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros e Noah A. Smith. Gramáticas de redes neurais recorrentes. In *Proc. of NAACL*, 2016.

[9] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats e Yann N. Dauphin. Aprendizado de sequência a sequência baseado em convolução. *arXiv preprint arXiv:1705.03122v2*, 2017.

[10] Alex Graves. Geração de sequências com redes neurais recorrentes. *arXiv preprint arXiv:1308.0850*, 2013.

[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren e Jian Sun. Aprendizado residual profundo para reconhecimento de imagens. In *Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition*, páginas 770–778, 2016.

[12] Sepp Hochreiter, Yoshua Bengio, Paolo Frasconi e Jürgen Schmidhuber. Fluxo de gradiente em redes recorrentes: a dificuldade de aprender dependências de longo prazo, 2001.

[13] Sepp Hochreiter e Jürgen Schmidhuber. Long short-term memory. *Neural computation*, 9(8):1735–1780, 1997.

[14] Zhongqiang Huang e Mary Harper. Treinamento autônomo de gramáticas PCFG com anotações latentes entre idiomas. In *Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing*, páginas 832–841. ACL, agosto de 2009.

[15] Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer e Yonghui Wu. Explorando os limites do modelamento de linguagem. *arXiv preprint arXiv:1602.02410*, 2016.

[16] Łukasz Kaiser e Samy Bengio. A memória ativa pode substituir a atenção? In *Advances in Neural Information Processing Systems, (NIPS)*, 2016.

[17] Łukasz Kaiser e Ilya Sutskever. GPUs neurais aprendem algoritmos. In *International Conference on Learning Representations (ICLR)*, 2016.

[18] Nal Kalchbrenner, Lasse Espeholt, Karen Simonyan, Aaron van den Oord, Alex Graves e Koray Kavukcuoglu. Tradução automática neural em tempo linear. *arXiv preprint arXiv:1610.10099v2*, 2017.

[19] Yoon Kim, Carl Denton, Luong Hoang e Alexander M. Rush. Redes de atenção estruturada. In *International Conference on Learning Representations*, 2017.

[20] Diederik Kingma e Jimmy Ba. Adam: Um método para otimização estocástica. In *ICLR*, 2015.

[21] Oleksii Kuchaiev e Boris Ginsburg. Truques de fatoração para redes LSTM. *arXiv preprint arXiv:1703.10722*, 2017.

[22] Zhouhan Lin, Minwei Feng, Cicero Nogueira dos Santos, Mo Yu, Bing Xiang, Bowen Zhou e Yoshua Bengio. Uma incorporação de sentença autoatenciva estruturada. *arXiv preprint arXiv:1703.03130*, 2017.

[23] Minh-Thang Luong, Quoc V. Le, Ilya Sutskever, Oriol Vinyals e Lukasz Kaiser. Aprendizado de sequência a sequência multitarefa. *arXiv preprint arXiv:1511.06114*, 2015.

[24] Minh-Thang Luong, Hieu Pham e Christopher D Manning. Abordagens eficazes para tradução automática neural baseada em atenção. *arXiv preprint arXiv:1508.04025*, 2015.

---

[25] Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. Building a large annotated corpus of english: The penn treebank. *Computational linguistics*, 19(2):313–330, 1993.

[26] David McClosky, Eugene Charniak, and Mark Johnson. Effective self-training for parsing. In *Proceedings of the Human Language Technology Conference of the NAACL, Main Conference*, pages 152–159. ACL, June 2006.

[27] Ankur Parikh, Oscar Täckström, Dipanjan Das, and Jakob Uszkoreit. A decomposable attention model. In *Empirical Methods in Natural Language Processing*, 2016.

[28] Romain Paulus, Caiming Xiong, and Richard Socher. A deep reinforced model for abstractive summarization. *arXiv preprint arXiv:1705.04304*, 2017.

[29] Slav Petrov, Leon Barrett, Romain Thibaux, and Dan Klein. Learning accurate, compact, and interpretable tree annotation. In *Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the ACL*, pages 433–440. ACL, July 2006.

[30] Ofir Press and Lior Wolf. Using the output embedding to improve language models. *arXiv preprint arXiv:1608.05859*, 2016.

[31] Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units. *arXiv preprint arXiv:1508.07909*, 2015.

[32] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. *arXiv preprint arXiv:1701.06538*, 2017.

[33] Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. *Journal of Machine Learning Research*, 15(1):1929–1958, 2014.

[34] Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. End-to-end memory networks. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors, *Advances in Neural Information Processing Systems 28*, pages 2440–2448. Curran Associates, Inc., 2015.

[35] Ilya Sutskever, Oriol Vinyals, and Quoc VV Le. Sequence to sequence learning with neural networks. In *Advances in Neural Information Processing Systems*, pages 3104–3112, 2014.

[36] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. *CoRR*, abs/1512.00567, 2015.

[37] Vinyals & Kaiser, Koo, Petrov, Sutskever, and Hinton. Grammar as a foreign language. In *Advances in Neural Information Processing Systems*, 2015.

[38] Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google’s neural machine translation system: Bridging the gap between human and machine translation. *arXiv preprint arXiv:1609.08144*, 2016.

[39] Jie Zhou, Ying Cao, Xuguang Wang, Peng Li, and Wei Xu. Deep recurrent models with fast-forward connections for neural machine translation. *CoRR*, abs/1606.04199, 2016.

[40] Muhua Zhu, Yue Zhang, Wenliang Chen, Min Zhang, and Jingbo Zhu. Fast and accurate shift-reduce constituent parsing. In *Proceedings of the 51st Annual Meeting of the ACL (Volume 1: Long Papers)*, pages 434–443. ACL, August 2013.

---

[25] Mitchell P Marcus, Mary Ann Marcinkiewicz e Beatrice Santorini. Construindo um grande corpus anotado de inglês: O Penn Treebank. *Linguística Computacional*, 19(2):313–330, 1993.

[26] David McClosky, Eugene Charniak e Mark Johnson. Auto-treinamento eficaz para análise sintática. In *Anais da Conferência de Tecnologia de Linguagem Humana da NAACL, Conferência Principal*, páginas 152–159. ACL, junho de 2006.

[27] Ankur Parikh, Oscar Täckström, Dipanjan Das e Jakob Uszkoreit. Um modelo de atenção decomponível. Em *Métodos Empíricos em Processamento de Linguagem Natural*, 2016.

[28] Romain Paulus, Caiming Xiong e Richard Socher. Um modelo profundo reforçado para sumarização abstrativa. *Pré-impressão arXiv arXiv:1705.04304*, 2017.

[29] Slav Petrov, Leon Barrett, Romain Thibaux e Dan Klein. Aprendendo anotação de árvore precisa, compacta e interpretável. In *Anais da 21ª Conferência Internacional sobre Linguística Computacional e 44ª Reunião Anual da ACL*, páginas 433–440. ACL, julho de 2006.

[30] Ofir Press e Lior Wolf. Usando o embedding de saída para melhorar modelos de linguagem. *Pré-impressão arXiv arXiv:1608.05859*, 2016.

[31] Rico Sennrich, Barry Haddow e Alexandra Birch. Tradução automática neural de palavras raras com unidades subword. *Pré-impressão arXiv arXiv:1508.07909*, 2015.

[32] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton e Jeff Dean. Redes neurais outrageosamente grandes: a camada mistura de especialistas esparsamente ativada. *Pré-impressão arXiv arXiv:1701.06538*, 2017.

[33] Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever e Ruslan Salakhutdinov. Dropout: uma maneira simples de evitar que redes neurais sofram overfitting. *Journal of Machine Learning Research*, 15(1):1929–1958, 2014.

[34] Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston e Rob Fergus. Redes de memória de ponta a ponta. Em C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama e R. Garnett, editores, *Avanços em Sistemas de Informação de Processamento Neural 28*, páginas 2440–2448. Curran Associates, Inc., 2015.

[35] Ilya Sutskever, Oriol Vinyals e Quoc VV Le. Aprendizado de sequência para sequência com redes neurais. Em *Avanços em Sistemas de Informação de Processamento Neural*, páginas 3104–3112, 2014.

[36] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens e Zbigniew Wojna. Repensando a arquitetura inception para visão computacional. *CoRR*, abs/1512.00567, 2015.

[37] Vinyals & Kaiser, Koo, Petrov, Sutskever e Hinton. Gramática como língua estrangeira. Em *Avanços em Sistemas de Informação de Processamento Neural*, 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. Sistema de tradução automática neural do Google: preenchendo a lacuna entre a tradução humana e a máquina. *Pré-impressão arXiv arXiv:1609.08144*, 2016.

[39] Jie Zhou, Ying Cao, Xuguang Wang, Peng Li e Wei Xu. Modelos recorrentes profundos com conexões de avanço rápido para tradução automática neural. *CoRR*, abs/1606.04199, 2016.

[40] Muhua Zhu, Yue Zhang, Wenliang Chen, Min Zhang e Jingbo Zhu. Análise constituinte shift-reduce rápida e precisa. In *Anais da 51ª Reunião Anual da ACL (Volume 1: Artigos Longos)*, páginas 434–443. ACL, agosto de 2013.

---

<img src="images/13-1.png" style="zoom:70%; display: block; margin: 0 auto;" />

Figure 3: An example of the attention mechanism following long-distance dependencies in the encoder self-attention in layer 5 of 6. Many of the attention heads attend to a distant dependency of the verb ‘making’, completing the phrase ‘making...more difficult’. Attentions here shown only for the word ‘making’. Different colors represent different heads. Best viewed in color.

---

Figura 3: Um exemplo do mecanismo de atenção seguindo dependências de longa distância na autoatenção do codificador na camada 5 de 6. Muitas das cabeças de atenção atendem a uma dependência distante do verbo ‘making’, completando a frase ‘making...more difficult’. As atenções aqui mostradas são apenas para a palavra ‘making’. Cores diferentes representam diferentes cabeças. Melhor visualizado em cores.

---

Figure 4: Two attention heads, also in layer 5 of 6, apparently involved in anaphora resolution. Top: Full attentions for head 5. Bottom: Isolated attentions from just the word 'its' for attention heads 5 and 6. Note that the attentions are very sharp for this word.

---

Figura 4: Duas cabeças de atenção, também na camada 5 de 6, aparentemente envolvidas na resolução da anáfora. Superior: Atenuações completas para a cabeça 5. Inferior: Atenuações isoladas apenas da palavra 'its' para as cabeças de atenção 5 e 6. Observe que as atenuações são muito agudas para esta palavra.

---

| | | | | | | | | | | | |
| :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
| The | The | The | The | | | | | | | | |
| Law | Law | Law | Law | | | | | | | | |
| will | will | will | will | | | | | | | | |
| never | never | never | never | | | | | | | | |
| be | be | be | be | | | | | | | | |
| perfect | perfect | perfect | perfect | | | | | | | | |
| , | , | , | , | | | | | | | | |
| but | but | but | but | | | | | | | | |
| its | its | its | its | | | | | | | | |
| application | application | application | application | | | | | | | | |
| should | should | should | should | | | | | | | | |
| be | be | be | be | | | | | | | | |
| just | just | just | just | | | | | | | | |
| - | - | - | - | | | | | | | | |
| this | this | this | this | | | | | | | | |
| is | is | is | is | | | | | | | | |
| what | what | what | what | | | | | | | | |
| we | we | we | we | | | | | | | | |
| are | are | are | are | | | | | | | | |
| missing | missing | missing | missing | | | | | | | | |
| , | , | , | , | | | | | | | | |
| in | in | in | in | | | | | | | | |
| my | my | my | my | | | | | | | | |
| opinion | opinion | opinion | opinion | | | | | | | | |
| . | . | . | . | | | | | | | | |
| <EOS> | <EOS> | <EOS> | <EOS> | | | | | | | | |
| <pad> | <pad> | <pad> | <pad> | | | | | | | | |

---

| | | | | | | | | | | | |
| :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
| A | A | A | A | | | | | | | | |
| Lei | Lei | Lei | Lei | | | | | | | | |
| nunca | nunca | nunca | nunca | | | | | | | | |
| será | será | será | será | | | | | | | | |
| perfeita | perfeita | perfeita | perfeita | | | | | | | | |
| , | , | , | , | | | | | | | | |
| mas | mas | mas | mas | | | | | | | | |
| sua | sua | sua | sua | | | | | | | | |
| aplicação | aplicação | aplicação | aplicação | | | | | | | | |
| deve | deve | deve | deve | | | | | | | | |
| ser | ser | ser | ser | | | | | | | | |
| justa | justa | justa | justa | | | | | | | | |
| - | - | - | - | | | | | | | | |
| esta | esta | esta | esta | | | | | | | | |
| é | é | é | é | | | | | | | | |
| o que | o que | o que | o que | | | | | | | | |
| nos | nos | nos | nos | | | | | | | | |
| falta | falta | falta | falta | | | | | | | | |
| , | , | , | , | | | | | | | | |
| na | na | na | na | | | | | | | | |
| minha | minha | minha | minha | | | | | | | | |
| opinião | opinião | opinião | opinião | | | | | | | | |
| . | . | . | . | | | | | | | | |
| <EOS> | <EOS> | <EOS> | <EOS> | | | | | | | | |
| <pad> | <pad> | <pad> | <pad> | | | | | | | | |

---

<img src="images/15-0.png" style="zoom:70%; display: block; margin: 0 auto;" />

Figure 5: Many of the attention heads exhibit behaviour that seems related to the structure of the sentence. We give two such examples above, from two different heads from the encoder self-attention at layer 5 of 6. The heads clearly learned to perform different tasks.

---

Figura 5: Muitos dos cabeçotes de atenção exibem comportamento que parece estar relacionado à estrutura da frase. Apresentamos dois exemplos acima, provenientes de dois cabeçotes diferentes da autoatenção do codificador na camada 5 de 6. Os cabeçotes claramente aprenderam a realizar tarefas distintas.

---
