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.

---

Sofern eine ordnungsgemäße Quellenangabe erfolgt, erteilt Google hiermit die Erlaubnis, die Tabellen und Abbildungen in diesem Dokument ausschließlich für journalistische oder wissenschaftliche Zwecke zu vervielfältigen.

---

# Attention Is All You Need

---

# Aufmerksamkeit ist alles, was Sie brauchen

---

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

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

Łukasz Kaiser*
Google Brain
lukaszkaiser@google.com

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

illia.polosukhin@gmail.com

## Abstract

---

## Zusammenfassung

---

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.

---

Die führenden Sequenz-zu-Sequenz-Transduktionsmodelle basieren auf komplexen rekurrenten oder konvolutionellen neuronalen Netzen, die einen Encoder und einen Decoder umfassen. Die besten Modelle verbinden zudem den Encoder und den Decoder über einen Aufmerksamkeitsmechanismus (Attention-Mechanismus). Wir schlagen eine neue, einfache Netzwerkarchitektur vor, den Transformer, der ausschließlich auf Aufmerksamkeitsmechanismen basiert und vollständig auf Rekurrenz und Konvolutionen verzichtet. Experimente zu zwei maschinellen Übersetzungsaufgaben zeigen, dass diese Modelle in der Qualität überlegen sind, besser parallelisierbar und deutlich weniger Trainingszeit benötigen. Unser Modell erzielt auf der WMT-2014-Aufgabe zur Englisch-Deutsch-Übersetzung einen BLEU-Wert von 28,4 und verbessert damit die bisherigen besten Ergebnisse, einschließlich Ensembles, um mehr als 2 BLEU-Punkte. Auf der WMT-2014-Aufgabe zur Englisch-Französisch-Übersetzung erreicht unser Modell nach dreieinhalb Tagen Training auf acht GPUs einen neuen Single-Model-State-of-the-Art-BLEU-Wert von 41,8, was nur ein kleiner Bruchteil der Trainingskosten der besten Modelle aus der Literatur ist. Wir zeigen, dass sich der Transformer gut auf andere Aufgaben verallgemeinern lässt, indem er erfolgreich für das englische Constituency-Parsing sowohl mit großen als auch mit begrenzten Trainingsdaten eingesetzt wird.

---

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

---

$^*$Gleichwertiger Beitrag. Die Reihenfolge der Auflistung ist zufällig. Jakob schlug vor, RNNs durch Self-Attention zu ersetzen, und initiierte die Bemühungen, diese Idee zu evaluieren. Ashish entwickelte zusammen mit Illia die ersten Transformer-Modelle und war in jedem Aspekt dieser Arbeit entscheidend beteiligt. Noam schlug skalierte Dot-Product-Aufmerksamkeit, Multi-Head-Aufmerksamkeit und die parameterfreie Positionsrepräsentation vor und war die andere Person, die fast in jedem Detail involviert war. Niki entwarf, implementierte, optimierte und evaluierte unzählige Modellvarianten in unserem ursprünglichen Codebase und tensor2tensor. Llion experimentierte ebenfalls mit neuartigen Modellvarianten, war für unseren initialen Codebase sowie effizientes Inference und Visualisierungen verantwortlich. Lukasz und Aidan verbrachten unzählige lange Tage damit, verschiedene Teile von tensor2tensor zu entwerfen und zu implementieren, wodurch unser früheres Codebase ersetzt wurde, die Ergebnisse erheblich verbessert und unsere Forschung massiv beschleunigt wurden.

$^\dagger$Arbeit durchgeführt während der Zeit bei Google Brain.

$^\ddagger$Arbeit durchgeführt während der Zeit bei Google Research.

---

## 1 Introduction

---

## 1 Einleitung

---

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.

---

Rekurrente neuronale Netze, insbesondere Long Short-Term Memory [13] und gated Recurrent Neural Networks [7], haben sich als State-of-the-Art-Ansätze bei Sequenzmodellierungs- und Transduktionsproblemen wie Language Modeling und maschineller Übersetzung [35, 2, 5] etabliert. Seither wurden zahlreiche Anstrengungen unternommen, um die Grenzen rekurrenter Sprachmodelle und Encoder-Decoder-Architekturen weiter voranzutreiben [38, 24, 15].

Rekurrente Modelle faktorisieren die Berechnung typischerweise entlang der Symbolpositionen der Eingabe- und Ausgabesequenzen. Durch die Zuordnung der Positionen zu Schritten in der Berechnungszeit erzeugen sie eine Folge von versteckten Zuständen $h_t$, die als Funktion des vorherigen versteckten Zustands $h_{t-1}$ und der Eingabe für die Position $t$ definiert ist. Diese inhärent sequenzielle Natur schließt die Parallelisierung innerhalb einzelner Trainingsbeispiele aus, was bei längeren Sequenzlängen kritisch wird, da Speicherbeschränkungen das Batching über Beispiele hinweg begrenzen. Jüngste Arbeiten haben durch Faktorisierungstricks [21] und bedingte Berechnung [32] erhebliche Verbesserungen der Recheneffizienz erzielt und dabei im letzteren Fall auch die Modellleistung gesteigert. Die grundlegende Einschränkung der sequenziellen Berechnung bleibt jedoch bestehen.

Aufmerksamkeitsmechanismen (Attention Mechanisms) sind zu einem integralen Bestandteil überzeugender Sequenzmodellierungs- und Transduktionsmodelle in verschiedenen Aufgaben geworden und ermöglichen die Modellierung von Abhängigkeiten ohne Berücksichtigung ihrer Entfernung in den Eingabe- oder Ausgabesequenzen [2, 19]. In allen bis auf wenigen Fällen [27] werden solche Aufmerksamkeitsmechanismen jedoch in Verbindung mit einem rekurrenten Netzwerk eingesetzt.

In dieser Arbeit schlagen wir den Transformer vor, eine Modellarchitektur, die auf Rekurrenz verzichtet und stattdessen vollständig auf einen Aufmerksamkeitsmechanismus zurückgreift, um globale Abhängigkeiten zwischen Eingabe und Ausgabe herzustellen. Der Transformer ermöglicht eine deutlich höhere Parallelisierbarkeit und erreicht nach dem Training für nur zwölf Stunden auf acht P100-GPUs einen neuen State of the Art in der Übersetzungsqualität.

---

## 2 Background

---

## 2 Hintergrund

---

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].

---

Das Ziel, die sequenzielle Berechnung zu reduzieren, bildet auch die Grundlage des Extended Neural GPU [16], ByteNet [18] und ConvS2S [9], die alle Faltungsneuronale Netze als grundlegende Bausteine verwenden und verborgene Darstellungen parallel für alle Eingabe- und Ausgabepositionen berechnen. In diesen Modellen wächst die Anzahl der Operationen, die erforderlich ist, um Signale von zwei beliebigen Eingabe- oder Ausgabepositionen in Beziehung zu setzen, mit dem Abstand zwischen den Positionen – linear bei ConvS2S und logarithmisch bei ByteNet. Dies erschwert das Erlernen von Abhängigkeiten zwischen weit entfernten Positionen [12]. Beim Transformer wird dies auf eine konstante Anzahl von Operationen reduziert, albeit auf Kosten einer verringerten effektiven Auflösung aufgrund der gemittelten aufmerksamkeitsgewichteten Positionen; diesem Effekt wirken wir durch Multi-Head Attention entgegen, wie in Abschnitt 3.2 beschrieben.

Selbst-Aufmerksamkeit, manchmal auch als intra-Aufmerksamkeit bezeichnet, ist ein Aufmerksamkeitsmechanismus, der verschiedene Positionen eines einzelnen Sequenzabschnitts miteinander in Beziehung setzt, um eine Darstellung der Sequenz zu berechnen. Selbst-Aufmerksamkeit wurde erfolgreich in einer Vielzahl von Aufgaben eingesetzt, darunter Leseverstehen, abstraktive Zusammenfassung, textuelle Folgerung und das Erlernen aufgabenunabhängiger Satzdarstellungen [4, 27, 28, 22].

End-to-End-Speichernetzwerke basieren auf einem rekurrenten Aufmerksamkeitsmechanismus anstelle von sequenzausgerichteter Rekursion und haben sich als leistungsfähig bei einfachen Fragenbeantwortungs- und Sprachmodellierungsaufgaben erwiesen [34].

Nach bestem Wissen unsererseits ist der Transformer jedoch das erste Transduktionsmodell, das vollständig auf Selbst-Aufmerksamkeit zurückgreift, um Darstellungen seiner Eingabe und Ausgabe zu berechnen, ohne sequenzausgerichtete RNNs oder Faltungen zu verwenden. In den folgenden Abschnitten werden wir den Transformer beschreiben, die Selbst-Aufmerksamkeit begründen und ihre Vorteile im Vergleich zu Modellen wie [17, 18] und [9] diskutieren.

---

## 3 Model Architecture

---

## 3 Modellarchitektur

---

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.

---

Die meisten wettbewerbsfähigen neuronalen Sequenztransduktionsmodelle besitzen eine Encoder-Decoder-Architektur [5, 2, 35]. Dabei bildet der Encoder eine Eingabesequenz von Symbolrepräsentationen $(x_1, ..., x_n)$ auf eine Sequenz kontinuierlicher Repräsentationen $\mathbf{z} = (z_1, ..., z_n)$ ab. Auf Basis von $\mathbf{z}$ erzeugt der Decoder dann eine Ausgabesequenz $(y_1, ..., y_m)$ von Symbolen schrittweise, d. h. ein Element nach dem anderen. In jedem Schritt ist das Modell autoregressiv [10], wobei es die zuvor generierten Symbole als zusätzliche Eingabe verwendet, um das nächste zu erzeugen.

---

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

Figure 1: The Transformer - model architecture.

---

Abbildung 1: Die Transformer-Modellarchitektur.

---

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.

---

Der Transformer folgt dieser allgemeinen Architektur, die gestapelte Selbst-Aufmerksamkeitsmechanismen und punktweise, vollständig verbundene Schichten sowohl für den Encoder als auch für den Decoder verwendet, wie in den linken und rechten Hälften von Abbildung 1 dargestellt.

---

## 3.1 Encoder and Decoder Stacks

---

## 3.1 Encoder- und Decoder-Stapel

---

**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 $i$.

---

**Encoder:** Der Encoder besteht aus einem Stapel von $N = 6$ identischen Schichten. Jede Schicht verfügt über zwei Sub-Schichten. Die erste ist ein Multi-Head-Selbst-Aufmerksamkeitsmechanismus, und die zweite ist ein einfaches, positionsweises vollverbundenes Feed-Forward-Netzwerk. Wir verwenden um jede der beiden Sub-Schichten herum eine Residualverbindung [11], gefolgt von einer Layer-Normalisierung [1]. Das heißt, die Ausgabe jeder Sub-Schicht ist $\text{LayerNorm}(x + \text{Sublayer}(x))$, wobei $\text{Sublayer}(x)$ die Funktion ist, die von der Sub-Schicht selbst implementiert wird. Um diese Residualverbindungen zu erleichtern, erzeugen alle Sub-Schichten im Modell sowie die Embedding-Schichten Ausgaben der Dimension $d_{\text{model}} = 512$.

**Decoder:** Der Decoder besteht ebenfalls aus einem Stapel von $N = 6$ identischen Schichten. Zusätzlich zu den beiden Sub-Schichten in jeder Encoderschicht fügt der Decoder eine dritte Sub-Schicht ein, die eine Multi-Head-Aufmerksamkeit über die Ausgabe des Encoder-Stapels durchführt. Ähnlich wie beim Encoder verwenden wir Residualverbindungen um jede der Sub-Schichten herum, gefolgt von einer Layer-Normalisierung. Wir modifizieren auch die Selbst-Aufmerksamkeits-Sub-Schicht im Decoder-Stapel, um zu verhindern, dass Positionen auf nachfolgende Positionen achten können. Diese Maskierung, kombiniert mit der Tatsache, dass die Ausgabe-Embeddings um eine Position versetzt sind, stellt sicher, dass die Vorhersagen für die Position $i$ nur von den bekannten Ausgaben an Positionen kleiner als $i$ abhängen können.

---

## 3.2 Attention

---

## 3.2 Aufmerksamkeit

---

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

---

Eine Aufmerksamkeitsfunktion kann als Abbildung einer Anfrage und einer Menge von Schlüssel-Wert-Paaren auf eine Ausgabe beschrieben werden, wobei die Anfrage, die Schlüssel, die Werte und die Ausgabe alle Vektoren sind. Die Ausgabe wird als gewichtete Summe berechnet

---

<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.

---

Abbildung 2: (links) Skalierte Punktprodukt-Aufmerksamkeit. (rechts) Multi-Head-Aufmerksamkeit besteht aus mehreren parallel ausgeführten Aufmerksamkeitslagen.

---

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

---

der Werte, wobei das jedem Wert zugewiesene Gewicht durch eine Kompatibilitätsfunktion der Abfrage mit dem entsprechenden Schlüssel berechnet wird.

---

## 3.2.1 Scaled Dot-Product Attention

---

## 3.2.1 Skalierte Punktprodukt-Aufmerksamkeit

---

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 $\frac{1}{\sqrt{d_k}}$.

---

Wir bezeichnen unsere spezielle Aufmerksamkeit als „Skalierte Punktprodukt-Aufmerksamkeit“ (Abbildung 2). Die Eingabe besteht aus Queries und Keys der Dimension $d_k$ sowie Values der Dimension $d_v$. Wir berechnen die Punktprodukte der Query mit allen Keys, teilen jedes durch $\sqrt{d_k}$ und wenden eine Softmax-Funktion an, um die Gewichte auf den Values zu erhalten.

In der Praxis berechnen wir die Aufmerksamkeitsfunktion für eine Menge von Queries gleichzeitig, die gemeinsam in einer Matrix $Q$ gepackt sind. Die Keys und Values werden ebenfalls in den Matrizen $K$ und $V$ zusammengefasst. Wir berechnen die Ausgabe-Matrix wie folgt:

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

Die zwei am häufigsten verwendeten Aufmerksamkeitsfunktionen sind additive Aufmerksamkeit [2] und Punktprodukt-(multiplikative) Aufmerksamkeit. Punktprodukt-Aufmerksamkeit ist identisch mit unserem Algorithmus, abgesehen vom Skalierungsfaktor $\frac{1}{\sqrt{d_k}}$. Additive Aufmerksamkeit berechnet die Kompatibilitätsfunktion unter Verwendung eines Feed-Forward-Netzwerks mit einer einzigen versteckten Schicht. Obwohl beide Ansätze theoretisch ähnliche Komplexität aufweisen, ist die Punktprodukt-Aufmerksamkeit in der Praxis deutlich schneller und speichereffizienter, da sie mithilfe hochoptimierter Code-Bibliotheken für Matrixmultiplikationen implementiert werden kann.

Während sich die beiden Mechanismen bei kleinen Werten von $d_k$ ähnlich verhalten, übertrifft additive Aufmerksamkeit die unskalierte Punktprodukt-Aufmerksamkeit bei größeren Werten von $d_k$ [3]. Wir vermuten, dass bei großen Werten von $d_k$ die Punktprodukte betragsmäßig groß werden und die Softmax-Funktion in Bereiche drängen, in denen sie extrem kleine Gradienten $^4$ aufweist. Um diesem Effekt entgegenzuwirken, skalieren wir die Punktprodukte mit $\frac{1}{\sqrt{d_k}}$.

---

## 3.2.2 Multi-Head Attention

---

## 3.2.2 Multi-Head-Aufmerksamkeit

---

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

---

Anstatt eine einzelne Aufmerksamkeitsfunktion mit $d_{\text{model}}$-dimensionalen Schlüsseln, Werten und Anfragen durchzuführen, haben wir festgestellt, dass es vorteilhaft ist, die Anfragen, Schlüssel und Werte $h$-mal mit unterschiedlichen, gelernten linearen Projektionen auf die Dimensionen $d_k$, $d_k$ und $d_v$ linear zu projizieren. Für jede dieser projizierten Versionen von Anfragen, Schlüsseln und Werten führen wir dann die Aufmerksamkeitsfunktion parallel aus, was zu $d_v$-dimensionalen

---

$^4$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 $d_k$.

---

$^4$Um zu veranschaulichen, warum die Skalarprodukte groß werden, nehmen wir an, dass die Komponenten von $q$ und $k$ unabhängige Zufallsvariablen mit Erwartungswert 0 und Varianz 1 sind. Dann hat ihr Skalarprodukt $q \cdot k = \sum_{i=1}^{d_k} q_i k_i$ den Erwartungswert 0 und die Varianz $d_k$.

---

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.

---

Ausgabewerte. Diese werden verkettet und erneut projiziert, was zu den endgültigen Werten führt, wie in Abbildung 2 dargestellt.

Multi-Head-Aufmerksamkeit ermöglicht es dem Modell, gleichzeitig Informationen aus verschiedenen Repräsentationsunterräumen an unterschiedlichen Positionen zu berücksichtigen. Bei einem einzelnen Aufmerksamkeitskopf würde eine Mittelwertbildung dies verhindern.

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

Die Projektionen sind Parametermatrizen $W_i^Q \in \mathbb{R}^{d_{\text{model}} \times d_k}$, $W_i^K \in \mathbb{R}^{d_{\text{model}} \times d_k}$, $W_i^V \in \mathbb{R}^{d_{\text{model}} \times d_v}$ und $W^O \in \mathbb{R}^{hd_v \times d_{\text{model}}}$.

In dieser Arbeit verwenden wir $h = 8$ parallele Aufmerksamkeitslagen bzw. -köpfe. Für jeden davon setzen wir $d_k = d_v = d_{\text{model}}/h = 64$ ein. Aufgrund der reduzierten Dimension jedes Kopfes ist die gesamte Rechenkosten ähnlich wie bei der Single-Head-Aufmerksamkeit mit voller Dimensionalität.

---

## 3.2.3 Applications of Attention in our Model

---

## 3.2.3 Anwendungen der Aufmerksamkeit in unserem Modell

---

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.

---

Der Transformer verwendet Multi-Head-Aufmerksamkeit auf drei verschiedene Arten:

*   In den „Encoder-Decoder-Aufmerksamkeits“-Schichten stammen die Queries aus der vorherigen Decoder-Schicht, und die Keys und Values kommen aus der Ausgabe des Encoders. Dies ermöglicht es jeder Position im Decoder, über alle Positionen in der Eingabesequenz zu achten. Dies ahmt die typischen Encoder-Decoder-Aufmerksamkeitsmechanismen in Sequence-to-Sequence-Modellen nach, wie in [38, 2, 9] beschrieben.
*   Der Encoder enthält Self-Attention-Schichten. In einer Self-Attention-Schicht stammen alle Keys, Values und Queries aus derselben Stelle, in diesem Fall aus der Ausgabe der vorherigen Schicht im Encoder. Jede Position im Encoder kann auf alle Positionen in der vorherigen Schicht des Encoders achten.
*   Ebenso ermöglichen Self-Attention-Schichten im Decoder, dass jede Position im Decoder auf alle Positionen im Decoder bis einschließlich dieser Position achten kann. Wir müssen einen linken Informationsfluss im Decoder verhindern, um die autoregressive Eigenschaft zu wahren. Wir implementieren dies innerhalb der skalierten Dot-Product-Aufmerksamkeit, indem wir alle Werte in der Eingabe der Softmax-Funktion, die illegalen Verbindungen entsprechen, maskieren (auf $-\infty$ setzen). Siehe Abbildung 2.

---

## 3.3 Position-wise Feed-Forward Networks

---

## 3.3 Positionsweise Feed-Forward-Netze

---

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 $d_{ff} = 2048$.

---

Neben den Aufmerksamkeits-Teilschichten enthält jede Schicht unseres Encoders und Decoders ein vollvernetztes Feed-Forward-Netzwerk, das separat und identisch auf jede Position angewendet wird. Dieses besteht aus zwei linearen Transformationen mit einer ReLU-Aktivierung dazwischen.

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

Obwohl die linearen Transformationen über verschiedene Positionen hinweg gleich sind, verwenden sie von Schicht zu Schicht unterschiedliche Parameter. Eine andere Beschreibung davon ist als zwei Faltungen mit der Kernelgröße 1. Die Dimensionalität von Eingabe und Ausgabe beträgt $d_{\text{model}} = 512$, und die innere Schicht hat die Dimensionalität $d_{ff} = 2048$.

---

## 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 $\sqrt{d_{\text{model}}}$.

---

Ähnlich wie bei anderen Sequenztransduktionsmodellen verwenden wir gelernte Embeddings, um die Eingabe-Token und Ausgabe-Token in Vektoren der Dimension $d_{\text{model}}$ umzuwandeln. Wir verwenden außerdem die übliche gelernte lineare Transformation und die Softmax-Funktion, um die Decoder-Ausgabe in vorhergesagte Wahrscheinlichkeiten für das nächste Token umzuwandeln. In unserem Modell teilen wir dieselbe Gewichtsmatrix zwischen den beiden Embedding-Schichten und der vor-Softmax-linearen Transformation, ähnlich wie in [30]. In den Embedding-Schichten multiplizieren wir diese Gewichte mit $\sqrt{d_{\text{model}}}$.

---

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.

---

Tabelle 1: Maximale Pfadlängen, Komplexität pro Schicht und minimale Anzahl sequenzieller Operationen für verschiedene Schichttypen. $n$ ist die Sequenzlänge, $d$ die Repräsentationsdimension, $k$ die Kernelgröße der Faltungen und $r$ die Größe der Nachbarschaft bei eingeschränkter Selbst-Aufmerksamkeit.

---

| 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) | $O(r \cdot n \cdot d)$ | $O(1)$ | $O(n/r)$ |

---

| Layer-Typ | Komplexität pro Layer | Sequentielle Operationen | Maximale Pfadlänge |
| :---: | :---: | :---: | :---: |
| Self-Attention | $O(n^2 \cdot d)$ | $O(1)$ | $O(1)$ |
| Rekurrent | $O(n \cdot d^2)$ | $O(n)$ | $O(n)$ |
| Faltungsnetzwerk (Convolutional) | $O(k \cdot n \cdot d^2)$ | $O(1)$ | $O(log_k(n))$ |
| Self-Attention (eingeschränkt) | $O(r \cdot n \cdot d)$ | $O(1)$ | $O(n/r)$ |

---

## 3.5 Positional Encoding

---

## 3.5 Positionskodierung

---

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.

---

Da unser Modell keine Rekursion und keine Faltung enthält, müssen wir dem Modell Informationen über die relative oder absolute Position der Token in der Sequenz zuführen, damit es die Reihenfolge der Sequenz nutzen kann. Zu diesem Zweck fügen wir den Eingabe-Embeddings an den unteren Enden der Encoder- und Decoder-Stapel „Positionsencodings“ hinzu. Die Positionsencodings haben dieselbe Dimension $d_{\text{model}}$ wie die Embeddings, sodass beide addiert werden können. Es gibt viele Möglichkeiten für Positionsencodings, sowohl gelernte als auch feste [9].

In dieser Arbeit verwenden wir Sinus- und Kosinusfunktionen unterschiedlicher Frequenzen:

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

wobei $pos$ die Position und $i$ die Dimension ist. Das heißt, jede Dimension des Positionsencodings entspricht einer Sinuskurve. Die Wellenlängen bilden eine geometrische Progression von $2\pi$ bis $10000 \cdot 2\pi$. Wir haben uns für diese Funktion entschieden, weil wir die Hypothese aufstellten, dass sie es dem Modell ermöglichen würde, leicht das Aufmerksamkeitsmechanismus nach relativen Positionen zu lernen, da sich für jeden festen Offset $k$ $PE_{pos+k}$ als lineare Funktion von $PE_{pos}$ darstellen lässt.

Wir experimentierten außerdem mit der Verwendung gelernter Positions-Embeddings [9] und stellten fest, dass beide Versionen nahezu identische Ergebnisse lieferten (siehe Tabelle 3, Zeile (E)). Wir wählten die sinusförmige Version, weil sie es dem Modell möglicherweise ermöglicht, auf Sequenzlängen zu extrapolieren, die länger sind als die während des Trainings beobachteten.

---

## 4 Why Self-Attention

---

## 4 Warum 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 $(x_1, ..., x_n)$ 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

---

In diesem Abschnitt vergleichen wir verschiedene Aspekte von Self-Attention-Schichten mit den häufig zur Abbildung eines variablen Sequenzen von Symbolsymbolrepräsentationen $(x_1, ..., x_n)$ auf eine weitere Sequenz gleicher Länge $(z_1, ..., z_n)$ verwendeten rekurrenten und Faltungsschichten, wobei $x_i, z_i \in \mathbb{R}^d$, wie etwa einer verborgenen Schicht in einem typischen Encoder oder Decoder der sequenziellen Transduktion. Zur Motivation unseres Einsatzes von Self-Attention betrachten wir drei desiderata.

Das erste ist die gesamte Rechenkomplexität pro Schicht. Das zweite ist der Grad der Parallelisierbarkeit der Berechnungen, gemessen an der minimalen Anzahl sequentieller Operationen, die erforderlich sind.

Das dritte ist die Pfadlänge zwischen langreichweitigen Abhängigkeiten im Netzwerk. Das Erlernen langreichweitiger Abhängigkeiten ist eine zentrale Herausforderung bei vielen Aufgaben der sequenziellen Transduktion. Ein entscheidender Faktor, der die Fähigkeit beeinflusst, solche Abhängigkeiten zu lernen, ist die Länge der Pfade, die Vorwärts- und Rückwärtssignale im Netzwerk zurücklegen müssen. Je kürzer diese Pfade zwischen beliebigen Kombinationen von Positionen in den Eingabe- und Ausgabesequenzen sind, desto einfacher ist es, langreichweitige Abhängigkeiten zu lernen [12]. Daher vergleichen wir auch die maximale Pfadlänge zwischen beliebigen Eingabe- und Ausgabepositionen in Netzwerken, die aus den verschiedenen Schichttypen zusammengesetzt sind.

Wie in Tabelle 1 vermerkt, verbindet eine Self-Attention-Schicht alle Positionen mit einer konstanten Anzahl sequentiell ausgeführter Operationen, während eine rekurrente Schicht $O(n)$ sequentielle Operationen erfordert. Hinsichtlich der Rechenkomplexität sind Self-Attention-Schichten schneller als rekurrente Schichten, wenn die Sequenz

---

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.

---

Die Länge $n$ ist kleiner als die Darstellungsdimensionalität $d$, was bei Satzdarstellungen, die von state-of-the-art-Modellen im maschinellen Übersetzen verwendet werden, am häufigsten der Fall ist, wie etwa Word-Piece-[38]- und Byte-Pair-[31]-Darstellungen. Um die Rechenleistung für Aufgaben mit sehr langen Sequenzen zu verbessern, könnte die Selbst-Aufmerksamkeit darauf beschränkt werden, nur eine Nachbarschaft der Größe $r$ in der Eingabesequenz zu berücksichtigen, die um die jeweilige Ausgabeposition zentriert ist. Dies würde die maximale Pfadlänge auf $O(n/r)$ erhöhen. Wir planen, diesen Ansatz in zukünftigen Arbeiten weiter zu untersuchen.

Eine einzelne Faltungsschicht mit Kernel-Breite $k < n$ verbindet nicht alle Paare von Eingabe- und Ausgabepositionen. Dies erfordert einen Stapel von $O(n/k)$ Faltungsschichten im Fall zusammenhängender Kernel oder $O(\log_k(n))$ im Fall dilatierter Faltungen [18], wodurch die Länge der längsten Pfade zwischen beliebigen zwei Positionen im Netzwerk erhöht wird. Faltungsschichten sind im Allgemeinen um den Faktor $k$ teurer als rekursive Schichten. Separable Faltungen [6] reduzieren die Komplexität jedoch erheblich auf $O(k \cdot n \cdot d + n \cdot d^2)$. Selbst bei $k = n$ entspricht die Komplexität einer separablen Faltung der Kombination aus einer Selbst-Aufmerksamkeitsschicht und einer punktuellen feed-forward-Schicht, dem Ansatz, den wir in unserem Modell verwenden.

Als zusätzlichen Vorteil könnten Selbst-Aufmerksamkeiten zu interpretierbareren Modellen führen. Wir untersuchen die Aufmerksamkeitsverteilungen unserer Modelle und präsentieren sowie diskutieren Beispiele im Anhang. Nicht nur lernen einzelne Aufmerksamkeitsköpfe eindeutig, verschiedene Aufgaben auszuführen, sondern viele scheinen auch Verhalten zu zeigen, das mit der syntaktischen und semantischen Struktur der Sätze zusammenhängt.

---

## 5 Training

This section describes the training regime for our models.

---

Dieser Abschnitt beschreibt das Trainingsverfahren für unsere Modelle.

---

## **5.1** Training Data and Batching

---

## **5.1 Trainingsdaten und Batching**

---

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.

---

Wir haben auf dem Standard-WMT-2014-Datensatz für Englisch-Deutsch trainiert, der aus etwa 4,5 Millionen Satzpaares besteht. Die Sätze wurden mit Byte-Pair-Encoding [3] kodiert, das ein gemeinsames Quell-Ziel-Vokabular von etwa 37.000 Token verwendet. Für Englisch-Französisch nutzten wir den deutlich größeren WMT-2014-Datensatz für Englisch-Französisch, der aus 36 Millionen Sätzen besteht, und teilten die Token in ein Wordpiece-Vokabular mit 32.000 Einträgen auf [38]. Satzpares wurden basierend auf ihrer ungefähren Sequenzlänge gruppiert. Jeder Trainingsbatch enthielt eine Menge an Satzpaares, die ungefähr 25.000 Quell-Token und 25.000 Ziel-Token umfasste.

---

## **5.2 Hardware and Schedule**

---

## **5.2 Hardware und Zeitplan**

---

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).

---

Wir haben unsere Modelle auf einer Maschine mit 8 NVIDIA P100 GPUs trainiert. Für unsere Basismodelle, die die im gesamten Papier beschriebenen Hyperparameter verwenden, dauerte jeder Trainingsschritt etwa 0,4 Sekunden. Wir haben die Basismodelle insgesamt für 100.000 Schritte oder 12 Stunden trainiert. Für unsere großen Modelle (beschrieben in der untersten Zeile von Tabelle 3) betrug die Schrittzeit 1,0 Sekunde. Die großen Modelle wurden für 300.000 Schritte (3,5 Tage) trainiert.

---

## 5.3 Optimizer

---

## 5.3 Optimierer

---

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:

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

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 $warmup\_steps = 4000$.

---

Wir verwendeten den Adam-Optimizer [20] mit $\beta_1 = 0.9$, $\beta_2 = 0.98$ und $\epsilon = 10^{-9}$. Wir variierten die Lernrate während des Trainings gemäß der Formel:

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

Dies entspricht einer linearen Erhöhung der Lernrate für die ersten *warmup_steps* Trainingsschritte und einer darauffolgenden Verringerung proportional zur inversen Quadratwurzel der Schrittnummer. Wir verwendeten $warmup\_steps = 4000$.

---

## 5.4 Regularization

---

## 5.4 Regularisierung

---

We employ three types of regularization during training:

---

Wir verwenden während des Trainings drei Arten von Regularisierung:

---

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.

---

Tabelle 2: Der Transformer erzielt auf den Newstest2014-Tests für Englisch-Deutsch und Englisch-Französisch bessere BLEU-Werte als die bisherigen State-of-the-Art-Modelle, und das zu einem Bruchteil der Trainingskosten.

---

| Model | EN-DE | EN-FR | EN-DE | EN-FR |
| :---: | :---: | :---: | :---: | :---: |
| | BLEU | | Training Cost (FLOPs) | |
| 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 $P_{drop} = 0.1$.

---

**Residual Dropout** Wir wenden Dropout [33] auf die Ausgabe jeder Sub-Schicht an, bevor sie zur Eingabe der Sub-Schicht addiert und normalisiert wird. Darüber hinaus wenden wir Dropout auf die Summen der Embeddings und der Positionsencodings in beiden Stapeln des Encoders und Decoders an. Für das Basismodell verwenden wir eine Rate von $P_{drop} = 0.1$.

---

**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.

---

**Label Smoothing** Während des Trainings verwendeten wir Label Smoothing mit dem Wert $\epsilon_{ls} = 0.1$ [36]. Dies verschlechtert die Perplexität, da das Modell lernt, unsicherer zu sein, verbessert jedoch die Genauigkeit und den BLEU-Score.

---

## 6 Results

---

## 6 Ergebnisse

---

## 6.1 Machine Translation

---

## 6.1 Maschinelle Übersetzung

---

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 $^5$.

---

Auf der WMT-2014-Aufgabe zur Übersetzung von Englisch nach Deutsch übertrifft das große Transformer-Modell (Transformer (big) in Tabelle 2) die zuvor besten berichteten Modelle (einschließlich Ensembles) um mehr als 2,0 BLEU und erzielt damit einen neuen State-of-the-Art-BLEU-Score von 28,4. Die Konfiguration dieses Modells ist in der letzten Zeile von Tabelle 3 aufgeführt. Das Training dauerte 3,5 Tage auf 8 P100-GPUs. Selbst unser Basismodell übertrifft alle zuvor veröffentlichten Modelle und Ensembles bei einem Bruchteil der Trainingskosten aller wettbewerbsfähigen Modelle.

Auf der WMT-2014-Aufgabe zur Übersetzung von Englisch nach Französisch erreicht unser großes Modell einen BLEU-Score von 41,0 und übertrifft damit alle zuvor veröffentlichten Einzelmodelle, und zwar bei weniger als $1/4$ der Trainingskosten des vorherigen State-of-the-Art-Modells. Das für die Englisch-nach-Französisch-Übersetzung trainierte Transformer-(big)-Modell verwendete eine Dropout-Rate von $P_{drop} = 0,1$ statt 0,3.

Für die Basismodelle verwendeten wir ein einzelnes Modell, das durch Mittelung der letzten 5 Checkpoints erhalten wurde, die in 10-Minuten-Abständen gespeichert wurden. Für die großen Modelle mittelten wir die letzten 20 Checkpoints. Wir verwendeten Beam Search mit einer Beam-Größe von 4 und einer Längenstrafe $\alpha = 0,6$ [38]. Diese Hyperparameter wurden nach Experimenten auf dem Entwicklungssatz gewählt. Wir setzten die maximale Ausgabelänge während der Inferenz auf Eingabelänge + 50, beendeten jedoch frühzeitig, wenn möglich [38].

Tabelle 2 fasst unsere Ergebnisse zusammen und vergleicht unsere Übersetzungsqualität und Trainingskosten mit anderen aus der Literatur bekannten Modellarchitekturen. Wir schätzen die Anzahl der Gleitkommaoperationen, die zum Trainieren eines Modells verwendet wurden, indem wir die Trainingszeit, die Anzahl der verwendeten GPUs und eine Schätzung der anhaltenden Gleitkomma-Leistungsfähigkeit jeder GPU in einfacher Genauigkeit multiplizieren $^5$.

---

## 6.2 Model Variations

---

## 6.2 Modellvariationen

---

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

---

Um die Bedeutung der verschiedenen Komponenten des Transformers zu bewerten, haben wir unser Basismodell auf unterschiedliche Weise variiert und dabei die Leistungsänderungen bei der Englisch-Deutsch-Übersetzung gemessen.

---

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

---

$^5$Wir verwendeten Werte von 2,8, 3,7, 6,0 und 9,5 TFLOPS für K80, K40, M40 bzw. P100.

---

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.

---

Tabelle 3: Variationen der Transformer-Architektur. Nicht aufgeführte Werte entsprechen denen des Basismodells. Alle Metriken beziehen sich auf den Entwicklungsdatensatz für die Englisch-Deutsch-Übersetzung, newstest2013. Die angegebenen Perplexitäten sind pro Wordpiece gemäß unserer Byte-Pair-Encoding-Methode und dürfen nicht mit pro-Wort-Perplexitäten verglichen werden.

---

| | $N$ | $d_{\text{model}}$ | $d_{\text{ff}}$ | $h$ | $d_k$ | $d_v$ | $P_{drop}$ | $\epsilon_{ls}$ | 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 |

---

| | $N$ | $d_{\text{model}}$ | $d_{\text{ff}}$ | $h$ | $d_k$ | $d_v$ | $P_{Drop}$ | $\epsilon_{ls}$ | Train-Schritte | PPL (dev) | BLEU (dev) | Parameter $\times 10^6$ |
| :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
| Basis | 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) | | Positionale Einbettung anstelle von Sinusoiden | | | | | | | | 4,92 | 25,7 | |
| Groß | 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.

---

Entwicklungssatz, newstest2013. Wir verwendeten Beam Search, wie im vorherigen Abschnitt beschrieben, jedoch ohne Checkpoint-Averaging. Diese Ergebnisse präsentieren wir in Tabelle 3.

In den Zeilen (A) von Tabelle 3 variieren wir die Anzahl der Attention-Heads sowie die Dimensionen der Attention-Key- und Value-Vektoren, wobei die Rechenmenge konstant gehalten wird, wie in Abschnitt 3.2.2 beschrieben. Während Single-Head-Attention um 0,9 BLEU schlechter abschneidet als die beste Einstellung, nimmt die Qualität auch bei zu vielen Heads ab.

In den Zeilen (B) von Tabelle 3 stellen wir fest, dass eine Verringerung der Attention-Key-Größe $d_k$ die Modellqualität beeinträchtigt. Dies deutet darauf hin, dass die Bestimmung der Kompatibilität nicht einfach ist und eine ausgefeiltere Kompatibilitätsfunktion als das Dot-Produkt vorteilhaft sein könnte. In den Zeilen (C) und (D) beobachten wir weiterhin, wie erwartet, dass größere Modelle besser abschneiden und Dropout sehr hilfreich ist, um Overfitting zu vermeiden. In Zeile (E) ersetzen wir unsere sinusoidale Positionsencodierung durch erlernte Positionsembeddings [9] und beobachten nahezu identische Ergebnisse im Vergleich zum Basismodell.

---

## 6.3 English Constituency Parsing

---

## 6.3 Englische Konstituentenanalyse

---

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

---

Um zu bewerten, ob der Transformer auf andere Aufgaben verallgemeinern kann, führten wir Experimente zur englischen Konstituentenparserung durch. Diese Aufgabe stellt spezifische Herausforderungen dar: Die Ausgabe unterliegt starken strukturellen Einschränkungen und ist deutlich länger als die Eingabe. Darüber hinaus konnten RNN-Sequence-to-Sequence-Modelle in Szenarien mit kleinen Datensätzen keine state-of-the-art Ergebnisse erzielen [37].

Wir trainierten einen 4-Schichten-Transformer mit $d_{model} = 1024$ auf dem Wall Street Journal (WSJ)-Teil des Penn Treebank [25], etwa 40.000 Trainingsätze. Wir trainierten ihn zudem in einem semi-supervised Setting, wobei wir die größeren Corpora mit hohem Konfidenzniveau und BerkleyParser mit ungefähr 17 Millionen Sätzen verwendeten [37]. Für das WSJ-only-Setting verwendeten wir ein Vokabular von 16.000 Tokens, für das semi-supervised Setting ein Vokabular von 32.000 Tokens.

Wir führten nur eine geringe Anzahl von Experimenten durch, um den Dropout (sowohl Attention- als auch Residual-Dropout, Abschnitt 5.4), die Lernraten und die Beam-Größe auf dem Entwicklungsdatensatz des Abschnitts 22 auszuwählen; alle anderen Parameter blieben unverändert gegenüber dem englisch-deutschen Basisübersetzungsmodell. Während der Inferenz

---

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

---

Tabelle 4: Der Transformer verallgemeinert gut auf die englische Konstituentenparserung (Ergebnisse stammen aus Abschnitt 23 des 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 |

---

| Parser | Training | WSJ 23 F1 |
| :---: | :---: | :---: |
| Vinyals & Kaiser et al. (2014) [37] | Nur WSJ, diskriminativ | 88,3 |
| Petrov et al. (2006) [29] | Nur WSJ, diskriminativ | 90,4 |
| Zhu et al. (2013) [40] | Nur WSJ, diskriminativ | 90,4 |
| Dyer et al. (2016) [8] | Nur WSJ, diskriminativ | 91,7 |
| Transformer (4 Schichten) | Nur WSJ, diskriminativ | 91,3 |
| Zhu et al. (2013) [40] | semi-supervised | 91,3 |
| Huang & Harper (2009) [14] | semi-supervised | 91,3 |
| McClosky et al. (2006) [26] | semi-supervised | 92,1 |
| Vinyals & Kaiser et al. (2014) [37] | semi-supervised | 92,1 |
| Transformer (4 Schichten) | semi-supervised | 92,7 |
| Luong et al. (2015) [23] | Multi-Task | 93,0 |
| Dyer et al. (2016) [8] | generativ | 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.

---

Die maximale Ausgabelänge wurde auf Eingabelänge + 300 erhöht. Wir verwendeten eine Beam-Größe von 21 und $\alpha = 0.3$ sowohl für die WSJ-nur-Einstellung als auch für das semi-supervised Setting.

Unsere Ergebnisse in Tabelle 4 zeigen, dass unser Modell trotz fehlender aufgabenspezifischer Abstimmung überraschend gut abschneidet und bessere Ergebnisse liefert als alle zuvor berichteten Modelle mit Ausnahme der Recurrent Neural Network Grammar [8].

Im Gegensatz zu RNN-Sequence-to-Sequence-Modellen [37] übertrifft der Transformer den Berkeley-Parser [29], selbst wenn nur auf dem WSJ-Trainingsdatensatz mit 40K Sätzen trainiert wird.

---

## 7 Conclusion

---

## 7 Fazit

---

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.

---

In dieser Arbeit stellen wir den Transformer vor, das erste Sequenz-zu-Sequenz-Modell, das vollständig auf Aufmerksamkeitsmechanismen (Attention) basiert und die in Encoder-Decoder-Architekturen am häufigsten verwendeten rekursiven Schichten durch multi-head Selbst-Aufmerksamkeit ersetzt.

Für Übersetzungsaufgaben lässt sich der Transformer deutlich schneller trainieren als Architekturen, die auf rekursiven oder konvolutionellen Schichten basieren. Bei beiden WMT 2014-Übungsaufgaben Englisch-Deutsch und Englisch-Französisch erreichen wir einen neuen State-of-the-Art. In der ersten Aufgabe übertrifft unser bestes Modell sogar alle zuvor berichteten Ensembles.

Wir sind begeistert von der Zukunft aufmerksamkeitbasierter Modelle und planen, sie auf andere Aufgaben anzuwenden. Wir beabsichtigen, den Transformer auf Probleme mit Eingabe- und Ausgabemodalitäten jenseits von Text zu erweitern und lokale, eingeschränkte Aufmerksamkeitsmechanismen zu untersuchen, um große Eingaben und Ausgaben wie Bilder, Audio und Video effizient zu verarbeiten. Die Generierung weniger sequenziell zu gestalten ist ein weiteres Forschungsziel unsererseits.

Der Code, den wir zum Trainieren und Evaluieren unserer Modelle verwendet haben, ist verfügbar unter https://github.com/tensorflow/tensor2tensor.

**Danksagungen** Wir danken Nal Kalchbrenner und Stephan Gouws für ihre fruchtbaren Kommentare, Korrekturen und Inspirationen.

---

## **References**

---

## **Referenzen**

---

[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 und Geoffrey E Hinton. Layer Normalization. *arXiv preprint arXiv:1607.06450*, 2016.

[2] Dzmitry Bahdanau, Kyunghyun Cho und Yoshua Bengio. Neuronale maschinelle Übersetzung durch gemeinsames Erlernen von Alignment und Übersetzung. *CoRR*, abs/1409.0473, 2014.

[3] Denny Britz, Anna Goldie, Minh-Thang Luong und Quoc V. Le. Massive Exploration neuronaler Architekturen für maschinelle Übersetzung. *CoRR*, abs/1703.03906, 2017.

[4] Jianpeng Cheng, Li Dong und Mirella Lapata. Long Short-Term Memory-Netzwerke für das maschinelle Lesen. *arXiv preprint arXiv:1601.06733*, 2016.

---

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

[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 und 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 und Yoshua Bengio. Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling. *CoRR*, abs/1412.3555, 2014.

[8] Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros und Noah A. Smith. Recurrent Neural Network Grammars. In *Proc. of NAACL*, 2016.

[9] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats und 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 und Jian Sun. Deep Residual Learning for Image Recognition. In *Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition*, Seiten 770–778, 2016.

[12] Sepp Hochreiter, Yoshua Bengio, Paolo Frasconi und Jürgen Schmidhuber. Gradient Flow in Recurrent Nets: The Difficulty of Learning Long-Term Dependencies, 2001.

[13] Sepp Hochreiter und Jürgen Schmidhuber. Long Short-Term Memory. *Neural Computation*, 9(8):1735–1780, 1997.

[14] Zhongqiang Huang und Mary Harper. Self-Training PCFG Grammars with Latent Annotations Across Languages. In *Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing*, Seiten 832–841. ACL, August 2009.

[15] Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer und Yonghui Wu. Exploring the Limits of Language Modeling. *arXiv preprint arXiv:1602.02410*, 2016.

[16] Łukasz Kaiser und Samy Bengio. Can Active Memory Replace Attention? In *Advances in Neural Information Processing Systems, (NIPS)*, 2016.

[17] Łukasz Kaiser und 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 und Koray Kavukcuoglu. Neural Machine Translation in Linear Time. *arXiv preprint arXiv:1610.10099v2*, 2017.

[19] Yoon Kim, Carl Denton, Luong Hoang und Alexander M. Rush. Structured Attention Networks. In *International Conference on Learning Representations*, 2017.

[20] Diederik Kingma und Jimmy Ba. Adam: A Method for Stochastic Optimization. In *ICLR*, 2015.

[21] Oleksii Kuchaiev und 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 und 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 und Lukasz Kaiser. Multi-Task Sequence to Sequence Learning. *arXiv preprint arXiv:1511.06114*, 2015.

[24] Minh-Thang Luong, Hieu Pham und Christopher D Manning. Effective Approaches to Attention-Based Neural Machine Translation. *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.

---

<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.

---

Abbildung 3: Ein Beispiel für den Aufmerksamkeitsmechanismus, der langreichweitigen Abhängigkeiten in der Selbst-Aufmerksamkeit des Encoders in Schicht 5 von 6 folgt. Viele der Aufmerksamkeitsköpfe beziehen sich auf eine entfernte Abhängigkeit des Verbs „making“ und vervollständigen die Phrase „making...more difficult“. Die Aufmerksamkeiten werden hier nur für das Wort „making“ dargestellt. Verschiedene Farben repräsentieren verschiedene Köpfe. Am besten in Farbe betrachtet.

---

<img src="images/14-0.png" style="zoom:70%; display: block; margin: 0 auto;" />

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.

---

Abbildung 4: Zwei Aufmerksamkeitsköpfe, ebenfalls in Schicht 5 von 6, die offenbar an der Auflösung von Anaphern beteiligt sind. Oben: Vollständige Aufmerksamkeiten für Kopf 5. Unten: Isolierte Aufmerksamkeiten nur vom Wort „its“ für die Aufmerksamkeitsköpfe 5 und 6. Beachten Sie, dass die Aufmerksamkeiten für dieses Wort sehr scharf ausgeprägt sind.

---

<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.

---

Abbildung 5: Viele der Aufmerksamkeitsköpfe zeigen Verhalten, das mit der Struktur des Satzes in Zusammenhang steht. Wir geben zwei solche Beispiele oben an, von zwei verschiedenen Köpfen aus der Encoder-Selbstaufmerksamkeit in Schicht 5 von 6. Die Köpfe haben eindeutig gelernt, unterschiedliche Aufgaben zu erfüllen.

---
