# Motivation and Question
-## Motivation
+## State of the art
- Domain control in Neural Machine Translation
- neural networks getting pretty good in pure translation
- often extra information is needed (e.g. domain, tone)
- Boxer zvedá peníze od banky a vstupuje do ringu.
- Boxer bere peníze z banky a kupuje prsten.
+## How are domains encoded?
+- Context
+- Sentence Structure
+- Vocabulary
+
+## How does Prefix Constraints work?
+- Identify a domain by a token
+- Learn overlap, allow generalization
+- Understand the concept "Context"
## Language Relatedness
Languages share:
## Research Questions
1) How does prefix constraints impact the translation performance?
-2) Has the relatedness of the language pairs an effect on the performance?
+2) Has the relatedness of the language pairs an effect on the performance of prefix constraints?
3) How important is the relatedness between language pairs in multi domain translation?
# Method
## Corpus Data
-- distinct domains as used by \cite{kobus:16}
+- Distinct Domains as used by \cite{kobus:16}
- Medical
- Law
- Finance
-- language pairs with different relatedness
+- Language Pairs with different Relatedness
- English-German
- English-Czech
-- 2 batches:
+- Two Batches:
- without modifications
- with prefix constraints
# Results
-## Dirty Training
+## Unsuccesful Training
\begin{figure}[h!]
\includegraphics[scale=0.5]{img/train_bad.png}
\caption{Example of a training curve of an unsuccsessful training}
\label{fig:train_good}
\end{figure}
-## Trainings Overview
+## Training Overview
\begin{figure}[h!]
\includegraphics[scale=0.5]{img/optim_top_5.png}
\caption{BLEU score and the validation accuracy for the best five models}
## Domain Specialization
- "Prefix Constraints" decreased score on all sets
- Mixed data set without much improvement
-- Big loose in EMEA
+- Big loss in EMEA
## Cross Language Pair Comparison
\begin{figure}
# Discussion
-## How are domains encoded?
-- Context
-- Sentence Structure
-- Vocabulary
-
-## How does Prefix Constraints work?
-- Identify a domain by a token
-- Learn overlap, allow generalization
-- Understand the concept "Context"
-
## Findings
- domain distinction differs between languages
- domain distinction differences can not be predicted (yet)
\begin{center}
Thank you for your attention.\\[1ex]
- Any question?\\[5ex]
+ Any questions?\\[5ex]
\end{center}
\footnotesize