I’ve listed a few significant use cases of Sequence-to-Sequence modeling below (apart from Machine Translation, of course):

It’s time to get our hands dirty! There is no better feeling than learning a topic by seeing the results first-hand. We’ll fire up our favorite Python environment (Jupyter Notebook for me) and get straight down to business. Fast-forward to 2019, I am fortunate to be able to build a language translator for any possible pair of languages. What a boon Natural Language Processing has been! We can randomly print some actual vs predicted instances to see how our model performs: # print 15 rows randomly Lately there have been repeated uncertainties concerning the rules and conventions of dict.cc. Discussions in the forum often ended without results. That's why I created a new system for changing and extending the guidelines (the set of rules for vocabulary maintenance). Here is how it works: Over the years this vocabulary grew to an amazing 945,000 entries and will inevitably exceed the million mark in the foreseeable future. This was only made possible by the real heros at dict.cc, the 21,330 registered (and many more unregistered) contributors! I know many of the most committed people well enough by now to know how much energy and emotion they keep investing in dict.cc. Some of them started contributing to dict.cc at the very beginning ten years ago and are still here, contributing diligently!Viele Wörter schnell übersetzen im Textmodus You can look up words in a whole text without having to manually type or copy each individual word you don't understand in the search field: ­ Zusätzliche Informationen zum Suchergebnis werden Ihnen angezeigt, wenn Sie mit der Maus über den entsprechenden Eintrag fahren. Dort finden Sie z.B. Angaben zur Grammatik, hilfreiche Diskussionen aus dem Forum oder Beispiele. Please note that we have used ‘ sparse_categorical_crossentropy‘ as the loss function. This is because the function allows us to use the target sequence as is, instead of the one-hot encoded format. One-hot encoding the target sequences using such a huge vocabulary might consume our system’s entire memory. Genießen Sie präzise, natürlich klingende Übersetzungen, die von der PROMT Neural Machine Translation (NMT)-Technologie unterstützt werden, die bereits in I spent quite an amount of time thinking about these requests, doing some experiments and finding a system that hopefully suits all suggestions. The resulting set of features - called my.dict.cc - is now available in a beta version. "Beta" means it might still contain errors and may need some fine-tuning.

Even after 20 years, dict.cc is still a one-man business, and also my personal life project. I started dict.cc when I was 24, and now I'm 44 somehow. ;-) Almost half my life already, I wonder where the time went! And this project is not just a job, it's also meant to be my contribution to society, helping with easier access to education and helping to break down barriers between people who don't speak the same language. Still most of the technical innovations happen in the mobile world, on smartphones, tablets and other "smart" devices. New Android and iOS versions are being published constantly and new devices appear, meaning the dict.cc app has to be adapted continuously. Unfortunately the programming environment I've been using so far does not seem to keep up with all of the developments, leading to technical issues with recent Android versions. That's why I've been busy reprogramming the app from scratch in a new development environment. Most probably it will take me several more months of programming until I can start releasing this new app. Apart from links to DWDS (for German definitions) and canoonet (German inflection), the information menu also contains links to LEO's inflection tables (verb tenses, plural forms, etc.). ­ The test data could have been organized as multi-reference data sets but this would require to provide different sets in both translation directions. Removing alternative translations is also not a good option as this would take away a lot of relevant data. Hence, we decided to provide the data sets as they are, which implicitly creates multi-reference test sets but with the wrong normalization. License New releases are planned in the future and will be announced here. Development and test data will be updated regularly but the original test sets will stay in the release. Updates of the test data will be available through this devtest release and will not include any examples available in development data. Those data sets are also available from this git repository in the sub directory data/devtest/. The translation challenge

Implementation in Python using Keras

print('English Vocabulary Size: %d' % eng_vocab_size) English Vocabulary Size: 6453 # prepare Deutch tokenizer I spent the last days mainly adapting guidelines, fixing bugs and adding missing features, tags and the links to other online resources. The guidelines are now more or less stable. I'd ask all participants to have a look at the updated guidelines, as a few things have changed. Statistics: Currently there are 20 new languages being built up, each connected to both German and English. Including German-English this makes 41 language pairs, with most of them still in their early stages. At the moment there are 38,000 entries, all new language pairs combined. More languages can be added as soon as the system has stabilized. In principle, each category lists direct matches for the search term(s) first (if available, in this case 'das Suchen'), regardless of whether the word you are looking up is interpreted as an inflected word (the noun 'Suchen' is also the plural of 'die Suche') or an infinitive (verb 'suchen'). All other possible word forms appear further down (in case of the noun: results for 'die Suche', the singular of 'die Suchen'). Note that we will prepare tokenizers for both the German and English sentences: # function to build a tokenizer

Next, vectorize our text data by using Keras’s Tokenizer() class. It will turn our sentences into sequences of integers. We can then pad those sequences with zeros to make all the sequences of the same length. Don't use any dev or test data for training (dev can be used for validation during training as an early stopping criterion). Some other statistics: 799,194 German-English translation pairs and 391,932 translation pairs in the new languages sum up to a total of over a million translations! Additionally there are 333,873 voice recordings and 286,232 inflections. 13,867 of 94,865 registered users have already contributed to the vocabularies, voice recordings or in the translation forum. Entering a keyword followed by the [Tab] key (once or several times) uses your preferred search engine.The test and development data sets are released with the CC-BY 2.0 FR license (see also the terms of use of the original data at https://tatoeba.org/en). New: "UPDATED!" is displayed if the original posting was changed within the current verfication round. A list of language pairs with training data but no test data from Tatoeba (paired with English only)