Johannes Leveling, Drooms’ Team Lead for Machine Learning, shares his insights on AI and machine learning within Drooms’ translation tool.
1. How would you best describe the Drooms translation tool?
Drooms offers document translation within the user interface to enable the “gisting” of information. Gisting allows the user to be able to understand the broader topic of a document. Currently, translation from German, Spanish, French, Italian, and Dutch into English (and vice versa) is supported. As Drooms expands into other markets and countries, more languages for machine translation will be added.
2. What translation software is on the market and what makes Drooms’ translation feature different?
Most people are familiar with web services and applications such as Google Translate, Bing Microsoft Translator, or DeepL to translate plain text. Even for mobile applications there are tools that translate text into a picture in real time. Our translation solution is based on Deep Learning, a branch within machine learning that employs neural networks. When Deep Learning was first researched for machine translation about ten years ago, it immediately showed a huge performance boost in terms of translation quality. Our setup for machine translation is actually derived from the setup that Google published for their participation in a machine translation evaluation benchmark and shows similar or better performance on test data sets.
3. What are the benefits of Drooms’ translation feature?
Translating a document in a Drooms data room means that the document never leaves the premises, i.e. the document is translated “in house” and stays within the secure data room. Any document that can be uploaded and converted in our processing pipeline (e.g. PDF, DOC, PNG etc.) can be translated, as long as it contains meaningful text.
4. Is the tool easy to use? What steps are involved?
Translating documents within the Drooms platform couldn’t be easier. Within the Drooms data room, click on the document to open the document view. Click on the earth icon above the document view. Then, select your target language and click ‘Translate’. Done!
5. Can you tell us more how AI/machine learning software works in this instance?
In our setup, the neural network for machine translation comprises layers of neurons, which encode the input text into an internal representation (basically activations of neurons) and decode the internal representation into words in the target language. To train the neural network, we use millions of parallel sentences (sentences in source and target language). Each sentence in the source language becomes an input for the neural network and the output of the neural network is compared with the reference sentence in the target language. Looking at the differences between the actual and the expected output, the connections between neurons are slightly adjusted to improve the translation. In general, compared to traditional purely statistical approaches for machine translation, deep learning approaches can better handle linguistic phenomena such as long-range syntactic dependencies, synonyms, or compounds, which results in a more fluent translation.
6. Is there anything else you would like to share?
Of course, I think it is worth noting that we do not provide a legally binding translation with a fully automatic machine translation. This would require a translation professional as well as legal experts.
7. What else is in the pipeline?
We are currently working on an overlay for Machine Translation, i.e. we will use the same background and layout as in the original document page and show the translation instead of the original text.