Future application: translator and dedicated student tools

Outline of the tool

The tool will segment the text the user is interested in and display each segment separately like so (see underneath the figure for explanation):

The work proceeds segment by segment:

• ✅: Completed segments influence the results of the segments following them due to the contextual nature of the model and the online generation of the text. This is a critical feature of such a tool. In general, it is very difficult to specify linguistic rules that formalize English sentences such as translate the interlocutor as male or the following sentences have the same referrent. However, the model involved here is not syntactic, it is semantic. It should "understand" the corrections made by the translator and change the outputs of subsequent translations. Of course, it will not do so perfectly and some additional machine learning work may be required to get this to work well.
• ✏️: The section currently being worked on should be word segmented. The user can tap/click on a word and see its part-of-speech and a short, preferably single word, translation. This requires a new dictionary. We're hoping to use unsupervised machine translation to discover this dictionary, with human expert corrections incorporated later.
• 🤖💬: Future sections show candidate machine translations. As the user makes corrections to the translation, the translations of subsequent sections update automatically. The level of ambiguity of the translation should be color coded.

The tool will be configurable, for example the user should be able to configure the section segmentation strategy. It is not intended to be a simple, one-click translator like what we have in modern web browsers.

Research needed

We first need to complete several research steps before we can develop this app.

• Similar requirements to the language teaching tools:

  • A model for Tibetan word segmentation, part-of-speech tagging and named entity recognition. This is likely a single model that combines all three tasks and uses morpheme-level or even character-level tokenization. Developing this model is part of our current roadmap for this year.

  • A simplified contextual dictionary that has only a short list of possible English words for a given Tibetan word (not long sentences) and that takes the part-of-speech tag into account. We would like to first use unsupervised machine translation to discover a first draft of this dictionary automatically, and then partner with linguists to complete it.

• Translation models:

  • A stylistic translation model for use by translators such as 84,000.

  • A literal translation model for use by students.

Delivery methods

This would be a desktop app. It would be best to use a universal framework so that we can easily build one app for both Windows and Mac, as well as in-browser.