Dear Gerard, Scott, Lucie and Amir & everyone,

Thank you for the helpful responses! 

Gerard - great to hear about your work and thank you for the reference on Cebuano Wikipedia. We weren't familiar with that case, but had similar fears. We are putting our machine translated content on a separate site (Wikibabel) rather than Wikipedia, and we never expect it to show up nearly as high in search engine results, and so the native content will always take precedence. 

Scott - for the time being, the Wikibabel experiment will not have a Wikidata or Lexicographical portion, as for now this is an experiment to see if machine translated content in certain categories is useful given the state of machine translation currently (and we are doing this on a separate site in order to avoid any contamination). If we'll measure good user engagement and survey results, we'd love to think about how to integrate this better. We'd love to eventually have an editing interface on top of the translation and with existing native content if exists and has fewer details. If this editing interface becomes popular, then we'll start accumulating a dataset which may be useful to machine translation services. We, however, are not developing anything novel in the machine translation algorithm space. 

Lucie - ah, that's what this is! We noticed the large number of recent one sentence articles and were wondering what project that was. Those are awesome! Both in terms of information availability and because it allows us to measure relative interest within the generated set. I would love to discuss further your plans to expand beyond the introductory sentence, and if we can be helpful in any way. Thank you for the publication links as well.

Amir - thank you for the pointers to these projects. Your 2 points of feedback, if I understand correctly:
1. Machine translation might not be good enough to yield useful information. 
2. People can translate the pages for free.
Those are excellent points that we've thought about deeply before starting the projects (though we find a much higher translation quality recently than you perhaps). Here's what we think:
1. This is exactly the central question of our experiment, and very much still open. Machine translation (or at least Google translate) has improved significantly in the last 12 months or so. The quality of the translation, particularly when there is context (longer sentences) has improved leaps and bounds. For the Wikibabel project, we spot checked with Swahili speakers that some pages translate very well (not perfect human level, but very understandable with a few awkward turns of phrase) and some are bad enough to not be useful. Given how little information there is on the Internet in Swahili, particularly on technical topics (easier to translate in some ways) and that there aren't many participants in the Swahili Wikipedia, we hypothesize that the best X% of translations would be useful, and that we can measure and tell the difference between well translated pages and not from page analytics and surveys. 
We are, however, careful to have this in a separate site (Wikibabel), rather than checking any of this into wikipedia, because as you mentioned that would be misleading.
2. That is absolutely true, and we're fundamentally solving a discoverability problem. If a Swahili speaker currently Searches for a term in Swahili, they will not land on the English results for it. They will get some potentially bad results in their language and potentially give up. In general, they would need to know about Wikipedia, or some other good source, that it's better in English and that Google translate exists. Some of the folks we're targeting are fairly new to the internet, so this is not a low bar.
Moreover, if this is successful we're very interested to being added to free mobile data services such as Free Basics. Those services don't work well with Google Translate as it requires heavy javascript that isn't included in the free data package yet, and some phones they surveyed as common have trouble handling javascript well.

Thank you!
Olya & the Wikibabel crew

On Mon, Jun 18, 2018 at 12:43 PM, Info WorldUniversity <> wrote:
Hi Olya, Lucie, and Wikidatans, 

Very interesting projects. And thanks for publishing, Lucie - very helpful!

With regard to Swahili, Arabic (both African languages!) and Esperanto, and leveraging Google Translate / GNMT, I've been looking at this Google GNMT gif  image - and wondering how the triplets of the Linked Open Data of Wikidata structured Knowledge Base (KB) would stream through this in multiple smaller languages? 

I couldn't deduce from this paper - - here, for example ... 

2.1 Encoding the Triples The encoder part of the model is a feed-forward architecture that encodes the set of input triples into a fixed dimensionality vector, which is subsequently used to initialise the decoder. Given a set of un-ordered triples FE = {f1, f2, . . . , fR : fj = (sj , pj , oj )}, where sj , pj and oj are the onehot vector representations of the respective subject, property and object of the j-th triple, we compute an embedding hfj for the j-th triple by forward propagating as follows: hfj = q(Wh[Winsj ;Winpj ;Winoj ]) , (1) hFE = WF[hf1 ; . . . ; hfR−1 ; hfR ] , (2) where hfj is the embedding vector of each triple fj , hFE is a fixed-length vector representation for all the input triples FE. q is a non-linear activation function, [. . . ; . . .] represents vector concatenation. Win,Wh,WF are trainable weight matrices. Unlike (Chisholm et al., 2017), our encoder is agnostic with respect to the order of input triples. As a result, the order of a particular triple fj in the triples set does not change its significance towards the computation of the vector representation of the whole triples set, hFE .

... whether this would address streaming triplets through GNMT?

Would this? And since Swahili, Arabic and Esperanto, are all active languages in - - no further coding on the GNMT side would be necessary. (I'm curious how best for WUaS to grow small languages not yet in either Wikipedia/Wikidata's 287-301 languages or in GNMT's ~100+ languages?). 
How could your Wikidata / Wikibabel work interface with Google GNMT more fully with time, building on your great Wikidata coding/papers?


On Mon, Jun 18, 2018 at 5:17 AM, Gerard Meijssen <> wrote:
On average there is little or no support for subjects that have to do with Africa. When I check the articles for politicians for instance, I find that even current presidents let alone ministers are missing in African Wikipedias. So it is wonderful that there have been projects that deal with gaps but what if there is hardly anything?

What this approach brings us is at least information. Basic information in lists, info boxes maybe an additional line of text.

What we apparently have not done is learn from the Cebuano experience. The biggest issue was not the quality of the new information, it is the integration with Wikidata. Everything is new and it did not link with what we already knew. What we bring in this way is integrated information and as long as data is not saved as an article, the quality provided improves as  Wikidata gains better intel.

If anything, the experience of the Welsh Wikipedia brings us more than gapfinder or tiger editathon because of this is more in line with this approach.

On 18 June 2018 at 13:19, Amir E. Aharoni <> wrote:

2018-06-18 2:12 GMT+03:00 Olya Irzak <>:
Dear Wikidata community,

We're working on a project called Wikibabel to machine-translate parts of Wikipedia into underserved languages, starting with Swahili.

In hopes that some of our ideas can be helpful to machine translation projects, we wrote a blogpost about how we prioritized which pages to translate, and what categories need a human in the loop:

Rumor has it that the Wikidata community has thought deeply about information access. We'd love your feedback on our work. Please let us know about past / ongoing machine translation related projects so we can learn from & collaborate with them.

I'm not sure how has the Wikidata community think deeply about it.

One project that does something related to what you're doing is GapFinder ( ). As far as I know, the GapFinder frontend is not developed actively, but the recommendation API behind it is being actively maintained and developed, but you should ask the Research team for more info (see ).

Project Tiger is also doing something similar:

As a general comment, displaying machine-translated text in a way that appears that is had been written by humans is misleading and damaging. I don't know any Swahili, but in languages that I can read (Russian, Hebrew, Catalan, Spanish, French, German), the quality of machine translation is at its best good as an aid during writing a translation by a human, and it's never good for actually reading. I also don't understand why do you invest credits into pre-machine-translating articles that people can machine-translate for free, but maybe I'm missing something about how your project works.

Wikidata mailing list

Wikidata mailing list


- Scott MacLeod - Founder & President  
- World University and School

- CC World University and School - like CC Wikipedia with best STEM-centric CC OpenCourseWare - incorporated as a nonprofit university and school in California, and is a U.S. 501 (c) (3) tax-exempt educational organization. 

Wikidata mailing list