I had my numbers wrong. The percentage of affected USER_CONTEXT_MENU_TRANSLATE is actually
529/3707 = ~14%

Another thing that is peculiar with these protos is that the RankerResponse is SHOW while the ranker version is 0, but I guess this is explained by the fact that SHOW is the default value of the enum. In hindsight, NOT_QUERIED would have been a better default value, but it is too late to change now. Let's take this into consideration for the next proto design.

So I suppose the only fields that are explicitly defined in these cases are event_type and event_timestamp_sec.

I recent count shows that only about 0.2% of data as an empty source language string for "explicit events" (events used to train the model).

However, this seems to happen more often with USER_ACCEPT (0.75%) than the other negative outcomes. This seems odd to me.

https://plx.corp.google.com/script/#a=qo%7Ci=google%253A%253Ascript_ae._29b300_09ae_464a_809a_a868cd8952b3

It would be good to understand what these cases are. The fact that the UI is shown to the user and that these events are not "UNSUPPORTED_LANGUAGE" probably points to something wrong in how we get the source_language and target_language for inference.

Looking at numbers where ranker is not enabled, empty source language happens 0.1% of the time for explicit events, but 1% when the event type is USER_ACCEPT. So, it looks like there is a correlation between both facts, but it is hard to say which one causes which.

I am adding this bug to issues with TEPs described in b/72396986

After investigation, we found that most of these events are TEPs that have not been initialized. So, it means that there is a code path that can show translate UI without going through TranslateManager.InitiateTranslation.

The simple solution to this is to verify that the TEP has been initialized when we call RecordTranslateEvent. We can also filter these events out on the server-side when computing metrics and building training datasets.


There is still a small number of events (0.06%) that have an empty source language, but where target_language is not empty. In this case, the TEP has been properly initialized, but for some reason, either TranslateDownloadManager::IsSupportedLanguage('') returns True, or some other code path eventually shows the UI. These events are very rare, so we can probably just ignore them.

The following revision refers to this bug:
  https://chromium.googlesource.com/chromium/src.git/+/d59d15f88feaaaa616bbd049f105dc4ceaa46a75

commit d59d15f88feaaaa616bbd049f105dc4ceaa46a75
Author: Mathieu Perreault <mathp@chromium.org>
Date: Mon Jan 29 21:18:02 2018

[Translate] Provide a better default value for proto field

Bug: 678689
Change-Id: I9f47f8f59f86621327d5ace0b2431b1fc48ff900
Reviewed-on: https://chromium-review.googlesource.com/889823
Reviewed-by: Robert Kaplow <rkaplow@chromium.org>
Commit-Queue: Mathieu Perreault <mathp@chromium.org>
Cr-Commit-Position: refs/heads/master@{#532591}
[modify] https://crrev.com/d59d15f88feaaaa616bbd049f105dc4ceaa46a75/third_party/metrics_proto/translate_event.proto