Same words repeated several times in translation

(Starkadur) #1

I have trained OpenNMT to translate text from English to Icelandic. Once in a while I get strange translations where the same English word is repeated. Here is an example:

src : “When, therefore, the signal for the polonaise resounded through the saloons, I went home”
tgt: “Þegar, signal signal signal signal i i i i i i”

If I skip the three las words (“I went home”), the translation will be:
“Þegar, gefur merki um polonaise …”

This is far from being a perfect translation, but at least somewhere close (but I wonder why I get three dots?)

Does anyone know what could possible be the reason for this behavior?

This is the content of rest_config.yml:

  • model: ‘/NMTModels/en-fr/model-enfr_epoch600_3.03.t7’
    replace_unk: true
    mode: aggressive
    joiner_annotate: true
    case_feature: true
    segment_case: true
    beam_size: 5