Why does neural machine translation sometimes generate translations with high BLEU scores but lack coherence?

💡 Explanation

The attention mechanism focuses on individual word pairs, maximizing short-term dependencies, but fails to capture long-range contextual information which leads to incoherence. Therefore, high BLEU scores can be achieved due to accurate local translations, rather than reflecting overall text quality as affected by statistical model overfitting.

• If a cochlear implant encodes speech information using a limited number of electrodes and current levels, which consequence follows for the recipient's perception of vowels?
• Why does agglutinative morphology in Turkish facilitate efficient information transfer compared to isolating morphology in Mandarin Chinese?
• Why does the real-time translation system of a telepresence robot often misinterpret the phrase 'put that there' when used by a remote operator?
• Why does a chatbot utilizing a Markov chain exhibit incoherence more readily than one using a transformer network?
• Why does an inexperienced translator using statistical machine translation sometimes produce fluent but culturally inappropriate translations, even with a large corpus?
• An engineer adjusts character spacing within a road sign’s text under varied luminance conditions. Which failure mode becomes likely when kerning is inconsistently applied across different letter pairings?