Chinese Internet search giant Baidu has achieved a breakthrough in real-time machine translation system with the ability to translate 2 languages at once. Termed ‘simultaneous translation with anticipation & controllable latency’ (STACL), the system has been endowed with anticipation capabilities & controllable latency, the tech giant announced in a recent blog post.
Calling it “a major breakthrough in natural language processing” (NLP), Baidu said the artificial intelligence (AI) powered tool can identify the difference in the word order between the source & target languages. The system has been devised with quick response time when applied to real-world applications for real-time machine translation or interpretation.
As opposed to consecutive interpretation where the translator pauses to start translation, simultaneous interpretation is the need of the hour, as here, the translation begins as soon as the speaker starts the speech & ends as soon as the speaker ends it, thus saving quite an amount of time owing to speed & accuracy. However, with growing demand & scarce availability of simultaneous interpreters, the Beijing company stressed on the challenge to create automated systems to expand the access to simultaneous & reliable translation.
Inspired by human simultaneous interpreters, Baidu took on the challenge to create such a real time machine translation system. However, unlike human translators who anticipate or predict words that a speaker may be about to communicate in a few seconds into the future, the automated model directly predicts the target language words in the translation.
Essentially, it seamlessly fuses translation & prediction in a single “wait-k” model.
Here, “translation is always ‘k’ words behind the speaker’s speech to allow some context for prediction.” The AI powered model has been trained to use the available prefix of the source sentence at each step (along with the translation so far) to decide the next word in translation.
The idea has been explained in this way in the blogpost:
- In Chinese, Bùshí Zǒngtǒng zài Mòsīkē yǔ Pǔjīng huìwù, means “President Bush meets with Putin in Moscow”.
- Here, the verb huìwù (meet) is positioned at the end which is akin to a German or Japanese verb. In contrast, in English, the verb “meets” appears before.
- This discrepancy in word order in languages proves to be a hindrance to both human simultaneous interpreters & machine translation systems.
- Therefore most “real-time” translation systems still use the conventional non-simultaneous full-sentence translation techniques. This not only is time-consuming but also puts the user out of sync with the speaker.
- Coming back to the aforementioned sentence, Baidu’s AI translation system, it claims, accurately predicts that the next word would be “meet” because Bush is likely “meeting” someone in Moscow.
The Baidu AI model is being trained from vast amount of training data having related sentence structures so as to predict with an intelligent accuracy.
Nevertheless, the tech company clarified that it had no intention of replacing human interpreters whose services will always be sought for many years to come. This was just an attempt to make simultaneous translation more accessible, it stated.
Image Credit: Baidu
