Tencent and Alibaba’s AI Models Speak Chinese Better than Your Teacher

AI models speak Chinese better with improved NLP technology. Now no more faux pas with the Chinese

雷雨天气 请勿拨打手机 Thunderstorms do not call cell phones', reads a Chinese signboard. Funny but in case you do not get the context, definitely it will cost your life. Losing in translation happens with every language and most of the time it is not a pleasant experience. It will not anymore, at least with Chinese, because from now onwards, the Chinese language can be translated better as researchers have made AI speak Chinese better. An Artificial intelligence model, released by Tencent Holdings and Alibaba Group Holding, as per a benchmark test measuring natural language processing, can understand Chinese better than a human being. The Chinese Language Understanding Evaluation (CLUE) benchmark, is set based on tasks that calibrate a machine's ability to understand and respond to Chinese text in the same way as humans do. Tencent's and Alibaba's AI models scored high in the evaluation. It is the first time that an AI model has defeated humans ever since researchers succeeded in creating this benchmark three years ago.

CLUE's website, according to Monday's ranking, ranked Tencent's "Huanyuan AI model" first with a score of 86.685, followed by Alibaba's AI model, Alismind, ranked 2^nd with a score of 86.685, both performing better than the humans who scored 86.678, while Oppo, the Chinese smartphone maker, and the food delivery giant Meituan's models were ranked 4^th and 5^th respectively. Alibaba in its recent blog said, "the top scorers have never surpassed humans so far despite the fact that the rankings change frequently".

By and large English translated Chinese is a kind of continuous bag of words that requires only an MLM (Masked Language Model), a common pre-training model for machine translation. It essentially recovers a few input tokens in the vocabulary space by replacing them with masking tokens, [MASK]. Though the output has a simple formulation, it can fairly represent the contextual information around the masked token. The meaning of the sentence can be understood even though certain words in the statement are disorganized. The current NLP models are the refined version giving the model the ability to identify the context. The CLUE benchmark against which the models are being measured is an open-ended, community-driven project that brings together 9 tasks spanning several well-established single-sentence/sentence-pair classification tasks, as well as machine reading comprehension, all on original Chinese text. It reports scores using an exhaustive set of current state-of-the-art pre-trained Chinese models, 9 in total.

Chinese AI companies have come a long way, so much so that they are about to set themselves as global leaders, and NLP technology is not an exception. Be it Alibaba's virtual voice assistants and spam detectors, or Baidu's content-generating bot Du Xiaoxiao, the industry is well ahead of its peers. Nevertheless, in many cases, as reported by Alabama University, AIs that performed better than humans in reading and comprehension lagged in detecting instances of shuffled words. Alibaba! Kindly take note