LANGUAGE: AI’S GREATEST CHALLENGE
Richard Socher, Salesforce’s Chief Scientist, have taken some time to write about the greatest challenge in the field of Artificial Intelligence. He writes that despite natural language processing or NLP, language is still the main challenge in AI technology. That’s because despite the programming capabilities that makes human-machine communication possible, we still need to address sentiment analysis, joint multi-tasking learning, and question answering as the nuances of AI language.
First, sentiment analysis revolves around context. And while we can program machines to distinguish between positive, negative, or neutral statements, we are yet to discover a way for them to understand the emotional components of language with accuracy. This is especially if the sentiment in language is geared towards brand impression and evaluation.
Second, joint multi-tasking learning. Currently, machines are programmed to process and act on input data because those data are simple. The challenge not only lies in combining a set of simple data into order to allow machines to process complex data. This especially becomes more complex with the introduction of language qualifiers. For example, machines need to learn the differences between ‘good,’ ‘better,’ and ‘best.’
Finally, question answering still has a long way to go. We know Siri answers questions. We know Google Assistant does the same thing. But what they lack is contextual understanding so that they can respond with accuracy. For example, the word ‘data’ can refer either to mobile internet consumption or a set of information. This arbitrary nature of language needs to be built in in machines so that we get the right answers.
To date, and despite Salesforce Research has already invented the Natural Language Decathlon, a single model capitalizing on the power of question answering the 10 toughest task of NLP, Socher recognizes that NLP has more opportunities for improvement in order to advance the ideals of human-machine communication.