Moving from natural language processing to natural language understanding
Imagine a world where satisfying human-computer dialogues exist. With the resurgence of interest in natural language processing (NLP) and understanding (NLU) – that day may not be far off.
In order to provide more satisfying interactions with machines, researchers are designing smart systems that use artificial intelligence (AI) to develop better understanding of human requests and intent.
Last year, OpenAI used a machine learning technique called reinforcement learning to teach agents to design their own language. The AI agents were given a simple set of words and the ability to communicate with each other. They were then given a set of goals that were best achieved by cooperating (communicating) with other agents. The agents independently developed a simple ‘grounded’ language.
Human language is said to be grounded in experience. People grasp the meaning of many basic words by interaction – not by learning dictionary definitions by rote. They develop understanding in terms of sensory experience — for example, words like red, heavy, above.
Abstract word meanings are built in relation to more concretely grounded terms. Grounding allows humans to acquire and understand words and sentences in context.
The opposite of a grounded language is an inferred language. Inferred languages derive meaning from the words themselves and not what they represent. In AI trained only on textual data, but not real-world representations, these methods lack true understanding of what the words mean.
It happens. Even if the researcher gives the agents simple English words the agent inevitably diverges to its own, unintelligible language. Recently researchers at Facebook, Google and OpenAI all experienced this phenomenon!
Agents are reward driven. If there is no reward for using English (or human language) then the agents will develop a more efficient shorthand for themselves.
When researchers at the Facebook Artificial Intelligence Research lab designed chatbots to negotiate with one another using machine learning, they had to tweak one of their models because otherwise the bot-to-bot conversation “led to divergence from human language as the agents developed their own language for negotiating.” They had to use what’s called a fixed supervised model instead.
The problem, there, is transparency.
