Anyone who has ever tried to have a conversation with the virtual assistant from Apple, Google or Amazon knows: this is not possible. The system still recognizes simple requests, but as soon as you deviate from the script it is trained for, you get stuck in a network of error communication. The same goes for conversations with chatbots on websites, and automatic translation services should not be blindly trusted. The computer obviously does not understand you properly.
Andrea Martin: “I do not think machines will ever understand us as we expect humans to.”
And this is not surprising, says Andrea Martin, cognitive researcher and leader of a research group at the Max Planck Institute in Nijmegen. “For these computer systems, meaning or interpretation does not matter at all,” she explains. “They do not understand language, they just always predict the next word. They do this on the basis of the training data: large amounts of text from, for example, Wikipedia or Reddit. And then they look not only at the words before the word to be predicted, but also at those after that, that is, in the future. It’s not like how the human brain understands language. “
What makes human language so special that it is difficult to translate into an understanding computer? “The human brain has a paradox when it comes to language,” Martin says. “We build our language skills by discovering statistical regularities in the language we hear around us, but our everyday language also shows that we are constantly breaking away from statistical knowledge.” Every day we produce and understand utterances – words, gestures, phrases – that we have never heard before or have only used in other contexts.
Selected by the editorial staff
Winning letter from an old oak tree
The war as a threat to our daily bread
Self-sacrifice as proof of love for the motherland
For example, look at relatively new words, such as “enemy” or “rejection officer”. The first time you heard or read these words, you could immediately deduce the meaning. Martin: “Such words do not come randomly, they follow the rules of language. We know these rules and can therefore use them to derive the meaning. ” At the exit -and of ‘unven’ you see that it is a verb, and the prefix develop indicates a rejection. How to get to the meaning of ‘breaking a friendship’.
This wonderful combination of pattern recognition and creativity is also the root of how young children learn languages. “Children distill the underlying rules based on the linguistic input they receive. They then produce new utterances based on it. If necessary, they can then adjust the derived rule. “For example, a small child can form a participle ‘desired’ himself; in line with ‘caught’, ‘lubricated’ and ‘cycled’, as he has heard before. After a while it will feel that it must, after all, be ‘desired’ and adapt its personal mental language model accordingly.
Computers cannot handle languages as creatively and flexibly. Martin: “A computer system is limited to providing the most likely response to a given input. So this is the answer that most often appeared in the training data. They will never say or understand anything they have not heard before, even if it differs slightly from the training data. “
Realistic language model
Martin and her research team are working on a theoretical model that treats language as the human brain does. How do you translate that paradox into such a language model? “It’s important that the theory we develop fits into what we know about language processing from different areas, such as linguistics, cognitive neuroscience and psychology,” says Martin. “It’s a lot: from information about how language works, to how we can understand and pronounce a sentence in its entirety. But also, for example, how brain cells become active and then recharge. That knowledge has never before been gathered in one model.”
Martin’s realistic language model must therefore in the first instance be able to imitate known effects from the scientific literature. Martin: “Words that are more predictable should therefore also be recognized more quickly. But at the same time, the model must also be able to handle unpredictable words or new sentences. If it can not do that, then something is not right somewhere and we have to go back to the drawing board. “
An MEG scanner measures the magnetic fields that occur naturally through brain activity.
In addition, the model must also be able to largely explain the results of several new experiments with a MEG scanner. This scanner detects the magnetic fields created in the brain by brain activity down to milliseconds. Martin: “We now have a large dataset of MEG recordings of adults listening to an audiobook of fairy tales by the Brothers Grimm. Fortunately, because of the corona, it is now difficult to get test subjects. ”
For example, based on the MEG data, researchers can determine how predictable a syllable or word is based on the preceding words. Martin’s language model should lead to the same conclusion. If not, it needs to be adjusted. “We also look at how different grammatical structures are related to neural responses in the brain,” says Martin. “We hope to take the first steps in understanding how the brain holds both statistical and rule-based information.”
Martin already released the first version of the theoretical model in 2016. And the second, updated version came out in 2020. “It also includes neuroscientific knowledge,” she explains. “As it is in science, this model may never be complete. Every few years we can provide an update if more information is known or if we see that the principles in theory do not agree with the experimental data. The model is my life’s work . ”
Martin does not expect her model to be used directly in technical applications. ‘Too much is still unknown to it. But cognitive and neuroscientific knowledge is necessary if we are to make the interaction with a computer more humane. Understanding the principles of language processing will help you come up with better solutions. Or as the German physicist Max Planck said: “Insight must precede application.”
Will there ever be computers that really understand us? “It depends on what you mean by understanding. If you mean by following instructions given by words or gestures, they already do so to some degree. But if you want more than to translate an utterance into a command or a predetermined answer, then I think it will be difficult.Although machines can ever really interpret a sentence, they may not know which concepts the words refer to, how these concepts are connected in a large network with other knowledge, and how the recipient perceives the meaning. It could be more human than it is now, but I do not think machines will ever understand us as we expect humans to. “