Why is it so hard to wreck a nice beach?

Unfortunately, my Finnish is a bit rusty, so I did what I always do in my times of need: I turned to Google. Using the Google machine translator, I typed in what I had intended to be the first sentence of this week’s column: Machines are not better at languages than people. I translated it into German. Then I translated the German sentence into Japanese and the Japanese into Norwegian. After only five seconds, I had written an entire four sentences. It was incredible; I had finally achieved the ultimate goal of AI journalism: My computer was actually writing the column for me.

But then I made the mistake of translating my Norwegian output back into English: The car better than the people in the language. Was my computer trying to tell me something?

Relax, this isn’t one of those AI columns preaching the superiority of robots over humans. The next 1,000 words are not devoted to convincing you that your Honda would get a better grade than you in SPA 101. Instead, I want to look at the capabilities of AI technology in processing language, a realm where people, actually, still significantly outperform computers.

The vast majority of humans know at least one language, if not more. We can speak, read, write and understand it, we can tell jokes and decipher foreign accents and bad handwriting, and we can recognize and adopt different tones of voice, from friendly to hostile, mocking or sarcastic. Our grasp of language far surpasses anything computers have achieved because words trigger all sorts of associated images and experiences in our heads. We use language to produce mental scenarios far beyond the surface meaning of the words.

This is tricky to teach computers, who still aren’t that good at understanding human speech. When I call an airline to check a flight’s arrival time, and the automated operator asks me for the flight number, I find that if I speak very slowly and enunciate as carefully as possible, then the results are usually fairly good. If I speak too quickly or mumble even a little, however, I get the departure time for the flight to Houston rather than the arrival time for the flight from Bombay and end up in a furious — not to mention futile — fight with the maddeningly calm robot voice from hell.

To be fair, these speech-recognition technologies are slowly improving, and many cell phones now feature fairly reliable voice dialing tools. And let’s give the machines a little credit. After all, they’re doing the work to learn our native language, while very few of us are up to the task of communicating with them in 0s and 1s.

But if we can’t even teach computers how to recognize speech — especially since they can’t really distinguish these lessons from ones on how to wreck a nice beach — how can we possibly hope to impart on them the richness and nuance of words and language?

The crucial tool for attacking this challenge may be the lexical database WordNet, first developed at the University’s Cognitive Science Laboratory in 1985 by psychology professor emeritus George Miller. WordNet, which contains definitions for some 150,000 nouns, verbs, adjectives and adverbs, stores thousands of semantic relations between different words: The word “Bloodhound” refers to a type of the word “dog,” the words “student” and “pupil” are synonyms, and the word “Bombay” refers to a part of the word “India,” which refers to a part of the word “Asia.”

Miller initially designed this web of semantic relationships to study how humans understand language.

“One of the questions that really motivated [Miller] to start WordNet was, ‘How do people organize the tens of thousands of concepts, and the words that express these concepts, in their mind?’ ” said Christiane Fellbaum GS ’80, a linguistics lecturer and senior research scholar in the computer science department who currently maintains WordNet.

“[Miller] was working from the idea that what we know about a concept is based on its relations to other concepts,” she added. “So, in WordNet, words are organized in a hierarchical fashion, based on their relations to other words, with more specific and less specific concepts arranged in a sort of tree.”

One such tree might start at the root “animals” and have a “mammal” branch that, in turn, would have a “monotreme” branch, leading to a “duck-billed platypus” branch, and so on. As you go further down the tree, reaching more and more specific concepts, you can simply add on more specific details — like that platypuses live in Australia — to the more general body of knowledge associated with higher branches — that mammals are vertebrate animals with sweat glands, for instance.

Get the best of the ‘Prince’ delivered straight to your inbox. Subscribe now »

WordNet has since branched out to become an invaluable tool in the AI subfield of natural language processing, which focuses on teaching computers to understand and generate human languages.

The rigorous mapping of semantic relationships between words like “foot” and “leg” means that computers can actually disambiguate the meaning of words with multiple possible definitions based on their context. So, if a computer reads the sentence “his foot hurt, but his leg felt fine,” it may, at first, be uncertain whether the word “foot” is referring to a length of 12 inches or the body part below the ankle. If the computer uses WordNet, however, it can determine that the nearby word “leg” is closely semantically related to the latter meaning, thereby concluding that the sentence refers to the body part, not the unit of measure.

This process is closely related to the ways we use context to understand words with multiple meanings, but our disambiguation abilities are so advanced and so unbelievably fast that we have yet to fully understand or appreciate their complexity. Even if computers were capable of disambiguating and analyzing words at the sophisticated level of humans, though, there would still be a wide gap between their understanding of language and ours.

“What does it mean to understand something? There’s a slightly superficial but sort of helpful distinction between word knowledge versus world knowledge,” Fellbaum said. “Word knowledge includes things like definition, part of speech and semantic relations, but world knowledge is what we know from our experience, our understanding of how and why things happen. At the moment, AI applications are pretty much confined to word knowledge.”

WordNet has helped researchers so tremendously in developing this kind of word knowledge that people all over the world have begun building similar databases in other languages. Fellbaum estimates there are between 50 and 60 such projects currently underway, united through the Global WordNet Association she founded with Dutch scholar Piek Vossen in 2000.

The development of WordNets in so many different languages opens up the possibility of using them to build more powerful, accurate automated translation tools, Fellbaum said.

So perhaps one day I will be able to use a tool like the Google machine translator to provide for all my readers across the globe who leap out of bed each Friday morning anxious to read the newest technocrat, technokrat or teknokraatti. For now, though, I’m sticking to English because no, the car not better than the people in the language. To my loyal Finnish fan club: Olen pahoillani.

This is the fifth in a series of articles examining current and emerging artificial intelligence technologies and their impact on today’s world.