Adaptive Disambiguation
By Justin Hall, Mon Dec 22 08:30:00 GMT 2003

The struggle to predict the next word in mobile text messages foreshadows a larger struggle to shape communications and rewire our minds.

Compose a message in your mind and pick up a phone to encode your thoughts to text. Chances are, the phone will work to complete your thoughts. Most of the time, the phone can be eerily accurate. But shortly there should be a computer stutter fueling frustration. A peek into the future of text prediction immediately conjures ever-closer computer-human cooperation, collaborative thinking between people and their devices. The aim is always to reduce that stutter and make fast texting seamless.

The most rudimentary mobile phone text-entry systems require "multi-tap" – writing "HELLO" involves tapping 4-4-3-3-5-5-5-right-5-5-5-6-6-6. T9 was the early fore-runner in text prediction systems. Developed by Tegic, later purchased by AOL, T9 stands for "Text on Nine Keys." The technology was developed by researchers in the Seattle area working to help people with disabilities communicate more freely. As you type, the sequence is compared to possible letter combinations in a built-in dictionary. If the predicted word is not the one you wanted, you can choose another from a list of other possible words without having to retype. If the desired word doesn't exist in the machine dictionary, you can press buttons to enter multi-tap mode and manually adds the word to the message.

These systems use standard dictionaries to check their words. Criticism of these dictionary-based disambiguating systems focuses on the rigidity of these dictionaries – people feel constrained by the small vocabularies. It's not terrifically friendly to try typing "cool" only to see "book" pop up repeatedly.

Adaptive Advantage

Leading edge mobile technology is often pushed towards business folk and adopted early by geeks and teenagers. Those groups employ non-standard language and wording, demanding devices that can predict not just text, but trends. Elizabeth Nathan, Senior Product Manager for T9 at AOL, says that their product has evolved to accommodate these users. If a user manually enters a word, it becomes part of their personal dictionary. T9 accepts words made of mixed alphanumerics ("2nite" "cula8r"), popular in chat and SMS. T9 now ships with some common online slang built-in.

T9 is combining informal English with dynamic dictionaries: "In early versions of T9, ‘book' always came before ‘cool' because according to our analysis of the English language it was used more frequently and the order determined by our analysis was static." Now the phone actually pays attention to which words the user prefers and pushes those words towards the top of the stack.

T9's adaptive innovations are not all on the machine side – the users have adapted as well. Elizabeth Nathan speaks of power T9 users writing and reading text differently: "I might send you a message that says ‘Call me at good.' A power T9 user knows that ‘good' has the same key sequence of keystrokes as ‘home' and that ‘home' is in the second position and the sender just forgot to hit the NEXT key one time to get that word to display."

Texting with text prediction at a snappy pace requires some learning then – learning how the interface works, and the system learning how you work. A T9 competitor, AirTx, proposes "Solutions that read your mind." According to Jim Hudak, VP of Sales and Marketing, AirTx is constructing an entire suite of technology observation and assistance. Starting with text prediction, AirTx can load a different dictionary for each application (for example, you might tend to use more formal language in email, rather than SMS). They use a 60 kilobyte database, with about 6000 words in it – your common vocabulary.

But AirTx anticipates not only next word, but also next action. They've had some computer intuition systems up and running, but a usability expert at the company recommended they tone it down for consumers; "to avoid the big brother effect: ‘oh wow - it's noon and my phone just asked me if I want to get lunch with my girlfriend'."

Most of us do have predictable patterns to our communications and even our scheduling. More technology might be the only way to manage all the communications made possible by the mobile internet. T9 and AirTx are not alone in their endorsement of this kind of assistance; Motorola recently announced a new version of their iTap software which will predict whole sentences based on past writing.

Power Steering Versus Cruise Control

Anticipating user needs and building an interface to present popular choices sounds a bit like Microsoft's Bob. Bob's progeny, the paper clip in Microsoft Word, has been popularly reviled by computer users – many people don't seem to want a dancing office supply offering help with each stray click.

Eatoni Ergonomics is taking the opposite approach to word and sentence prediction. CEO Howard Gutowitz explains that Eatoni's LetterWise solution works backwards – not predicting words, but predicting letters. Most of the words we write in our native languages follow predictable patterns ("u" after "q" being one of the most predictable in romance languages). Unlike AirTx or T9 storing dictionaries on the devices, Eatoni's software is a list of probabilities. As users multi-tap, the phone predicts the most likely appropriate next letter from the letters associated with that button. Gutowitz employs a driving analogy: "We're interested in power steering, power brakes, features which extend your control, mastery, efficacy and joy in driving, not cruise control or a car which purports to drive itself to a destination."

Gutowitz is proud of the fact that the LetterWise software doesn't change or adapt. Users learn the rules of the system and it works the same each time; a faster way of texting, Gutowitz believes. According to Eatoni studies, roughly 15% of text message content is non-standard – names, addresses. Gutowitz claims that even experienced users of T9-type systems stumble on those non-standard names, slowing them to multi-tap speeds. "Have you ever tried to enter a URL with T9?" he asks.

Both the T9-type word prediction systems and Eatoni's probability metrics intertwine language and intention. It's as though our pens have begun to wrap themselves around our fingers, drawing ink directly from our arteries. We express a feeling, the intention to write a note, and the computer steps up to form words and sentences loosely based on our impulse to communicate. We may correct the computer output, but increasingly, software is our scribe, filtering our words through computer conveniences. It's a process that began with mechanized type; turning over textual reproduction to machines. Few of us scratch out our characters by hand; soon it seems only poets will take the time to write full sentences.

This projection might seem absurd to people who compose from their desks. But increasingly, mobile devices are taking the place of laptops and desktops as people's primary point of interaction with their correspondence. Early text prediction systems have helped faster text messaging.

But, new mobile internet devices offer the promise of posting to the web from the road. This new era of personal publishing on the fly has excited our sense of potential as well as increasing our frustration with current modes of entering data on tiny screens with miniscule keypads. Greater participation in the internet at large demands better human/machine co-writing technology.

What do we stand to lose and gain when we rely on machine to help us express ourselves? Perhaps our frustration with misplaced words will be replaced with more profound frustration at being led in the wrong direction by software purporting to read our minds. There's likely to be an ever greater struggle as we compose more text on our phones, not just to plan a date with a friend, but also to share with the world the subtlety of our experience as we hurtle down the road.

A warm thank you to Shelley Lloyd for her research and writing assistance with this piece.