ENGLISH is arguably the world鈥檚 most widely understood language. Even
computers can understand it, thanks to speech recognition software. Now
programmers at IBM have cracked a far harder problem鈥攈ow to make computers
understand Chinese.
Modern Standard Chinese, or Putonghua, is a far more difficult language for
computers than English. This is because tone is a vital component of meaning in
Chinese. What English speakers think of as syllables can have a high, rising,
falling, low or neutral tone. The one-syllable word 鈥渕a鈥� can mean 鈥渞ough鈥�,
鈥渉orse鈥�, 鈥渕other鈥�, 鈥渃urse鈥� or 鈥渜uestion鈥�, depending on the tone. So, although
there are 400 of these syllables, the total number of syllables rises to 1400
when tone is taken into account.
Now Julian Chen and his colleagues at IBM鈥檚 Thomas Watson Laboratory in New
York and IBM鈥檚 Beijing Research Laboratory have found a way to make computers
distinguish between them. Chen says the key is to treat the sounds in Chinese
like English phonemes. In English, phonemes are the sounds that distinguish one
word from another. 鈥淧鈥� and 鈥渂鈥� are separate phonemes because they help English
speakers to distinguish words like 鈥減ear鈥� and 鈥渂ear鈥�. So are 鈥渆鈥� and
鈥渦鈥濃�攖hey differentiate 鈥渂ed鈥� from 鈥渂ud鈥�.
Advertisement
The computer program developed by Chen and his colleagues looks for
distinctive main vowel sounds in Chinese. The tone of this part of a word is
used as a reference by the software to determine whether the tone of the word as
a whole is high, low, neutral, rising or falling. 鈥淲e rely on the pitch of the
vowel to determine tone,鈥� says Chen.
The syllables in every word are split into two parts and analysed. 鈥淢a鈥� would
be split into 鈥渕鈥� and 鈥渁鈥�. The vowel, 鈥渁鈥�, is scrutinised for tonal information
and becomes the reference for the other phoneme, 鈥渕鈥�. If the 鈥渕鈥� is higher than
the 鈥渁鈥�, the syllable 鈥渕a鈥� means 鈥渃urse鈥�; if it is lower it means
鈥渉orse鈥濃�攁nd so on
(see Diagram).
Chen has built up a library of 161 of these vowel sounds. Also in the library
are 104 tonemes, which are like phonemes but help the computer to work out the
tone of the sound.
The team then built a statistical model of the Chinese language to help the
computer make sense of speech. It works out the different probabilities of a
word appearing, given the collection of vowels and tones it has heard and the
words it believes appear before and after it.
Chen started his model with a small electronic vocabulary. This was made into
a more representative model with the help of 400 speakers of languages from all
the different regions of China. Having such a diverse mix was important because
many of the regional dialects of standard Chinese are very different from that
spoken in Beijing.