Machine-assisted translating still isn’t perfect. When you input a text into Google Translate the results are often… special. Wouldn’t it be great if computers were able to spit out perfect results?
Dieke Oele says there is room for improvement. After a master in Antwerp in computational linguistics – also known as computer linguistics – Oele started her research as part of the QTLeap project (Quality Translation by Deep Language Engineering Approaches). Eight European universities are working together on the a project to develop better automated translation systems. Their efforts have led to the creation of an IT help desk chat system that can provide answers in several languages.
For the past few years, Oele has been focused on homonyms, or words that have several meanings. For example, ‘figure’. In the sentence, ‘there was a nice figure on the cheque I received’ the word means something completely different than it does in the sentence ‘she has a nice figure’. Context is essential for understanding homonyms.
But try teaching context to a computer. ‘I use what I call word embedding to do this. These are long numerical sequences that represent the concept of certain words. By reproducing each word in the sentence as an embedding, you can do “maths” with words’, Oele explains.
Oele uses her method to calculate the similarity of one word to other words in the same sentence. She then has the computer select the meaning with the highest score, thus determining context with a simple calculation.
Unfortunately, she hasn’t gotten the results she hoped for. ‘The error margin in determining the meaning is still too large’, she says. Because the computer didn’t always link the correct translation to the relevant word, the translation system did not actually improve.
During this time Oele also took part in a pun contest for computational linguistics. Contestants had to teach their systems to recognise and understand puns. Consider the sentence: ‘I used to be a banker, but I lost interest’, Oele says – the trick is to make the system understand both meanings of the word ‘interest’. ‘We scored better than most others, but our scores were still too low. It was an even more difficult thing to do than finding just a single meaning.’ Oele says her team performed well and even made it to second place. ‘But there wasn’t a lot of competition.’
Oele currently works in Germany, where she is using her knowledge of automated word recognition to develop an automated personal assistant – not unlike Siri or Alexa – for Samsung. If you want to open a new savings account and ask your phone where to find a bank, you don’t want it to send you to the nearest river or lake because it doesn’t recognize the question. ‘Ultimately they want to roll out these kinds of systems for different machines, like refrigerators and washing machines. In a few years, you might be able to ask your fridge if you’re out of milk.’