鈥淥F ALL the gin joints in all the towns in all the world, she walks into mine.鈥� It鈥檚 a classic quote from the film Casablanca, but can a computer grasp the magic of such memorable lines?
鈥淲hen I started this experiment, people said 鈥榠t鈥檚 cultural, a computer can鈥檛 catch it鈥�,鈥� says at Cornell University in Ithaca, New York.
The doubters may have to think again. Danescu-Niculescu-Mizil and colleagues have taught a computer to identify memorable quotes with an accuracy approaching that of humans. It means computers might one day help writers test their latest catchy lines.
Advertisement
The Cornell team amassed quotes from the (IMDb), which contains a list of lines flagged by users as memorable. The context in which a line is uttered can make a quote more notable, so as a control, the team paired each notable quote with an ordinary one from the same context. It was the same length and spoken by the same character at around the same point in the film.
The computer analysed the pairs of quotes 鈥� around 2200 in total 鈥� for language patterns, unusual words, and word combinations. Unusual words were defined as those that rarely cropped up in text taken from news publications.
The computer managed to identify several characteristics peculiar to the memorable quotes, creating a model that could identify them. 鈥淭he phrases contain surprising combinations of words, but at the same time they have a syntactic structure that is common, so they are easy to use,鈥� says Danescu-Niculescu-Mizil, who will present at a conference on in Jeju, South Korea, in July.
聯Memorable quotes contain surprising combinations of words but their structure makes them easy to use聰
The analysis also showed that memorable lines often have a property the team dubs 鈥済enerality鈥�: they can be widely used because they don鈥檛 contain words that tie them to a specific context. For example, the 鈥� 鈥淵ou鈥檙e going to need a bigger boat鈥� 鈥� might not be judged had Roy Scheider said 鈥渢he鈥� bigger boat.
The model was able to distinguish between memorable and non-memorable quotes with 64 per cent accuracy. Humans .
The team suggests that political campaigners could use the model to assess their slogans.
Take the test at
