杏吧原创

How does a computer program read a novel? Jessica Griggs spoke to Jon Adams to find out

CAN literary criticism ever be considered a science? Perhaps computer programs that can read and interpret literature will provide the answer, an idea that has been exploring. He spoke at the London School of Economics鈥檚 this month.

You fed the novel by Robert Hudson into a text analysis program. Why?

I wanted to see how a computer program that is usually used to analyse and summarise factual texts such as government reports or legal documents would cope with a work of fiction. The program I used was called Alceste, which works by looking at how often words occur near each other.

What was the result?

It managed to identify the central themes of the novel 鈥� football, business and romance. But it couldn鈥檛 distinguish between setting and plot; the fourth theme it picked out was 鈥渨alking in London鈥�, as the characters spend a lot of time walking around the city discussing philosophical ideas. As a reader, we would hardly register the setting as we would be concentrating on the dialogue, but the computer considers both equally.

Would the program be able to distinguish between literary works like Shakespeare and popular fiction such as Dan Brown鈥檚 novels?

No, the program is only capable of telling you what the book is about. That鈥檚 fine for factual prose or some works of fiction when the book is actually about what it鈥檚 nominally about, but many novels have multiple layers of meaning. Often the most obvious theme is not the most important, and this is completely missed by a computer.

Could we ever program a computer or robot to read between the lines to find the layered meanings in a novel?

It鈥檚 starting to happen now. You can do something called , where the program has the different forms and synonyms for every word. The next step would be to include idioms, which would bring the program closer to how people infer meaning from context. Take the British and Irish Lions rugby union team, for example 鈥� you know a rugby player can鈥檛 really be a lion so it must be a metaphor. You can imagine training a computer to pick up on metaphors, but it鈥檚 not clear whether what you would get out at the end could ever be more than a restatement of what you put in.

Do you think it鈥檚 useful for literary criticism to be treated scientifically?

Since literary criticism came into its own as a discipline, people have been trying to work out how to make it an impartial, technical exercise. In the 1960s, a Russian linguist called thought he had discovered a scientific way to measure what he called 鈥渓iterariness鈥�. He found that the French poet Baudelaire and the Russian poets he studied had this quality and it was all very exciting until the analysis was applied to advertising slogans and newspaper copy and got the same result. Trying to define literariness is a futile pastime.

Would a culture of scientific literary criticism hinder novelists? For example, a character in the book Small World by has his life鈥檚 work analysed by a machine that tells him his most commonly written word is 鈥済reasy鈥�. It stifles his creativity.

This is the 鈥済olden goose鈥� fear 鈥� that creativity won鈥檛 work if we take it apart. Yet maybe the flip side of computers reading novels 鈥� computers writing novels 鈥� could shed light on the creative process. We have some evidence that a computer would do a surprisingly good job 鈥� at least at coming up with a plot. You can already buy scriptwriting software that lays out your plot for you, which works because there are only a limited number of storylines that we find interesting.

For a robot, it would be easier to write novels than to read them. A computer writing novels would do much to demystify the process of creativity. I鈥檓 all for that, I just don鈥檛 think we are going to find conclusive scientific support to back up why we consider one book to be 鈥渓iterary鈥� but another not.

鈥淔or a robot or computer program, it鈥檚 easier to write novels than to read them鈥�