Building Natural Language Generation Systems (review)

Now that we know how to store complex knowledge in computers, how do we get it out again? Producing human-readable expository prose—as opposed to tables or fill-in-the-blank boilerplate texts—is a significant challenge. But much of the knowledge stored in computers wasn’t put there by human beings; it was produced or heavily modified by computation.

Natural language generation (NLG) is not the inverse of natural language understanding (NLU), because the goal is to produce, not just any grammatical text that encodes the knowledge, but rather a good one. Thus NLG is controlled by pragmatic concerns—topic, comment, relative importance, expository sequence—that are almost completely ignorable by NLU systems. An NLG system need not deploy the entire grammar of the language, and it certainly need not deal with ill-formed input or variant dialects.

A typical NLG task is distilling a table of numbers into a verbal description such as, ‘The month was cooler and drier than average, with the average number of rain days. The total rain for the year so far is well below average. There was rain on every day for eight days from the 11th to the 18th’ (8). The computer must judge what facts are worth reporting (e.g. the 8-day rainy spell) and the order in which they should be presented.

The whole process is, in fact, related to the planning of robot motion. Goals have to be selected and arranged into a coherent sequence. Planning can be done bottom-up (by combining individual facts into a larger structure) or top-down (by trying to fill in a specification for the complete text, parts of which may be optional). Either way, there are many alternatives to explore at each stage, and the planning process is crosscut by a second problem: generation of referring expressions. Natural language uses pronouns and definite NPs to avoid repetition; using them appropriately in generated text is nontrivial.

Though not a solved problem, NLG is maturing. Besides taking the reader through their own experiments, Reiter and Dale review the literature and describe several NLG systems that have been put to practical use.

I have two predictions about NLG. First, just as with machine translation, the most successful applications in the short term will be those in which even low-quality output is much better than the unprocessed alternative. Second, as the quality of NLG improves, ethical problems will arise. It will become easy to pass off computer output as the product of human thought or to make the computer seem smarter than it is. I have encountered this already with a consulting client who wanted to use random variation to conceal the simple mechanical process by which test scores were being turned into English descriptions. Joseph Weizenbaum (Computer power and human reason, Freeman, 1976) encountered it over 30 years ago with his classic ELIZA computer program, which carried on human-like conversations without understanding their content.