British, U.S., and Canadian researchers say they have developed a Web-based machine-language platform that solves crossword puzzles better than existing commercial products by using artificial neural networks, and it could potentially improve machines' language understanding.
Tests showed the system could more accurately answer single-word, short-word combinations, or sentence/phrase clues than commercial software, and also function as a "reverse dictionary" in which the user describes a concept and the platform yields possible descriptive words.
The researchers "trained" the software to recognize words, phrases, and sentences using definitions in six dictionaries and Wikipedia.
"We're seeing a lot more usage of deep learning, which is especially useful for language perception and speech recognition," notes the University of Cambridge's Felix Hill. He says definitions contain an important clue for helping models interpret and represent phrase and sentence meaning. "Our system can't go too far beyond the dictionary data on which it was trained, but the ways in which it can are interesting, and make it a surprisingly robust question-and-answer system--and quite good at solving crossword puzzles," Hill notes.
The platform has limitations, such as an inability to infer a user's intent or wider context when it receives a query.
From University of Cambridge
View Full Article
Abstracts Copyright © 2016 Information Inc., Bethesda, Maryland, USA