from the 80/90 artificial intelligence researchers have been interested in the natural language understanding by machines. Research that has led to semantic representations elaborate but the results, particularly in the context of machine translation is far from perfect. Faced with this particularly difficult problem, Web giants put on another approach, the Deep Learning. Rather than trying to create algorithms seeking to understand the subtleties of human languages, with the Deep Learning, machinery understand for themselves the language by reading books.
When the Deep Learning learns to read
In a post on Facebook, Zuckerberg raised the question of how we could learn to teach AI to read “Alice in Wonderland”. In his speech, he introduced the work of researchers from Facebook AI Research. They have “read” books to Deep Learning algorithm to prepare for a test frequently used in primary schools, the missing word in a text. This allows you to check a child to understand the text submitted to it by raising a word that can be a verb, a noun, a proper noun. Charge it, depending on the context and meaning, to find the missing words.
To prepare their algorithm to “The Children’s Book Test”, the researchers read 98 books in their algorithm, then confronted it with various texts. While it is easy for the algorithm to find adverbs and verbs, it is much more complex for common and proper nouns. The researchers found that the quality of the predictions varied depending on the number of words in the text submitted; They theorized the phenomenon of “ Goldilocks Principle .” To deliver the right answer, you have too much or too little text for the algorithm finds the right answer.
Maybe one day this Deep Learning algorithm will fit on M, AI digital assistance which may be embedded in future versions of Messenger.
Translation : Google Translate
“Can we teach AI to read Alice in Wonderland? “ Bill, Mark Zuckerberg of Facebook, February 18, 2016
“The Goldilocks Principle: Reading Children’s Books with Explicit Memory Representations” , Cornell University Library, November 7, 2015