Minimal supervision for language learning: bootstrapping global patterns from local knowledge

Machine Learning;Natural Language Processing;Language acquisition;Psycholinguistics

Connor, Michael

关键词: Machine Learning;    Natural Language Processing;    Language acquisition;    Psycholinguistics;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/29824/Connor_Michael.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

A fundamental step in sentence comprehension involves assigningsemantic rolesto sentence constituents. To accomplish this, the listenermust parse the sentence, find constituents that are candidate arguments, andassign semantic roles to those constituents. Each step depends on prior lexicaland syntactic knowledge. Where do children begin in solving this problem whenlearning their first languages? To experiment with different representationsthat children may use to begin understanding language, we have built a computational model for this early point in language acquisition.This system,BabySRL, learns from transcriptions of natural child-directed speech and makesuse of psycholinguistically plausible background knowledge and realisticallynoisy semantic feedback to begin to classify sentences at the level of ``whodoes what to whom.''Starting with simple, psycholinguistically-motivated representations ofsentence structure, the BabySRL is able to learn from full semantic feedback,as well as a supervision signal derived from partial semantic backgroundknowledge.In addition we combine the BabySRL with an unsupervised HiddenMarkov Model part-of-speech tagger, linking clusters with syntactic categoriesusing background noun knowledge so that they can be used to parse input for theSRL system.The results show that proposed shallow representations of sentencestructure are robust to reductions in parsing accuracy, and that thecontribution of alternative representations of sentence structure to successfulsemantic role labeling varies with the integrity of the parsing andargument-identification stages.Finally, we enable the BabySRL to improve bothan intermediate syntactic representation and its final semantic role classification. Using this system we show that it is possible for a simplelearner in a plausible (noisy) setup to begin comprehending simple semanticswhen initialized with a small amount of concrete noun knowledge and some simplesyntax-semantics mapping biases, before acquiring any specific verb knowledge.