Neural discovery of abstract inflectional structure

点击翻译按钮获取中文摘要

In many languages, words have different forms based on their grammatical role in a sentence. For instance, verbs in Standard English have different forms when the subject is "I/you/we/they" versus "she/he/it" ("I walk", "she walks"). Because many languages have a large number of infrequently used words and many forms of these words, speakers may not be able to memorize all of these forms and must predict some. This project uses computational techniques to measure how predictable the grammatical forms of words are and what information about words (e.g., their sound structure, meaning, or distribution of grammatical forms) contributes to form predictability. The project improves on previous techniques by accounting for cases where unpredictable forms appear in predictable places. For instance, the English verb "sing" has the irregular past tense "sang" and participle "sung". If we know that the verb "give" has an irregular past tense "gave", this (by analogy) increases the chance that the participle is irregular as well, even though the irregular forms involved are different. The project conducts a broad survey of world languages, as well as more focused study of two language families: Romance and Semitic. The project contributes to scientific understanding of the ways in which languages differ from one another and in what respects all human languages must be similar. The project develops practical prediction systems which can be used to improve the ability of technology applications (e.g. automated translators, speech transcription apps, or assistants like Alexa that generate fluent original speech) to handle rare word forms. It also trains a graduate student in advanced computational skills relevant to the technology industry, and the researchers plan to discuss the project in public events designed to engage the local community.This project investigates how inflectional organization facilitates or inhibits the task of predicting previously unobserved inflected forms of words. The extent to which distributional principles shape inflectional organization, facilitate prediction of grammatical forms, and interact with other aspects of morphological organization are typologically not well understood. Drawing on a long history of memory-rich analogical models in morphology, the project develops a computational model which distinguishes between abstract/distributional morphological operations and morphophonological surface form, allowing an independent analysis of the relative difficulty of predicting each dimension. The model is used to conduct a large-scale typological study. This award impacts society in three ways. First, products from this research contribute to computational tools for under-resourced languages, for which inflected form prediction is challenging. All research software created is made publicly available with open-access permissions. Second, the project provides interdisciplinary STEM training. Third, the project supports outreach activities that promote the importance of language diversity and scientific investigation of language via public events.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.