Natural Language Acquisition for Machines - Reinforcement Learning of Minimalist Grammars

语音操作的用户界面便于控制设备和家庭助理。指令和响应可以通过语音控制传递，而不是使用键盘和显示器作为输入输出接口。特别是，在不久的将来，智能家居开发和健康科学应用的智能用户界面变得越来越重要。为了用自然语言与用户交流，“认知主体”需要“心理词典”形式的语言知识，该词典存储单词的发音和拼写、可能的句法组合及其意义。最先进的技术需要手工制作这些数据库的专家，因此在很大程度上排除了技术上的日常应用。该项目的目的是在机器学习的过程中交互地获取心理词典。为了实现这一目标，该项目将把计算语言学的最新发现应用于语言技术。乔姆斯基创立的生成语法最简方案提供了一个复杂的自然语言语法理论，该理论已成功地应用于连接和形式化地实现句法、语义和音系学。最低限度语法的最大特点是通过积极的例子有效地学习。也承认语言习得的反面例子，可以避免说出不符合语法的概括。积极证据和消极证据的结合是“强化学习”的显著标志，从斯金纳开始，强化学习就成为了人工智能研究和行为自适应控制的基本方法。

Voice-operated user interfaces facilitate the control of devices and home assistants. Instructions and responses can be communicated via voice control, instead of using keyboards and displays as input- output interfaces. In particular, intelligent user interfaces for smart home developments and applications in health science become increasingly important in the near future. In order to communicate with a user in natural language, a "cognitive Agent" requires linguistic knowledge in form of a "mental lexicon" that stores word pronunciations and spellings, possible syntactic combinations and their meanings. State-of-the-art technology requires experts who handcraft these data bases, hence considerably excluding technical every-day applications. The project aims at acquiring the mental lexicon interactively during machine learning. In order to achieve this, the project will apply recent findings from computational linguistics in language technology. The "minimalist program" for generative grammar, founded by Chomsky, provides a sophisticated theory of natural language grammar that has successfully been applied for conjoining and formally implementing syntax, semantic and phonology. The most important feature of minimalist grammar for the suggested project is their effective learnability by means of positive examples. Admitting negative examples for language acquisition as well, could avoid the utterance of ungrammatical generalizations. The combination of positive and negative evidence is a distinguished hallmark for “reinforcement learning” which, going back to Skinner, became an essential method for artificial intelligence research and adaptive control of behavior.