Collaborative Research: CISE-MSI: DP: IIS: Hybrid-Architecture Symbolic Parser with Neural Lexicon

如果一个话语符合说话人的心理语法，那么它就是合乎语法的。但语法上的话语可能模棱两可。歧义，即对一个话语的多种表述和解释，是有效沟通中语言使用的不良副产品；然而，歧义在语言表现中是普遍存在的，因为心理语言表征具有层次结构，而语言外化由于感觉-运动的限制是线性的。这个混合架构符号解析器和神经词典系统（HASPNeL）判断给定的话语是否符合语法，并检测它是否在单词或句子层面上有歧义。HASPNeL也是一种遵循当前极简语法句法理论的人类语言计算认知模型，它必须满足易学性、可进化性和普适性条件。尽管在相关文献中已经记录了混合符号-概率体系结构的优点，但目前还没有任何基于该体系结构的可比较的系统。HASPNeL预计将影响教育和工业应用程序的发展（特别是适用于代表性不足的语言，因为没有足够的可用数据），并进一步研究和推进人类语言认知模型和技术。HASPNeL的混合架构包括一个功能统一解析器和结构生成器，使用符号人工智能方法进行编码，以及一个用于构建功能丰富的词典的机器学习标记器。一个带注释的合成语料库训练一个神经网络系统，该系统正确地识别和标记每个词汇项，并估计语料库中每个类别的可能性。为了考虑词汇歧义，具有不同类别、特征或含义的标记被分配不同的条目。该系统只解析符合语法的话语，通过产生尽可能多的语法表示来识别模棱两可的话语，并计算每个结构描述的可能性。HASPNeL能够通过对语法和词汇特征的微小参数调整来解释句法变化。该项目由MSI和促进竞争研究的既定计划（EPSCoR）共同资助。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

An utterance is grammatical if it conforms to the speaker's mental grammar. But grammatical utterances may be ambiguous. Ambiguity, the assignment of multiple representations and interpretations for one utterance, is an undesirable byproduct of language use for efficient communication; nevertheless, ambiguity is pervasive in language performance, since mental linguistic representations have a hierarchical structure, yet language externalization is linear due to sensory-motor constraints. This Hybrid-Architecture Symbolic Parser and Neural Lexicon system (HASPNeL) judges whether a given utterance is grammatical and detects whether it is ambiguous at the word or sentence level. HASPNeL is also a computational cognitive model of human language following current syntactic theory on Minimalist grammars, which must satisfy conditions of learnability, evolvability, and universality. Although the advantages of a hybrid symbolic-probabilistic architecture have been documented in relevant literature, there is not yet any comparable system based upon this architecture. HASPNeL is expected to impact the development of applications for education and industry (particularly applicable to underrepresented languages for which there is not enough available data), and to further research and advancement of human language cognitive models and technologies. HASPNeL's hybrid architecture comprises a feature-unification parser and structure generator, which is encoded using a symbolic AI approach, and a machine learning tagger that is used to construct a feature-enriched lexicon. An annotated synthetic corpus trains a neural network system that properly identifies and tags each lexical item and estimates the likelihood of each category within the corpus. To account for lexical ambiguity, tokens with different categories, features or meanings are assigned different entries. The system only parses grammatical utterances, recognizes ambiguous utterances by producing as many syntactic representations as there were possible interpretations, and calculates the likelihood for each structural description. HASPNeL is able to account for syntactic variation by minor parametric adjustments to grammatical and lexical features.This project is jointly funded by MSI and the Established Program to Stimulate Competitive Research (EPSCoR).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.