MUDCAT - MUltimodal Dimensions and Computational Applications of AbstracTness

抽象词和具体词之间的区别（比如“dream”和“banana”的对比）代表了一种与自然语言处理（NLP）目的高度相关的语义分类。在这种情况下，我们的项目MUDCAT从数据驱动和面向应用程序的角度研究了抽象化的概念。虽然关于抽象性的长期讨论已经发生在认知科学中，但我们解决并加强了现有定义，数据收集和特征中的关键问题，并拓宽和优化了NLP方法中有效利用的视角。迄今为止，抽象性的定义、集合和应用大多是在没有上下文化的情况下以词类型为基础进行的。相比之下，MUDCAT将开发、利用和应用抽象性的经验维度，同时关注基于标记的、与意义相关的视角，跨词类（名词、动词、形容词）、跨模态（文本、关联、特征、图像）和跨语言（英语、德语、意大利语）。在这种情况下，我们将收集新的人类生成的抽象规范，并利用跨语言迁移来推进规范生成的半自动算法。经验层的主要工作将从大规模语料库中识别和诱导依赖词类的抽象显著维度，同时考虑到句法星座形式的上下文条件（如子分类和修饰）。考虑到抽象在概念上与具体在多模态基础上有所区别，我们将超越文本维度，在自由词联想、特征属性生成和图像中收集和探索抽象的多模态方面。基于类和跨语言的聚类方法将研究多模态特征的语义和语言概括。最后，抽象的多模态跨语言经验知识将应用于已知或期望从抽象知识中获益的NLP任务。因此，我们将开发通用的计算方法，将我们增强的抽象性信息应用于语义挑战：具象语言识别作为具体-抽象映射任务，超长词检测作为语义共性任务。总体而言，MUDCAT将研究英语、德语和意大利语中抽象和具体词汇的定义和应用的跨语言可转移性，同时考虑到目标和上下文的模糊性。

The distinction between abstract and concrete words (such as "dream" in contrast to "banana") represents a semantic categorisation highly relevant for Natural Language Processing (NLP) purposes. In this vein, our project MUDCAT investigates the notion of abstractness from a data-driven and application-oriented point of view. While the most long-standing discussions about abstractness have taken place in the cognitive sciences, we address and enhance critical issues in existing definitions, data collections and characterisations, and broaden and optimise the perspective towards effective exploitation in NLP approaches.Up to date, definitions, collections and applications of abstractness have mostly been performed on a word-type basis without contextualisation. In contrast, MUDCAT will develop, exploit and apply empirical dimensions of abstractness while paying attention to a token-based, sense-related perspective across word classes (nouns, verbs, adjectives), across modalities (text, associations, features, images) and across languages (English, German, Italian). In this vein, we will collect novel human-generated norms on abstractness and exploit cross-lingual transfer to advance semi-automatic algorithms for norm generation. A major effort at the empirical layer will identify and induce word-class-dependent salient dimensions of abstractness from large-scale corpora, taking into account contextual conditions in the form of syntactic constellations (such as subcategorisation and modification). Considering that abstractness is conceptually distinguished from concreteness on multimodal grounds, we will go beyond the textual dimension and collect and explore multimodal facets of abstractness in free word associations, feature-property generation and images. Class-based and cross-lingual clustering approaches will investigate semantic and language generalisations of the multimodal characteristics. Finally, the multimodal cross-lingual empirical knowledge of abstractness will be applied to NLP tasks whose performance is known or expected to profit from abstractness knowledge. Accordingly, we will develop generic computational approaches to apply our enhanced abstractness information to semantic challenges: figurative language identification as concrete-abstract mapping task, and hypernymy detection as semantic generality task. Overall, MUDCAT will investigate the cross-lingual transferability in definitions and applications of abstract and concrete words for English, German and Italian, while taking ambiguity of targets and contexts into account.