CI-P: Towards the Creation of a Unified Repository for MultiLingual and CrossLingual Multiword Expressions

George Washington University, Washington DC

Investigators

Abstract

Single concepts that cross word boundaries, such as kick the bucket and traffic light, are typically referred to as multiword expressions (MWE). MWE usage is pervasive in natural languages. MWEs pose a significant challenge from a processing perspective. Explicitly identifying, classifying, and modeling MWEs in text is shown to improve NLP technologies. The objective of this proposal is to conduct exploratory research and build consensus towards the creation of a Unified Repository for Multilingual and Cross-lingual MWE cutting across different languages (Arabic, Chinese, English, Spanish, Persian), genres and domains while adding contextual references and links to existing lexical resources. Understanding the space of MWE across different languages will have a significant impact on multilingual and cross-lingual NLP applications as well as studies in Linguistics in general. Such a large-scale resource will provide for a disciplined investigation into language universals and language-specific studies by adding insights into shared and varying cultural and language concepts as grounds for extensive typological and etymological studies of how concepts are formed across peoples. Multiword Expressions (MWE) occupy a significant portion of the semantic space. With deeper understanding of MWE, their nature, behavior and usage, natural language processing (NLP) practitioners can build more robust systems to achieve the goal of Natural Language Understanding (NLU). Looking at a diverse set of languages simultaneously would lead to more insights into language universals and language specific phenomena having profound impact in how we design overall natural language solutions. The objective of this proposal is to conduct exploratory research and build consensus towards the creation of a Unified Repository for Multilingual and Cross-lingual MWE cutting across different genres and domains while adding contextual references and links to existing lexical resources such as WordNet. The vision is for a repository that is consistently annotated with links across 5 different languages: Arabic, Chinese, English, Spanish, and Persian. This grant will support the following: i) investigating universal linguistic information that is common across MWEs in various languages as well as identifying differentiating features that are language-specific or found in typologically related languages; ii) conducting pilot annotation studies on essential annotation tasks; iii) building the basic infrastructure for harvesting, storing, and annotating the data; iv) developing tools to bootstrap the envisioned resource from existing resources; and, v) Organizing two 2-day workshops to gather input from leading researchers in the field and to build consensus on key issues for such a resource.