TL;DR
This paper introduces IDION, a comprehensive web-based database of Modern Greek multiword expressions (MWEs), particularly verb MWEs (VMWEs). IDION aims to serve both human users and Natural Language Processing (NLP) systems by providing richly documented data on MWE syntax, semantics, and relations, addressing a critical need for structured linguistic resources in Modern Greek.
What Problem Does This Paper Address?
Developing robust NLP systems for any language requires detailed, structured lexical resources. For Modern Greek, a significant challenge lies in the comprehensive documentation of its multiword expressions (MWEs), which are phrases where the meaning isn't easily predictable from the individual words. These MWEs, especially idiomatic expressions and light verb constructions, are crucial for accurate language understanding. Existing resources are often fragmented or lack the depth needed for advanced NLP tasks. IDION aims to fill this gap by creating a unified, detailed, and accessible database of Modern Greek VMWEs, covering their syntactic flexibility, semantic nuances, and relationships with other MWEs.
Methods
IDION employs a bottom-up approach to document MWEs, starting with a curated list and a detailed documentation manual for editors. The development involves:
- Crowdsourcing: Leveraging the efforts of university students to collect and initially document MWEs, supplemented by expert editors.
- Rich Documentation Template: A detailed template is used to capture a superset of properties typically found in MWE databases, including syntactic flexibility (e.g., word order, interpolation, passivisation), semantic relations (synonymy, opposition, causality), morphological constraints, and usage examples.
- NLP-Oriented Representation: The data is structured to be theory-independent, facilitating its use in NLP applications. This includes encoding part-of-speech, phrasal categories, and various constraints.
- Semantic Relation Encoding: IDION goes beyond simple synonymy by defining terms like "semantic pair" for causative/non-causative relations and encoding "opposite," "stative," and "possessive" relations to build a network of concepts expressed by MWEs.
- Web Editor: A PHP-based web editor with a user-friendly interface and structured tabs facilitates the encoding process, ensuring consistency and ease of data management.
Datasets
The primary dataset is IDION, a newly developed web resource containing approximately 2000 Modern Greek verb MWEs (VMWEs), with about 850 fully documented. The data is collected from:
- Published collections (e.g., Sarantakos, 2013)
- Dictionaries (e.g., Lexigram)
- Native speaker intuitions
- Web retrieval (approximately 8 examples/VMWE)
- Introspective examples (less than 10%)
IDION aims to eventually include frequency information, distinguishing it from some existing databases that rely on large corpora.
Key Insights and Ideas
- Bottom-up Concept Discovery: Defining synonymous MWE sets in a bottom-up manner allows for the emergence of conceptual categories that might not be directly represented by existing verb predicates. This reveals deeper semantic structures in the language.
- Comprehensive Documentation: IDION's extensive template captures a wide array of linguistic properties, providing a richer resource than many existing MWE databases, which is crucial for nuanced NLP tasks.
- Bridging Human and Machine Understanding: The dual focus on human users and NLP systems ensures the resource is both linguistically informative and computationally usable.
- Addressing Emphasis in MWEs: The paper highlights that emphasis with MWEs has received limited attention and proposes encoding it as a feature, acknowledging its importance for capturing the full meaning of expressions.
- Dynamic Synonymy Sets: The transitive property applied to synonymy sets makes them dynamic, meaning they automatically update as new data is added, reducing manual effort and ensuring currency.
Key Results
- IDION currently contains about 2000 Modern Greek VMWEs, with 850 fully documented.
- The database encodes a wide range of syntactic and semantic properties, including syntactic flexibility, semantic relations (synonymy, opposition, causality, stative, possessive), polarity, style, and emphasis.
- A web editor has been developed to facilitate the ongoing documentation and expansion of the database.
- The paper demonstrates the encoding of MWE properties using a detailed template, exemplified by the VMWE "apo to stoma mu to pernis."
Limitations and Future Work
- Limited Corpus Data: The lack of large Modern Greek corpora makes it challenging to obtain frequency information for MWE alternants, a planned future enhancement.
- Ongoing Development: IDION is still under development, with plans to further populate it and add more types of MWEs (nominal, adjectival, adverbial).
- Future Priorities: The authors aim to develop a full network of semantic relations to define "concepts" and use the web resource to identify usage tendencies. They also plan to enhance the ability to encode the frequency of occurrence of VMWE alternants.
Summary updated