Ilfhocail: A Lexicon of Irish MWEs

Abigail Walsh, Teresa Lynn, Jennifer FosterACL 2019

TL;DR

This paper introduces "Ilfhocail," the first comprehensive lexicon of Irish Multiword Expressions (MWEs) designed for Natural Language Processing (NLP). It addresses the significant lack of MWE resources for minority languages like Irish, which hinders NLP development, by compiling and categorizing MWEs from various sources.

What Problem Does This Paper Address?

Multiword Expressions (MWEs) are crucial components of human language, but they pose a significant challenge for Natural Language Processing (NLP) systems. While research on MWEs has progressed for major languages, minority languages like Irish suffer from a severe lack of technological support and resources. This deficiency impedes the development of NLP tools for Irish, such as parsing, machine translation, and information extraction. Existing Irish treebanks have minimal MWE annotation, and while some theoretical linguistic studies exist, a unified, NLP-ready lexicon is absent. This paper aims to bridge this gap by creating such a lexicon.

Methods

The authors employed a semi-automatic approach to compile the Ilfhocail lexicon. They leveraged several existing Irish lexical resources, including the National Terminology Database (Tearma), Irish WordNet (Líonra Séimeantach na Gaeilge), idiom collections, and dictionaries. MWE entries were extracted from these sources in XML format. A key aspect of their methodology involved developing a lexicon structure with specific columns (GA-Head, GA, POS, EN, Source, ID) to organize the extracted data. They also addressed data cleaning by handling duplicate entries and redundant translations. Furthermore, they proposed a novel categorization scheme for Irish MWEs, drawing inspiration from existing schemes for other languages but adapting them to the specific properties of Irish. A subset of 530 entries was manually annotated to assess the quality of the extracted MWEs and to refine the categorization scheme.

Datasets

The paper describes the compilation of the Ilfhocail lexicon from the following sources:

  • An Bunachar Náisiúnta Téarmaíochta don Gaeilge (Tearma): Approximately 141,031 MWE entries.
  • Líonra Séimeantach na Gaeilge (Irish WordNet): 8,995 MWE entries.
  • Peadar Ó Laoghaire Idiom Collection: 420 MWE entries.
  • Pota Focal Gluais Tí (Pot of Words House Glossary): 375 MWE entries.
  • New English-Irish Dictionary & English-Irish Dictionary: A combined total of 105,358 MWE entries.
  • Foclóir Gaeilge-Béarla (Irish-English Dictionary): 38,775 sense entries.
  • An Foclóir Beag (The Small Dictionary): 771 MWE entries.

The paper also mentions a manually annotated sample of 530 entries for quality assurance and categorization refinement.

Key Insights and Ideas

  • The "Minority Language" Challenge: The paper highlights the significant resource gap for MWEs in minority languages, underscoring the need for dedicated efforts to build NLP tools for these languages.
  • Adaptability of Categorization Schemes: Existing MWE categorization schemes, while useful, often require adaptation to accommodate the unique linguistic features of languages like Irish, particularly regarding verb constructions and idiomatic expressions.
  • The Importance of Lexical Resources: The successful compilation of Ilfhocail demonstrates the value of leveraging diverse lexical resources, even those not explicitly designed for NLP, by extracting and unifying their MWE content.
  • Manual Annotation is Crucial: Despite semi-automatic extraction, manual annotation proved essential for quality control, identifying errors, and refining the categorization scheme, revealing nuances not captured by automated processes.

Key Results

The Ilfhocail lexicon, in its first release, contains 201,795 entries. A manual annotation of 530 entries revealed that Nominal MWEs constitute the largest proportion of the lexicon, largely due to the inclusion of the Tearma corpus. The annotation process also identified issues such as missing POS information and inconsistent headword assignment across different sources, which will be addressed in future versions. The paper also proposes a preliminary categorization scheme for Irish MWEs, distinguishing between non-verbal and verbal MWEs, and further subdividing verbal MWEs into categories like Light Verb Constructions, Verb-Particle Constructions, and Inherently Adpositional Verbs.

Limitations and Future Work

The authors acknowledge several limitations and outline future work:

  • Unifying POS Information: The POS tags extracted from various sources are inconsistent, and efforts will be made to unify them in the next release.
  • Removing Non-MWEs: The manual annotation identified entries that were deemed not to be true MWEs (e.g., highly productive expressions or those with too many non-lexicalized components), and these will be removed in future versions.
  • Automatic Categorization: The paper presents an initial attempt at defining an MWE categorization scheme. Future work will focus on developing automatic methods for categorizing MWE types.
  • Integration with PARSEME: The authors aim to align their categorization scheme with the PARSEME Annotation Guidelines to facilitate participation in future PARSEME Shared Tasks.
  • NLP Applications: Ilfhocail is intended to serve as a valuable data source for various Irish NLP applications, including automatic MWE identification in treebanks, improved Irish parsing technologies, and enhanced machine translation.

Summary updated

BibTeX
@inproceedings{walsh2019ilfhocail,
  author = {Abigail Walsh and Teresa Lynn and Jennifer Foster},
  title = {{Ilfhocail: A Lexicon of Irish MWEs}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
}
Figure