Proceedings of the Joint Workshop on Multiword Expressions and WordNet (MWE-WN 2019)

Agata Savary, Carla Parra Escartín, Francis Bond, Jelena Mitrović, Verginica Barbu MititeluACL 2019

TL;DR

This document comprises the proceedings of the Joint Workshop on Multiword Expressions and WordNet (MWE-WN 2019), bringing together researchers from two complementary fields. It showcases advancements in computational linguistics, particularly in understanding and processing multiword expressions (MWEs) and their integration with lexical-semantic networks like WordNet.

What Problem Does This Paper Address?

The workshop addresses the challenge of effectively integrating and processing multiword expressions (MWEs) within computational linguistic frameworks, especially in conjunction with lexical-semantic resources like WordNet. MWEs, such as "in the middle of nowhere" or "to kick the bucket," exhibit unique lexical, syntactic, semantic, and pragmatic properties that make them difficult to handle with traditional single-word processing methods. Wordnets, while rich in semantic relationships, often contain a significant proportion of MWEs, necessitating specialized approaches for their representation and utilization. The workshop aims to foster synergy between researchers working on MWEs and those focused on WordNets, as well as to explore connections with the Construction Grammar community.

Methods

The papers presented at the workshop cover a wide range of methodologies, reflecting the diverse nature of research in MWEs and WordNets. These include:

  • Representation and Integration: Methods for encoding MWEs within WordNet structures, exploring the consequences for lexical-semantic organization, and linking WordNets with existing MWE lexicons.
  • Processing and Modeling: Techniques for word sense disambiguation of both single-word and multiword expressions, cross-wordnet and cross-language comparisons of MWEs, and the use of MWEs in sense-annotated corpora.
  • Theoretical and Computational Studies: Research into computationally-applicable theoretical studies of MWEs and constructions from psycholinguistics, corpus linguistics, and formal grammars.
  • Annotation and Lexical Resources: Development of MWE and construction annotation schemes for corpora and treebanks, and representation of MWEs and constructions in manually or automatically constructed lexical resources.
  • Syntactic and Semantic Frameworks: Application of MWE and construction processing within various syntactic and semantic frameworks (e.g., CCG, CxG, HPSG, LFG, TAG, UD) and in end-user applications like information extraction, machine translation, and summarization.
  • Discovery and Identification: Original methods for the discovery and identification of MWEs and constructions.
  • Language Acquisition and Non-standard Language: Studies on MWEs and constructions in language acquisition and in non-standard language forms like tweets, forums, and spontaneous speech.
  • Evaluation: Techniques for evaluating annotation and processing methods for MWEs and constructions, including retrospective comparative analyses from shared tasks.
  • Machine Learning Approaches: Utilization of pretrained language models for identifying neologisms, cross-lingual transfer learning, multitask learning for capturing MWEs, and the impact of word representations on sequential neural MWE identification.

Datasets

The papers in this workshop utilize a variety of datasets, including:

  • Bulgarian and Romanian Wordnets: Used for studying verbal multiword expressions.
  • The Romanian Corpus: Annotated with Verbal Multiword Expressions.
  • OdeNet and MMORPH: German multiword expression resources.
  • English Noun Compound Representations: For systematic comparison.
  • BTB-WN: A resource for modeling MWEs.
  • FrameNet: Used for semantic modeling of adjective-noun collocations.
  • Modern Greek Multiword Expressions: Represented in the IDION database.
  • Irish MWEs: Documented in the Ilfhocail lexicon.
  • General Corpora: For various annotation, processing, and identification tasks.

The workshop does not appear to introduce a single, overarching new benchmark dataset but rather leverages existing resources and specialized annotated corpora for specific tasks.

Key Insights and Ideas

  • Synergy is Achievable: The workshop's focus on both MWEs and WordNets, along with connections to Construction Grammar, has led to a significant overlap in presented papers (6 out of 20 papers concerned both areas), suggesting the intended synergy effect is being realized.
  • MWEs are Pervasive: Over 50% of entries in the Princeton WordNet are MWEs, highlighting their fundamental importance in lexical resources and natural language understanding.
  • Context is Crucial for MWEs: The workshop explored various methods for identifying and processing MWEs, emphasizing the need for contextual information, whether from corpora, semantic networks, or linguistic frameworks.
  • Beyond Compositionality: A key theme is the non-compositionality of many MWEs, particularly idioms, which poses a significant challenge for semantic analysis and sentiment analysis.
  • Leveraging Lexical Resources: The integration of MWEs into WordNets and the development of specialized MWE lexicons are crucial for advancing NLP tasks.
  • Cross-lingual and Transfer Learning: Research is exploring how to transfer knowledge about MWEs across languages and tasks, a vital step for building robust multilingual NLP systems.

Key Results

The workshop featured 20 presented papers (12 long, 8 short) with an overall acceptance rate of 54%. Specific quantitative results are detailed within individual papers, but the overarching success is measured by the strong participation and the demonstrated synergy between the MWE and WordNet research communities. The invited talk by Aline Villavicencio, "When the Whole Is Greater Than the Sum of its Parts: Multiword Expressions and Idiomaticity," likely provided a high-level overview of the challenges and opportunities in this field.

Limitations and Future Work

The proceedings themselves do not explicitly detail limitations or future work for the workshop as a whole. However, the diverse range of topics suggests several avenues for future research:

  • Deeper Integration: Further exploration of how to seamlessly integrate MWEs into WordNet structures and other knowledge graphs.
  • Advanced Processing: Development of more sophisticated models for MWE identification, disambiguation, and translation, especially for low-resource languages and non-standard text.
  • Constructional Approaches: Continued investigation into the role of construction grammar in understanding MWEs and their acquisition.
  • Evaluation Benchmarks: The creation of standardized, comprehensive benchmarks for evaluating MWE processing systems across various tasks and languages.
  • Real-World Applications: Increased focus on deploying MWE processing techniques in practical NLP applications, such as improved machine translation, more accurate information extraction, and nuanced sentiment analysis.

Summary updated

BibTeX
@inproceedings{savary2019proceedings,
  author = {Agata Savary and Carla Parra Escartín and Francis Bond and Jelena Mitrović and Verginica Barbu Mititelu},
  title = {{Proceedings of the Joint Workshop on Multiword Expressions and WordNet (MWE-WN 2019)}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
}
Figure