TL;DR
This document presents the proceedings of the 17th Workshop on Multiword Expressions (MWE 2021), focusing on the role of multiword expressions (MWEs) in end-user applications. It highlights the ongoing tension between traditional linguistic modeling of MWEs and the rise of end-to-end neural models, exploring how MWEs can be better integrated into practical NLP systems.
What Problem Does This Paper Address?
Multiword expressions (MWEs) – phrases like "in the middle of nowhere" or "hot dog" – pose significant challenges for Natural Language Processing (NLP) due to their lexical, syntactic, and semantic idiosyncrasies. Their unpredictable nature makes them difficult to model, impacting tasks such as parsing, machine translation, and information extraction. This workshop specifically addresses the evolving landscape of MWE research, particularly the increasing importance of MWEs in end-user applications and the implications of end-to-end neural models, which may not explicitly represent MWEs. The core problem is how to effectively identify, represent, and utilize MWEs in modern NLP systems to improve their performance on real-world tasks.
Methods
This document is a collection of papers presented at a workshop, rather than a single research paper with a unified methodology. Therefore, the methods vary across the individual contributions. However, the workshop's scope covers a range of approaches:
- Traditional Linguistic Modeling: This includes theoretical work, corpus linguistics, psycholinguistics, and the development of computational frameworks (e.g., CCG, CxG, HPSG) for handling MWEs.
- MWE Identification and Discovery: Methods for automatically finding MWEs in text, often leveraging statistical and machine learning techniques.
- Representation of MWEs: Techniques for annotating and representing MWEs in resources like corpora, treebanks, and lexicons.
- End-User Application Integration: Research focused on how MWEs can be incorporated into practical NLP applications such as machine translation (MT), natural language understanding (NLU), summarization, and social media mining.
- End-to-End Neural Models: Exploration of how modern neural architectures, like transformers (e.g., BERT), can implicitly or explicitly capture MWE knowledge, and how to leverage them for MWE-related tasks.
- Corpus Development: Creation of new datasets and benchmarks, such as the PIE corpus for idiomatic sentence generation and paraphrasing.
Datasets
The document mentions the use of various datasets and resources, including:
- PARSEME shared tasks: These have provided datasets and evaluation measures that have advanced MWE identification.
- PIE: A Parallel Idiomatic Expression Corpus: A newly introduced corpus designed for idiomatic sentence generation and paraphrasing.
- Corpora, treebanks, e-lexicons, and WordNets: General resources used for MWE annotation and representation.
- Song lyrics: Used for data-driven idiom identification.
- German 'stehen unter'-LVCs: A specific corpus for studying light verb constructions.
Key Insights and Ideas
- The dual nature of MWEs in modern NLP: There's a clear tension between the need for explicit linguistic modeling of MWEs and the success of end-to-end neural models that might learn them implicitly. This workshop aims to bridge this gap.
- MWEs are crucial for end-user applications: Despite the rise of end-to-end models, MWEs remain critical for tasks like machine translation, summarization, and understanding nuanced language, especially in social media.
- Contextualized embeddings offer new possibilities: Models like BERT can capture idiomaticity, but understanding how they do so and how to best leverage this knowledge is an active area of research.
- Specialized corpora are vital: The development of resources like the PIE corpus is essential for advancing research in specific MWE-related tasks like generation and paraphrasing.
- Cross-lingual MWE understanding is challenging but important: Research is exploring how to encode both monolingual and cross-lingual knowledge of idiomaticity.
Key Results
As this is a proceedings document, specific aggregated results are not presented. However, the individual papers likely contain quantitative and qualitative findings related to their specific research questions. For example, papers on MWE identification would report precision, recall, and F1 scores, while papers on end-user applications would show improvements in task-specific metrics. The workshop's overall acceptance rate of 36% (7 out of 19 submissions) suggests a competitive selection process.
Limitations and Future Work
- Implicit vs. Explicit Modeling: The ongoing debate about whether MWEs should be explicitly modeled or learned implicitly by neural networks remains a key challenge. Future work will likely focus on finding the optimal balance.
- End-User Application Integration: While progress has been made, further research is needed to seamlessly integrate MWE handling into a wider range of end-user NLP applications.
- Cross-lingual MWEs: Developing robust methods for understanding and processing MWEs across different languages is a significant area for future exploration.
- Resource Development: Continued creation and refinement of specialized datasets and benchmarks for MWE research are crucial.
- Understanding Neural Model Behavior: Deeper analysis into how large language models capture and utilize MWE knowledge is needed.
Summary updated