TL;DR
This paper introduces a novel approach to modeling Multiword Expressions (MWEs) within the Bulgarian WordNet (BTB-WN) by extending the "catena" framework. The authors propose a method to represent both the figurative and literal meanings of MWEs, addressing the challenge of their semantic compositionality and variability.
What Problem Does This Paper Address?
Multiword Expressions (MWEs) – like "kick the bucket" or "kick the can" – pose a significant challenge in natural language processing. They are not simply a sum of their parts, often having idiomatic meanings that differ from the literal meanings of their constituent words. Furthermore, MWEs exhibit considerable variability in their structure and usage, making them difficult to classify and model consistently. Existing lexical resources and classification schemes for MWEs are often heterogeneous and lack a unified approach to handling their semantic compositionality. This paper tackles the problem of representing MWEs in a structured lexical resource (BTB-WN) by developing a formal model that captures their morphosyntactic properties, semantic compositionality, and the distinction between figurative and literal meanings.
Methods
The core of the paper's methodology lies in extending the "catena" framework, a dependency-based representation originally developed for idioms. The authors adapt this framework to create a more comprehensive lexical entry for MWEs. This new entry includes:
- Lexicon-Catena (LC): Represents the basic syntactic structure of the MWE.
- Semantics (SM): Captures the overall meaning of the MWE, including its figurative interpretation.
- Valency Frame: Details the grammatical requirements of the MWE's components.
- Constituent Word Literal Meanings (LM): A novel addition that explicitly annotates the literal meanings of the individual words within the MWE.
This approach allows for the representation of both the idiomatic meaning (in SM) and the literal meanings of constituent words (in LM), enabling the model to handle cases where MWEs can also have a compositional usage. The authors also leverage a classification scheme developed within the PARSEME COST Action to categorize MWEs into types like nominal, verbal, adjectival, and adverbial.
Datasets
The paper utilizes two primary subsets of the Bulgarian WordNet (BTB-WN):
- Current version of BTB-WN: From this subset, 981 examples of MWEs were extracted.
- Subset mapped to Bulgarian Wikipedia: This subset, containing 506 examples, was used to evaluate the impact of encyclopedic domain knowledge on MWE distribution.
No new benchmarks were introduced; rather, the existing BTB-WN resource was analyzed and extended.
Key Insights and Ideas
- Bridging Figurative and Literal Meanings: The most significant insight is the explicit modeling of both figurative and literal meanings within a single lexical entry. This is crucial for understanding MWEs that can be used compositionally in certain contexts.
- Catena Framework's Adaptability: The paper demonstrates the flexibility of the catena framework, showing its suitability for representing the complex morphosyntactic and semantic properties of MWEs beyond simple idioms.
- Importance of Lexical Semantics: The authors emphasize that understanding the literal meanings of constituent words is vital, even for highly idiomatic MWEs, as it allows for a more nuanced representation and potential for disambiguation.
- Bulgarian MWE Characteristics: The analysis reveals specific patterns in Bulgarian MWEs, such as the prevalence of nominal MWEs in Wikipedia-mapped data and the unique role of NN compounds.
Key Results
The paper presents a detailed classification and distribution analysis of MWEs within BTB-WN, categorizing them into nominal, verbal, adjectival, and adverbial types. The authors extracted 981 MWE examples from the core BTB-WN and 506 from the Wikipedia-aligned subset. The analysis highlights that nominal MWEs are dominant in the Wikipedia-mapped data, likely due to Wikipedia's focus on named entities and general concepts. The proposed catena-based modeling framework is presented as a feasible method for representing MWEs, demonstrating its potential for both Bulgarian and other languages.
Limitations and Future Work
The authors acknowledge that while their approach handles cases where MWEs have a compositional usage, the "handling of ambiguity when the respective element has more than one literal meaning" remains an open question. They also note that while no examples were found where MWE elements lacked a literal meaning, a "more balanced and incremental view on compositionality" is needed, as language is generative. Future work could involve extending the coverage of named entities through Wikipedia and other sources, and potentially incorporating "phrasets" (groups of phrases with similar meanings) into BTB-WN. The development of a special "chunk grammar" is also suggested for handling complex organization names.
Summary updated