Choosing features for classifying multiword expressions

Problem
The classification of Multiword Expressions (MWEs) remains a contentious issue within the computational linguistics community. Despite forty years of research, there is no satisfactory consensus on the most relevant features for defining large classes of MWEs. Current classifications often rely on "fuzzy" features—such as semantic weight, semantic decomposability (or analysability), and intuitive notions of syntactic flexibility—which lack clear-cut boundaries. These imprecise features lead to unreliable assignment of MWEs to classes, creating a barrier for computational applications like statistical parsing and machine translation, which require precise, reproducible definitions. Furthermore, recent trends often aggregate loosely correlated features into single, imprecise categories (e.g., "syntactic flexibility"), obscuring the distinct properties of individual expressions.

Methodology
The author adopts a critical review of existing literature, contrasting the "Lexicon-Grammar" (LG) tradition with more intuitive, semantic-based approaches. The methodology emphasizes:

1. Preference for Clear-Cut Features: Prioritizing features that are binary, formally definable, and reproducible (e.g., the applicability of specific syntactic operations, selected prepositions, and coreference constraints) over those relying on speaker intuition or semantic continua.
2. Reproducibility of Observation: Evaluating features based on inter-judge agreement. The paper argues that features requiring introspection (like semantic decomposability) suffer from low reproducibility, whereas formal syntactic tests (e.g., passivization, dative shift, or the insertion of possessive determiners) yield consistent results across observers.
3. Extensive Lexical Inventorying: Drawing on large-scale lexical databases (such as those developed by Gross and colleagues for French, English, and other languages) rather than sporadic examples. The author argues that only comprehensive scrutiny of the lexicon can reveal the true distribution and correlation of features, preventing the overestimation of correlations between loosely related properties.
4. Cross-Linguistic Validation: Utilizing findings from multiple languages (French, English, Italian, Korean, Greek, etc.) to ensure that proposed classification criteria are robust and adaptable, even if specific formal criteria vary by language.

Key Contributions

• Critique of Fuzzy Features: The paper demonstrates that features like "semantic weight" and "semantic decomposability" are ill-suited for computational classification because they exist on a continuum and lack reproducible observation procedures. It highlights that definitions based on these features often fail to classify valid MWEs (e.g., support-verb constructions without equivalent verbs) or create artificial uncertainty.
• Deconstruction of "Syntactic Flexibility": The author argues against treating "syntactic flexibility" as a single, unidimensional feature. Instead, it is shown to be a cluster of loosely correlated, independent features (e.g., optionality of slots, passivization, topicalization). Merging these into a single category reduces the accuracy of the model and fails to provide specific information about an MWE's behavior.
• Advocacy for Lexical Checking: The paper asserts that checking features against comprehensive lexical inventories is essential for validating hypotheses. It critiques the reliance on small, representative samples and the assumption that semantic properties predict syntactic behavior without empirical verification.
• Enhanced Classification Proposals:
  • Figure 1: A decision tree adapted from Baldwin & Kim (2010) but refined to use clear-cut criteria. It distinguishes between lexicalized and non-lexicalized expressions, identifies support-verb constructions using formal criteria, and utilizes Part-of-Speech (PoS) distinctions (including a specific category for prepositional phrases compatible with be) to resolve ambiguities.
  • Figure 2: An alternative classification that expands the definition of support-verb constructions to include copulative constructions (e.g., be angry, be on time), treating the copula as a support verb. This approach groups expressions based on the PoS of the core predicate (noun, adjective, or prepositional phrase) rather than just the presence of a verb.

Results
The analysis reveals that:

• Features derived from extensive lexical description (LG tradition) are more reliable and computationally exploitable than those derived from semantic intuition.
• Correlations between syntactic operations are often loose (e.g., Pearson's coefficient of 0.13 for optionality vs. passivization in one dataset), invalidating the practice of using a single "flexibility" score to define classes.
• Formal criteria, such as the inability to insert possessive determiners in support-verb constructions (e.g., He made his joke vs. He made your joke), provide precise distinctions that semantic weight cannot.
• The hypothesis that semantic decomposability predicts syntactic behavior is unsupported by systematic data; numerous counter-examples exist where analysable idioms lack syntactic variation and non-analysable idioms possess it.

Significance
The paper claims that the choice of features is the primary determinant of a classification's utility for computational use. By shifting from fuzzy, intuition-based features to clear-cut, reproducible, and formally defined features, the community can construct classifications that are stable and scientifically rigorous. The author argues that the labor-intensive work of lexical inventorying is not merely descriptive but foundational for discovering valid hypotheses and avoiding the pitfalls of unverifiable generalizations. The proposed enhanced classifications offer a framework that is adaptable to multiple languages and better suited for the practical needs of NLP, such as parsing and machine translation, by ensuring that class membership is defined by observable, factual properties rather than ambiguous semantic intuitions.