Modeling Multi Word Expressions and Complex Predicates

25 concepts21 papers

Natural Language Processing (NLP) has long treated the individual word or token as its atomic unit of meaning. From classic n-gram models to modern transformer-based embeddings, the underlying assumption is often compositional: sentences are built from words, and the meaning of a sentence can be derived by aggregating the meanings of its isolated tokens.

However, human language frequently violates this atomicity through Multi-Word Expressions (MWEs) and Complex Predicates. When a speaker uses an idiom like "kick the bucket," a phrasal slang term like "catfish," or a light verb construction like "take a shower," the semantic core cannot be cleanly localized to a single token. Instead, the meaning is non-compositional and emergent across multiple words. When models evaluate text on a strict word-by-word basis, they suffer from a systemic blindspot.

The Syntactic and Semantic Challenge of Complex Predicates

In many languages, core verbal concepts are not expressed by a single verb, but by a combination of a "light verb" (a semantically bleached verb such as do, make, or take) and a nominal or adjectival element. These are known as complex predicates or light verb constructions (LVCs).

Standard morphological lexicons and dependency parsers struggle with this structure because the primary semantic weight resides in the noun, while the syntactic inflection happens on the verb. For instance, in Persian computational linguistics, Sagot et al. highlight that traditional single-word lexical frameworks fail to capture how these multi-word units function as a single semantic block. In subsequent work on the PerLex 2 lexicon, Sagot et al. demonstrate that failing to model the argument structure of these complex predicates severely limits deep semantic modeling and advanced syntactic analysis.

This issue propagates directly into dependency parsing and semantic role labeling. As Seraji et al. observe, traditional parsing models struggle to capture the unified semantic nature of these units, leading to errors in identifying the core predicate-argument structure of a sentence.

Traditional Token-Level Parsing (Incorrect):
[He] -> subject -> [took] -> object -> [a shower]
                      |
               (Bleached Verb)

Holistic Complex Predicate Parsing (Correct):
[He] -> subject -> [took a shower] 
                     (Unified Predicate)

Similarly, in Modern Standard Arabic, Zaghouani et al. show that standard proposition banks encounter severe challenges when assigning semantic roles to complex predicates. Because the primary meaning resides in the noun rather than the verb, traditional argument models that anchor roles to a single verbal head break down.

The Ambiguity and Lexical Induction Blindspot

If a model cannot recognize that a sequence of words forms a single tight semantic unit, its ability to resolve ambiguity or induce word senses degrades. In Word Sense Induction (WSI), algorithms attempt to discover the distinct meanings of a term based on its context. However, Amrami et al. point out that current WSI frameworks are heavily restricted to single-word lemmas, leaving the disambiguation of non-compositional idioms entirely unsolved.

This limitation is not unique to English. When clustering web search results in Russian, Andrey Kutuzov demonstrates that restricting semantic graph construction to single nouns discards critical information contained in compound phrases. Without these complex units, systems cannot accurately distinguish between distinct search intents that are explicitly defined by specific word combinations.

To address the lack of multi-word benchmarks in low-resource settings, Nguyen et al. introduced Vietnamese datasets (ViCon and ViSim-400) specifically designed to evaluate semantic models on multi-word structures. The authors prove that single-word pairs fail to provide a comprehensive assessment of semantic competence in languages where compound expressions dominate.

Capturing Emergent Meaning in Figurative and Evolving Language

Figurative language is perhaps the ultimate test for multi-word modeling. Metaphors, slang, and neologisms rarely operate on isolated words; instead, their meaning emerges from the interaction of multiple tokens.

Le et al. argue that most figurative language models target single-word metaphors, leaving multi-word metaphorical expressions and idioms unaddressed. This prevents models from capturing meanings that are non-localized and emergent across multiple lexical units.

Similarly, slang and in-group language constantly innovate by repurposing multi-token phrases. Zemel et al. show that interpretation frameworks optimized for single-word semantic shifts fail to激 generalize to complex idioms, because they cannot account for the collective meaning of multiple tokens. This is mirrored in automated discovery: Aoki et al. note that token-level detection methods fail to identify phrasal slang or multi-word idiomatic expressions, while Christine de Kock demonstrates that standard lexicon induction methods miss phrasal innovations within online communities because they are structurally biased toward single-token terms.

Even when tracking the historical evolution of human knowledge, Yun et al. observe that restricting conceptual analysis to single words overlooks a massive portion of scientific and technical innovation, which is inherently expressed through multi-word phrases rather than isolated terms.

Shifting Architectures to Sequence and Joint Modeling

To overcome the single-token limitation, researchers are shifting away from word-by-word classification toward sequence labeling, graph-based architectures, and joint extraction frameworks.

In Complex Word Identification (CWI)—the task of spotting text that needs simplification—Gooding et al. demonstrate that linguistic complexity often resides within multi-word expressions rather than individual tokens. By framing CWI as a sequence labeling task rather than an isolated token classification problem, models can successfully capture the syntactic and semantic boundaries of complex phrases that function as a single unit of meaning.

Task Token-Level Approach (Fails) Sequence/Joint Approach (Succeeds)
Complex Word Identification Evaluates words in isolation, missing phrasal complexity. Uses sequence labeling to capture multi-word boundaries (Gooding et al.).
Event Extraction Assumes events are triggered by a single word. Uses transition-based models to link multi-word triggers (Zhang et al.).
Cross-Lingual Retrieval Maps single words across languages, breaking on idioms. Extends bilingual embeddings to represent larger semantic units (Ho et al.).

A similar paradigm shift is occurring in information extraction. In event extraction, traditional systems assume that events are triggered by single words. To break this limitation, Zhang et al. proposed a transition-based neural model that extracts entities and events jointly, allowing the system to identify and link multi-word triggers that span across multiple tokens.

Finally, this challenge extends to cross-lingual retrieval. Ho et al. highlight that bilingual word embeddings are typically restricted to individual words, making it impossible to accurately process or retrieve complex idioms and multi-word compounds in mixed-code queries (e.g., blending a native and target language). Extending these embedding spaces to represent larger, multi-word semantic units remains a critical frontier for deep language understanding.

Key Foundational Work

Computing Word-Pair Antonymy (2008) — contributed 1 concept elements to this direction.
Morphological Analysis (A Formal Approach) (1973) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_multi_word_expression_detection_2761581a,
  title = {multi-word expression detection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multi-word-expression-detection/rdr_2761581a935147dbbf199ecd1944abcc},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_multi_word_expression_detection_2761581a_v1,
  title = {multi-word expression detection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multi-word-expression-detection/rdr_2761581a935147dbbf199ecd1944abcc/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}

See something wrong? Report a mistake