Phrase Level Simplification of Multi Word Expressions

58 concepts60 papers1 proposal

Traditional natural language processing models often operate under a "unigram assumption": they treat language as a sequence of independent, single-word tokens. When a text simplification system encounters a difficult sentence, its default strategy is to find complex words and swap them with simpler synonyms.

However, human language does not always store meaning in individual words. Meaning frequently aggregates into Multi-Word Expressions (MWEs), idioms, and specialized phrases. When complexity resides within the phrasing itself rather than the individual words, the unigram approach breaks down entirely. For instance, trying to simplify the idiom "kick the bucket" by replacing individual words might yield "strike the pail"—completely destroying the intended meaning.

To build effective text simplification and understanding systems, researchers are shifting from single-word substitution to phrase-level processing.

The Tokenization Bottleneck

Why do standard models struggle with phrases? The problem lies in how words are represented mathematically. As Aina Garí Soler et al. observe in A comparative study of word embeddings and other features for lexical complexity detection in French, standard word-level embeddings do not readily represent multi-word expressions. Because these models map single tokens to isolated vectors, a compound phrase gets fractured during preprocessing. This omission is highly problematic for text simplification systems, as idiomatic or compound phrases are frequently the primary sources of linguistic difficulty for readers.

This fracturing creates a cascade of errors across downstream tasks. In lexical simplification pipelines, Goran Glavaš et al. point out in Simplifying Lexical Simplification: Do We Need Simplified Corpora? that a lack of capability to process multi-word expressions or idiomatic phrases severely limits a system's real-world effectiveness. Similarly, in machine translation and word-sense disambiguation, David Vickrey et al. show in Word-Sense Disambiguation for Machine Translation that preprocessing and aligning words individually fails to treat multi-word phrases as single, atomic units, leading to sub-optimal translations of complex terms.

The High Stakes of Specialized Phrasing

The limitation of single-word substitution becomes particularly acute in technical and medical domains, where complexity is almost inherently phrasal. Medical jargon, for example, rarely consists of isolated difficult words; instead, it relies on multi-word clinical terms.

In their work on health literacy, Tarek Sakakini et al. (Context-Aware Automatic Text Simplification of Health Materials in Low-Resource Domains) demonstrate that one-to-one word replacement cannot simplify complex medical instructions. Instead, systems require phrase-level rephrasing or expansion to translate clinical terms into patient-friendly language.

This is supported by Anna Hätty et al. in Predicting Degrees of Technicality in Automatic Terminology Extraction, who note that because a large portion of technical terminology is composed of multi-word expressions, extraction systems limited to single words provide an incomplete and impractical representation of a domain's specialized vocabulary.

Discarding Intent in Search and Logic

The necessity of phrase-level handling extends beyond simplification into information retrieval and logical inference. When a user inputs a search query, treating words independently discards the semantic connection between them.

Hui Fang demonstrates in A Re-examination of Query Expansion Using Lexical Resources that traditional query expansion frameworks miss critical context because they focus primarily on individual words rather than multi-word expressions. To capture true user intent, Michael Bendersky et al. (Parameterized concept weighting in verbose queries) show that expansion frameworks must move beyond single terms to integrate complex linguistic structures and named entities into the latent concept generation process.

Even formal reasoning systems suffer from this unigram constraint. In natural language inference, Masashi Yoshikawa et al. (Combining Axiom Injection and Knowledge Base Completion for Efficient Natural Language Inference) point out that logic-based models are often restricted to generating word-to-word lexical relations. Without the ability to produce complex phrasal axioms, these models cannot infer implications that depend entirely on multi-word expressions.

How Phrase-Level Processing Works

To overcome the unigram trap, modern architectures treat multi-word expressions as atomic units of meaning rather than combinations of independent words. This requires a shift in three core areas:

  1. Boundary Identification: Before any simplification or translation occurs, the system must recognize where a multi-word entity begins and ends. As Li Shao et al. suggest in Mining New Word Translations from Comparable Corpora, integrating phrase chunking or named entity recognition is essential to identify multi-token units before analysis.
  2. Joint Contextual Modeling: Instead of evaluating a word's complexity or meaning in isolation, models must evaluate the joint context of the phrase. This ensures that the non-compositional meaning of idioms or technical terms is preserved in the vector space.
  3. Many-to-Many Mapping: Simplification engines must be capable of asymmetric substitutions—mapping a multi-word expression to a single, simpler word, or expanding a dense technical term into an explanatory phrase.

By treating phrases as primary linguistic citizens, these approaches bridge the gap between literal word-swapping and true semantic simplification.

Key Foundational Work

Finding content-bearing terms using term similarities (1999) — contributed 1 concept elements to this direction.
Corpus-based Method for Automatic Identification of Support Verbs for Nominalizations (1995) — 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_simplification_6b218ad8,
  title = {multi-word expression simplification},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multi-word-expression-simplification/rdr_6b218ad899fa4a708f98937054cdef42},
  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_simplification_6b218ad8_v1,
  title = {multi-word expression simplification},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multi-word-expression-simplification/rdr_6b218ad899fa4a708f98937054cdef42/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake