The Lexical Overlap Bias in Text Embeddings

This direction is merged into another direction: lexical overlap bias (lexical-semantic conflation). Its last version stays readable and citable here.

Version 1 of 1104 papers126 conceptsfrozen 2026-10-04

Modern Natural Language Processing relies heavily on text embeddings—vector representations that map sentences into a high-dimensional space where "similar" meanings are close together. However, a recurring finding across recent research is that these models often "cheat." Instead of learning deep semantic relationships, they frequently rely on a lexical overlap bias: a heuristic that assumes two sentences mean the same thing simply because they share the same words.

This shortcut leads to significant failures in tasks like paraphrase detection, natural language inference (NLI), and semantic search, where surface-level similarity often masks fundamental logical contradictions.

The Bag of Words Shortcut

The intuition behind many embedding models is that if two sentences share a high percentage of tokens, they likely describe the same concept. While this is often true, it creates a "lazy" model. Yuan Zhang and colleagues exposed this vulnerability by creating the PAWS (Paraphrase Adversaries from Word Scrambling) dataset PAWS: Paraphrase Adversaries from Word Scrambling. PAWS consists of sentence pairs with high word overlap but different meanings (e.g., "The flight from New York to Florida" vs. "The flight from Florida to New York"). They found that state-of-the-art models often fail these tests, proving that high accuracy on standard benchmarks often masks a lack of true structural understanding.

This reliance on surface features is so pervasive that Nicole Peinelt and Maria Liakata argue that standard evaluation metrics are misleading. By identifying "non-obvious" cases where word overlap is low but meaning is identical (or vice versa), they demonstrated that models perform significantly worse when they cannot rely on simple keyword matching Aiming beyond the Obvious: Identifying Non-Obvious Cases in Semantic Similarity Datasets.

Syntactic Blindness and Role Swapping

A major facet of this bias is "syntactic blindness." Because models prioritize the presence of words over their arrangement, they struggle with relational roles. Yan Zhang introduced the SR (Syntax-Relational) benchmark to show that models often cannot distinguish between "Cats chase mice" and "Mice chase cats" How Well Do Text Embedding Models Understand Syntax?.

This problem extends to how models handle entities. Sahil Manchanda and Pannaga Shivaswamy identified a "name bias," where models prioritize character names over plot or semantic content What is in a name? Mitigating Name Bias in Text Embedding Similarity via Anonymization. If two stories share the same character names, a model might rate them as highly similar even if the events are entirely different. Conversely, identical plots with different names are often mapped to distant points in the vector space.

The Logic Gap: Negation and Antonyms

Perhaps the most damaging effect of lexical bias is the inability to process logical operators like negation. In a vector space dominated by word overlap, "I am happy" and "I am not happy" are nearly identical.

Teemu Vahtola and colleagues explored this using the SemAntoNeg benchmark, finding that many models fail to recognize that "not guilty" is a paraphrase of "innocent" It Is Not Easy To Detect Paraphrases: Analysing Semantic Similarity With Antonyms and Negation Using the New SemAntoNeg Benchmark. Similarly, Yash Mahajan and Eduardo Blanco discovered a "SentEval Paradox": models that perform well on downstream tasks often fail basic semantic alignment tests, such as distinguishing a sentence from its antonym ALIGN-SIM: A Task-Free Test Bed for Evaluating and Interpreting Sentence Embeddings through Semantic Similarity Alignment.

Strategies for Deep Semantic Understanding

To move beyond surface-level matching, researchers have proposed several architectural and training interventions:

  1. Structural Encoders: Early work by Richard Socher used Recursive Autoencoders (RAEs) to learn meanings based on the recursive structure of parse trees, rather than flat sequences Dynamic Pooling and Unfolding Recursive Autoencoders for Paraphrase Detection.
  2. Contrastive Learning with Hard Negatives: Tianyu Gao and the creators of SimCSE used NLI datasets to provide "hard negatives"—sentences that look similar but are logically contradictory—to force the model to distinguish between lexical overlap and semantic entailment SimCSE: Simple Contrastive Learning of Sentence Embeddings.
  3. Syntax-Semantic Disentanglement: Models like ParaBART, developed by James Y. Huang, explicitly separate "what" a sentence means from "how" it is structured using adversarial loss, making the resulting embeddings more robust to syntactic changes Disentangling Semantics and Syntax in Sentence Embeddings with Pre-trained Language Models.
  4. Difference-Aware Training: DiffCSE, proposed by Yung-Sung Chuang, adds an ELECTRA-style task that forces the model to detect exactly which words were changed, making it more sensitive to small but semantically critical modifications like negation DiffCSE: Difference-based Contrastive Learning for Sentence Embeddings.

By shifting the training signal from "do these sentences share words?" to "do these sentences share logic?", these approaches aim to build a more human-like conceptual understanding of language.

Papers in this version

Showing 60 of 104 papers.

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_the_lexical_overlap_bias_in_text_embeddi_1c825021,
  title = {The Lexical Overlap Bias in Text Embeddings},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/the-lexical-overlap-bias-in-text-embeddings/rdr_1c8250211fd94776a8c8a41e3e073571},
  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_the_lexical_overlap_bias_in_text_embeddi_1c825021_v1,
  title = {The Lexical Overlap Bias in Text Embeddings},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/the-lexical-overlap-bias-in-text-embeddings/rdr_1c8250211fd94776a8c8a41e3e073571/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}