The Cosine Similarity Trap: Why Embeddings Can’t Distinguish “War” from “Union”

2026-02-12 · Source: Naturallanguageprocessing on Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Intermediate, quick

Summary

Modern Natural Language Processing (NLP) models, particularly those relying on cosine similarity and embeddings, struggle to differentiate between complementary and adversarial relationships in word pairs. For instance, while humans perceive "Good–Evil" as conflictual and "Rama–Sita" as cohesive, standard embedding spaces treat both as "semantically related opposites." This limitation stems from embeddings being built on the assumption that words in similar contexts have similar vectors, encoding relatedness rather than the specific nature of the relationship. Vector arithmetic, such as adding embeddings, exacerbates this by performing semantic averaging, which flattens emergent relationships like tension or fit. This issue has practical implications for bias detection, cultural NLP, and representation learning, where understanding relational coherence beyond mere similarity is crucial.

Key takeaway

For NLP engineers developing or evaluating language models, recognize that standard embeddings conflate complementary and adversarial relationships. Your systems may misinterpret nuanced cultural or ethical contexts by treating "Good–Evil" and "Rama–Sita" similarly. Consider exploring methods that probe embeddings along specific semantic axes, like "Balance vs. Conflict," to capture directional and relational information beyond simple similarity, improving model accuracy in complex domains.

Key insights

Language models often cannot distinguish complementarity from conflict due to how embeddings encode relatedness, not relationship.

Principles

• Contextual similarity flattens relational structure.
• Vector addition performs semantic averaging, not interaction.

In practice

• Bias detection requires distinguishing balance from hostility.
• Cultural NLP needs relational coherence beyond co-occurrence.

Topics

• Word Embeddings
• Relational Semantics
• Semantic Similarity
• NLP Bias Detection
• Cultural NLP

Best for: NLP Engineer, AI Scientist, Research Scientist, AI Engineer, Machine Learning Engineer, Data Scientist

Related on AIssential

• Token Embeddings : From numbers to meaning — Token embeddings convert arbitrary token IDs into dense, learned vectors that capture semantic relationships through contextual co-occurrence.
• Emotion Isn’t Universal: Rethinking Sentiment Analysis for a Global Customer Base — Cultural context is critical for accurate sentiment analysis, as emotion expression varies globally.
• Discourse Relations and Connectives in Higher Text Structure — Local discourse analysis can reveal clues about higher text structure and global coherence.
• When Similar Means Different: Evaluating LLMs on Arabic--Hebrew Cognates — Current LLMs struggle with cross-lingual semantic disambiguation, especially false friends and loanwords, due to over-reliance on surface forms.
• From Comments to Trends: Studi NLP dan Computational Linguistics dalam Mendeteksi Topik Viral dari… — NLP and computational linguistics are crucial for detecting social media trends from vast, informal, and diverse comment data.
• How Contexto Actually Works: A Peek Under the Hood — Contexto uses NLP's vectorial representation and cosine similarity to rank words by semantic relatedness.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Naturallanguageprocessing on Medium.