Preposition Vs Semantic Chunking

🧩 Prepositional vs Semantic Chunking

Metric / FeaturePrepositional ChunkingSemantic Chunking DefinitionSplits text based on logical propositions (e.g., clauses, prepositions, conjunctions).Splits text based on semantic similarity using embeddings (e.g., sentence transformers). GranularityFine-grained: clause-level or logical unit segmentation.Variable: sentence-level or multi-sentence chunks based on meaning. Context PreservationModerate: may lose broader context across clauses.High: preserves semantic coherence across chunks. Chunk Size VariabilityLow to moderate (clauses are often short).High (semantic similarity may group multiple sentences). Computational CostModerate: requires syntactic parsing.High: requires embedding generation and clustering. Speed🟡 Medium🔴 Slower Accuracy in Retrieval TasksGood for logic-heavy or entailment tasks.Excellent for semantic search and RAG. Robustness Across DomainsLimited: depends on grammar and sentence structure.High: adapts well to diverse domains. Overlap SupportRarely used.Common: overlapping windows improve recall. Noise SensitivityHigh: punctuation errors or grammar issues degrade performance.Moderate: embeddings are more tolerant to noise. Best Use CasesReasoning, entailment, legal texts, logic-heavy documents.Semantic search, RAG, summarization, clustering. Example Input“Although it was raining, we went outside.” → [“Although it was raining”, “we went outside”]“We went outside despite the rain.” → grouped with similar weather-related sentences. Evaluation MetricsClause precision, syntactic coverage, logical completeness.Embedding coherence, retrieval recall, semantic overlap.

🧰 Top Libraries & Tools

LibraryPrepositional ChunkingSemantic Chunking spaCy✅ Dependency parsing for clause extraction.❌ Not designed for semantic chunking. NLTK✅ Grammar-based chunking with regex and POS tags.❌ Limited semantic capabilities. LangChain / LlamaIndex❌ Not built-in, but can be customized.✅ Native support for semantic chunking via embeddings. Transformers (HuggingFace)❌ Parsing not supported.✅ Embedding models like BGE, MiniLM, etc. Chunking_RAG (GitHub)🟡 Experimental support via agentic chunkers.✅ Semantic chunking with nomic-embed-text, etc.

Sources: 2

🧪 Example Use Case: C4 Dataset

Let’s say you’re working with the first million samples of C4:

Prepositional Chunking: Use spaCy to parse each sentence and extract clauses. Ideal for logic-heavy QA or entailment tasks.

Semantic Chunking: Use LangChain or LlamaIndex with BGE embeddings to group semantically similar sentences. Ideal for RAG or semantic search.