Most practical systems treat a sentence as:

• syntax (grammar structure)
• semantics (meaning)
• token sequence
• statistical pattern
• embedding/vector representation

Then they compare or recombine at multiple levels.

Example:

Sentence A:

“the expensive salmon”

Sentence B:

“I bought for my own dinner”

A practical system does NOT usually combine raw text blindly.

Instead it extracts structures like:

Grammar:

• DET + ADJ + NOUN
• SUBJECT + VERB + PURPOSE

Semantics:

• food item
• ownership
• purchase action
• personal consumption

Then it searches for compatible continuations.

1. SIMPLEST METHOD Token probability

LLMs mainly work using:

P(next_token | previous_tokens)

Example:

Input:

“the expensive salmon”

Possible continuations:

• was delicious
• smelled awful
• I bought yesterday
• I bought for my own dinner

The model ranks possibilities by probability.

This already creates recombination behavior.

2. EMBEDDING SIMILARITY Most practical modern method

Convert sentence → vector.

Example:

“The expensive salmon” → [0.12, -0.44, 0.88, …]

“I bought seafood for dinner” → [0.10, -0.39, 0.81, …]

Then compare using cosine similarity.

Common uses:

• semantic search
• RAG
• clustering
• deduplication
• recommendation
• sentence matching

Best practice:

• use transformer embeddings
• cosine similarity
• ANN/vector database

Common tools:

• sentence-transformers
• spaCy vectors
• FAISS
• ChromaDB
• Milvus

3. STRUCTURE MATCHING Very important for generation

Systems often learn reusable templates.

Example:

Pattern:

“the [ADJ] [NOUN] I bought for my own [NOUN]”

Can generate:

• the expensive salmon I bought for my own dinner
• the fancy wine I bought for my own party
• the rare book I bought for my own collection

This is compositional generation.

Very important concept:

• reusable latent structures

4. VALIDATION “How do we know it makes sense?”

Practical systems use multiple checks.

1. Grammar validation

• POS tagging
• dependency parsing
• grammar models

2. Semantic plausibility Example:

• “cat ate salmon” → plausible

• “rock ate salmon” → low plausibility

3. World knowledge Example:

• salmon is edible
• cats eat fish

4. Statistical likelihood If many similar patterns appeared during training, probability increases.

5. BEST PRACTICES Most practical systems

1. Use embeddings for semantic similarity NOT raw string matching.

Bad:

• exact text compare

Good:

• vector similarity

2. Separate:

• syntax
• semantics
• entities
• actions

This improves recombination quality.

3. Use chunk-level recombination NOT random word mixing.

Good chunks:

• noun phrases
• verb phrases
• clauses

Example:

• “the expensive salmon”
• “I bought yesterday”
• “for my own dinner”

4. Validate generated text before accepting.

Typical checks:

• grammar
• semantic coherence
• contradiction detection
• toxicity
• domain rules

5. Store latent representations instead of memorizing sentences.

Modern systems prefer: sentence → embedding/vector

rather than: sentence → raw lookup table

6. ADVANCED SYSTEMS

More advanced architectures combine:

• embeddings
• symbolic logic
• knowledge graphs
• memory retrieval
• syntactic trees
• world models
• probabilistic decoding

This allows:

• novelty
• consistency
• reasoning
• controllable generation

instead of simple copying.

7. IMPORTANT REALIZATION

Human language itself is highly compositional.

Humans also recombine learned fragments:

• adjective patterns
• verb patterns
• social scripts
• semantic templates

So:

“new sentence”

usually means:

• new arrangement of known structures

NOT:

• completely alien creation.