Skip to content

Phase 2 — Prompting Strategy

NER and Local Relation Extraction

Chain-of-Thought Approach

The NER prompt uses a two-step chain-of-thought (CoT) strategy before classification:

  1. Structural reasoning first — the model identifies logical connectives and argument structure (e.g., "X implies Y", "A contradicts B", "because", "therefore"). This primes the model to treat philosophical sentences as relational structures rather than isolated facts.
  2. Entity and relation classification — only after the structural pass does the model assign labels and extract triples.

This order matters for philosophical texts: a sentence like "Hegel's dialectic presupposes Kant's transcendental idealism" contains a SETZT_VORAUS relation between two KONZEPT nodes, but a model that does not first identify " presupposes" as a logical connective will miss it.

Alias Resolution within a Chunk

The prompt includes an explicit instruction to resolve within-chunk aliases: "Resolve pronouns or aliases that are clearly identifiable within this chunk." The LLM normalises the entity name it writes — so "he", "the philosopher", "I. Kant" all become "Kant" (or the canonical form it infers from context). This eliminates a whole class of phantom duplicate entities.

Cross-chunk aliases are not resolved here. They are handled downstream by Phase 3b Latent Graph Consolidation. See docs/workflow/3b_consolidation/index.md and docs/adr/0003-coreference-absorbed-into-fusion.md.

Default Prompt Template

The following is the prompt used in pipeline/prompts/default_prompts.py (NER_EXTRACTION_PROMPT). It is injected with {entity_types} and {relation_types} at runtime from SchemaConfig.

You are an expert in formal logic and the history of philosophy.
Extract all entities and their local relations from the following text chunk.

### Entity Types (use exactly these labels)
{entity_types}

### Relation Types (use exactly these labels)
{relation_types}

### Instructions
1. First identify logical connectives and argument structure (chain-of-thought step).
2. Then classify each entity span and extract local triples.
3. Resolve pronouns or aliases that are clearly identifiable within this chunk.
4. Output strictly valid JSON matching the schema below. No text outside the JSON.

### Output Schema
{
  "entities": [{"id": "<str>", "label": "<entity_type>", "name": "<canonical name>", "text": "<span>"}],
  "triples":  [{"subject_id": "<str>", "predicate": "<relation_type>", "object_id": "<str>", "confidence": <float 0-1>}]
}

### Input Text
{chunk_text}

This prompt is overridable per-run via Phase2Config.ner_prompt_template.

Key prompt design choices

Choice Rationale
Canonical name separate from text Allows normalisation (e.g., "the Königsberg philosopher" → "Kant") while preserving the original span in text for provenance
LLM-local id (e1, e2, …) Decouples triple references from stable graph IDs; remapping happens in LLMNERExtractor after parsing
confidence on triples Enables noise filtering in later phases; LLMs often assign lower confidence to inferred vs explicit relations
Strict JSON output instruction Required for astructured_predict to parse reliably; hallucinated prose before/after JSON breaks structured output

Entity Linking (Disambiguation)

Entity linking during Phase 2 matches extracted entities against canonical graph nodes using string containment (NameEntityLinker). Downstream cross-chunk deduplication and consolidation is handled non-generatively by Phase 3b Latent Graph Consolidation. See docs/workflow/3b_consolidation/index.md.

Tuning guidance

  • Dense philosophical texts — consider raising Phase2Config.batch_size to 100+ chunks if per-chunk extraction is fast; LLM latency dominates over I/O at small batch sizes.
  • German-language sources — entity types KONZEPT, SCHULE_STROEMUNG, etc. are German-labelled by design. If the LLM struggles with German labels, consider using English aliases in SchemaConfig and mapping back at output time.
  • Long chunks — chunks exceeding ~800 tokens have higher NER miss rates because the model attends less to entity spans at the end. The default Phase1Config.chunk_size = 1024 tokens is a reasonable upper bound; consider reducing to 768 for dense texts.