Literary Network · Knowledge & Entity Graph
Interactive semantic topology of literary characters, places, and interactions.Literary Knowledge Topology & Named Entity Recognition (NER) Methodology
| 1. NER Position in the NLP Hierarchy | Named Entity Recognition (NER) operates as a higher-order semantic layer in Natural Language Processing (NLP). Moving beyond low-level tokenization, lemmatization, and part-of-speech (POS) tagging, NER locates and classifies entity spans into structured ontological categories (characters, locations, geopolitical entities, facilities). In classical literary corpora characterized by archaic syntax, shifting narrative registers, and honorifics, standard statistical extractors require specialized calibration to avoid misattributing descriptive nouns as proper entities. |
| 2. Deterministic Co-occurrence (Dialogue Sliding Window ±2 Anchors) | Network topology is generated via an in-house deterministic sliding-window algorithm rather than opaque off-the-shelf relation extractors. Instead of treating each sentence as an isolated silo, the engine monitors co-occurrences across a local span of ±1 to ±2 adjacent sentence anchors, capturing rapid dialogue exchanges across speakers (e.g., query at anchor i, reply at anchor i+1). Edge weight reflects interaction frequency and proximity, while preserving the exact set of underlying sentence_id anchors for direct 1:1 inspection in the Parallel Reader. |
| 3. Two-Tier AI Verification & Canonization (spaCy + LLM) | To eliminate entity fragmentation and statistical hallucinations, the architecture implements a two-pass hybrid pipeline: (1) spaCy Linguistic NER conducts base extraction isolating entity mentions across four canonical classes (PERSON, LOC, GPE, FAC); (2) LLM Coreference & Alias Fusion resolves aliases (e.g., consolidating 'Sherlock', 'Holmes', and 'Mr. Holmes' into a single authoritative node), removes misclassified artifacts, and exports a pristine network.json topology tailored for D3.js force-directed physics. |