情報抽出と固有表現認識
自由文から人名・地名・関係を取り出す
本ページの本文は英語で提供されています。タイトルと導入は日本語化されています。
この能力とは何か
Takes natural-language text and returns structured fragments: typed entities (person, organisation, date, amount) plus the relations or events among them. Unlike classification it does not give one label to the whole text but locates spans; unlike schema-driven structured output, the extracted targets come from the text itself rather than a fully prescribed field list.
技術的にどう実現するか
Early systems used conditional random fields or rule-based sequence labelling, tagging tokens BIO-style; pre-trained encoders with a tagging head then became standard. Relation and event extraction are often framed as entity-pair classification or as generation of triples directly. More recently, large models perform few-shot or zero-shot extraction against a given schema, avoiding per-type annotation.
代表的な製品
5GPT-4o
2024ネイティブにマルチモーダルな汎用モデル。テキスト・画像・音声をひとつの入口で扱う
Qwen
2023多様な規模とマルチモーダル版を備えたオープンウェイトのモデル群
ERNIE
2019知識増強の事前学習から始まった中国語モデル、その初期の代表的版
Doubao
2023バイトダンスの汎用対話モデルとアプリ
GLM
2023自己回帰空白穴埋め事前学習から始まった中国語の汎用モデル
関連する組織
代表的な用途
- Clause and amount extraction from contracts and filings
- Résumé parsing and talent-pool building
- Drug and symptom recognition in clinical notes
- News events and knowledge-graph construction
どう評価するか
- Span-level F1
- A hit requires both boundary and type to be correct
- Relation F1
- Share of triples (head, relation, tail) matched exactly
- Exact-match rate
- Share of records whose fields are all correct
限界と難しさ
- Nested and overlapping entities are flattened by token-level tagging schemes
- Cross-sentence coreference is hard; pronouns bind to the wrong antecedent
- Domain terms and novel words are missed when the type was unseen in training