Multi-turn MRC Implementation for "A Study on the Information Extraction and Knowledge Injection for Machine Reading Comprehension"

基於 BERT、BERT-HAE、BERT-HAM 以及 ExCorD 的資源修改。 適用於英文多輪機器閱讀理解 QuAC。

本程式碼為論文 A Study on the Information Extraction and Knowledge Injection for Machine Reading Comprehension 於多輪機器閱讀理解的實作部分。實作分為四個部分。

(1) Baseline: BERT, BERT-HAE, ExCorD

(2) Information Extraction（資訊擷取）

(3) Knowledge Graph（知識注入）

(4) Ensemble（N-best 答案進行 Reranking）

Clutering Strategies

Information Extraction 架構

Knowledge Injection - WortNet 架構

Knowledge Injection - PLSA/PRGC 架構

Answer Re-ranking 示意圖

• tf.yml 模型訓練環境 (BERT-HAE, BERT-HAM)
• ex.yml 模型訓練環境 (ExCorD-BERT, ExCorD-RoBERTa, PRGC, K-BERT)
• ex2.yml 模型訓練環境 (ExCorD-DeBERTa)
• prep.yml 資料預處理環境 (Clustering, Ensemble, WordNet, PLSA)

• BERT-HAE: dialog/hea_clustering/bert_hae-master
• BERT-HAE + IE: dialog/hea_clustering/hae_kg
• BERT-HAE + KI: dialog/hea_clustering/hae_kg
• ExCorD: dialog/excord/excord-main
• ExCorD + IE: dialog/excord/excord-clus
• ExCorD + KI: dialog/excord/kbert-ex
• Ensemble: dialog/ensemble

• BERT-HAM + IE: dialog/attentive_cls
• PRGC-based KI: dialog/PRGC-main

• WordNet based KG: dialog/PRGC-main
• PLSA-based KG: plsa
• PRGC based KG: wordnet

• Information Extraction

bash cls/run_cls.sh

• Train & Prediction

bash run.sh

• Evaluate

bash eval.sh

Information Extraction 結果

Knowledge Injection 結果