Update README.md

Add English summary section

README.md

@@ -38,6 +38,16 @@ configs:
         path: nlp2026_extended_per_query_results.csv
 ---
 
+## English Summary
+
+This dataset accompanies our NLP2026 study on language resource design in the RAG era.
+
+Using a 321-episode Japanese dataset derived from the essay series *Shitto Mania*, we show that structured metadata substantially outperforms full text in retrieval tasks.
+
+**Key finding:** structured metadata alone achieves **11.1×** better retrieval performance than full text in TF-IDF retrieval (**59.0% vs 5.3% Recall@10**, STRUCT queries, n=300). This advantage also persists in dense retrieval (E5).
+
+Our central claim is simple: in the RAG era, language resources should be designed not only as readable text, but as **referenceable structures for AI systems**.
+
 # Shitto-Mania-Dic / 嫉妬辞書（Jealousy Dictionary）
 
 **Author cooperation (not supervision):** This dataset was created with the cooperation of essayist **Nami Saito** (material provision and comments). Final analysis/design/publication decisions are the responsibility of **Jun Ikematsu**.