Update README.md

README.md

@@ -66,6 +66,34 @@ print(cos_sim(text_embeddings[0], text_embeddings[1])) # text embedding similari
 print(cos_sim(text_embeddings[0], image_embeddings[0])) # text-image cross-modal similarity
 ```
 
+**notice: our emperical study shows that text-text cosine similarity is normally larger than text-image cosine similarity!**
+
+If you want to merge two scores, we recommended 2 ways:
+
+1. weighted average of text-text sim and text-image sim:
+
+```python
+# pseudo code
+alpha = 0.6 # text search
+beta = 0.4 # cross-modal search
+
+combined_scores = alpha * sim(query, document) + beta * sim(text, image)
+```
+
+2. apply z-score normalization before merging scores:
+
+```python
+# pseudo code
+query_document_sim_mean = np.mean(cos_sim_query_documents)
+query_document_sim_std = np.std(cos_sim_query_documents)
+text_image_sim_mean = np.mean(cos_sim_text_images)
+text_image_sim_std = np.std(cos_sim_text_images)
+
+query_document_sim_normalized = (cos_sim_query_documents - query_document_sim_mean) / query_document_sim_std
+text_image_sim_normalized = (cos_sim_text_images - text_image_sim_mean) / text_image_sim_std
+# sum normalized scores
+```
+
 ## Performance
 
 ### Text-Image Retrieval