Update app.py

app.py

@@ -7,7 +7,7 @@ semantic_model = SentenceTransformer('all-MiniLM-L6-v2')
 
 @st.cache(allow_output_mutation=True)
 def get_model(model):
-	return pipeline("fill-mask", model=model, top_k=100)#seto maximum of tokens to be retrieved after each inference to model
+	return pipeline("fill-mask", model=model, top_k=100)#set the maximum of tokens to be retrieved after each inference to model
 
 
 HISTORY_WEIGHT = 100 # set history weight (if found any keyword from history, it will priorities based on its weight)
@@ -15,28 +15,37 @@ HISTORY_WEIGHT = 100 # set history weight (if found any keyword from history, it
 st.caption("This is a simple auto-completion where the next token is predicted per probability and a weigh if appears in user's history")
 
 history_keyword_text = st.text_input("Enter users's history keywords (optional, i.e., 'Gates')", value="")
+#history_keyword_text=''
 
 text = st.text_input("Enter a text for auto completion...", value='Where is Bill')
-semantic_text = st.text_input("Enter users's history semantic (optional, i.e., 'Microsoft')", value="Microsoft")
+#text='Where is Bill'
 
-model = st.selectbox("choose a model", ["roberta-base", "bert-base-uncased"])
+semantic_text = st.text_input("Enter users's history semantic (optional, i.e., 'Microsoft or President')", value="Microsoft")
+#semantic_text='President'
 
-data_load_state = st.text('Loading model...')
+model = st.selectbox("choose a model", ["roberta-base", "bert-base-uncased"])
+#model='roberta-base'
 nlp = get_model(model)
+#data_load_state = st.text('Loading model...')
+
+
 
 if text:
-    data_load_state = st.text('Inference to model...')
+#    data_load_state = st.text('Inference to model...')
     result = nlp(text+' '+nlp.tokenizer.mask_token)
-    data_load_state.text('')
-    
-    if len(semantic_text):   
-        predicted_embeddings = model.encode(result['sequence'], convert_to_tensor=True)
-        semantic_history_embeddings = model.encode(semantic_text.spllit(','), convert_to_tensor=True)
+#    data_load_state.text('')
+    sem_list=[_.strip() for _ in semantic_text.split(',')]
+    if len(semantic_text):
+        predicted_seq=[rec['sequence'] for rec in result]
+        predicted_embeddings = semantic_model.encode(predicted_seq, convert_to_tensor=True)
+        semantic_history_embeddings = semantic_model.encode(sem_list, convert_to_tensor=True)
         cosine_scores = util.cos_sim(predicted_embeddings, semantic_history_embeddings)
     
     for index, r in enumerate(result):
         if len(semantic_text):
-            result[index]['score']*=cosine_scores[index][index]
+#            for j_index in  range(len(sem_list)):
+                if len(r['token_str'])>2: #skip spcial chars such as "?"
+                    result[index]['score']+=float(sum(cosine_scores[index]))
         if r['token_str'].lower().strip() in history_keyword_text.lower().strip() and len(r['token_str'].lower().strip())>1:
             #found from history, then increase the score of tokens
             result[index]['score']*=HISTORY_WEIGHT
@@ -44,5 +53,5 @@ if text:
     #sort the results        
     df=pd.DataFrame(result).sort_values(by='score', ascending=False)
     
-    #show the results as a table
+#    show the results as a table
     st.table(df)