Update app.py

app.py

@@ -1,8 +1,58 @@
-import streamlit as st
+import streamlit as st, random
 st.set_page_config(layout="wide")
 
+# Give some context
+st.html("""
+    <h1 style="text-align: center; margin: 0px; text-wrap: balance;">🔀 Word-level alignment between two sentences</h1>
+    <div style="text-align: center; color: gray; text-wrap: balance;">Supports English, French, Dutch, and German.</div>
+    <style>
+        .stButton { text-align: center; }
+    </style>
+""")
+
+# Create a layout with a columns on each side for padding
+_, mid_st, _ = st.columns([1, 2, 1])
+
+# Allow the user to reroll the example sentences
+reroll_button = mid_st.button("Try a new example!", key="reroll")
+if reroll_button:
+    example_sentences = [
+        # translations
+        ("The book, which was on the table, is now missing.", "Het boek, dat op de tafel lag, is nu verdwenen."),
+        ("If I had known, I would have acted differently.", "Si j'avais su, j'aurais agi différemment."),
+        ("She can speak three languages fluently.", "Sie kann drei Sprachen fließend sprechen."),
+        ("I wish I had more time to learn.", "Ich wünschte, ich hätte mehr Zeit zum Lernen."),
+        ("The children were playing while their parents were talking.", "De kinderen speelden terwijl hun ouders aan het praten waren."),
+        ("He would go to the gym every day if he had more energy.", "Il irait à la salle de sport tous les jours s'il avait plus d'énergie."),
+        ("By the time I arrived, she had already left.", "Als ich ankam, was zij al vertrokken."),
+        ("Despite the rain, they went for a walk.", "Malgré la pluie, ils sont allés se promener."),
+        ("If I were you, I wouldn't do that.", "Als ik jou was, zou ik dat niet doen."),
+        ("The movie, which I watched yesterday, was fantastic.", "Der Film, den ich gestern gesehen habe, war fantastisch."),
+        # paraphrases
+        ("She has a remarkable ability to solve problems quickly.", "Her problem-solving skills are impressive and rapid."),
+        ("Despite the fact that the project was delayed, they managed to finish it on time.", "Even though the project was delayed, they were able to complete it by the deadline."),
+        ("The teacher asked the students to submit their assignments by Friday.", "The students were required to hand in their assignments no later than Friday."),
+        ("I haven't seen him in years, and I wonder how he's doing.", "It's been years since I last saw him, and I'm curious about his well-being."),
+        ("He was hesitant to take the offer because it seemed too good to be true.", "He doubted the offer because it appeared to be too perfect to be genuine."),
+        ("She didn't have the necessary qualifications, but she still managed to get the job.", "Even though she lacked the required qualifications, she succeeded in securing the position."),
+        ("John said that he would be going to the meeting later.", "According to John, he planned to attend the meeting later."),
+        ("The weather was terrible, so we decided to cancel the outdoor event.", "Due to the poor weather, we chose to call off the outdoor event."),
+        ("They have lived in this city for a long time, and they're very familiar with it.", "Having resided in this city for many years, they know it quite well."),
+        ("The book was so captivating that I couldn't put it down until I finished it.", "I found the book so engrossing that I read it all the way through without stopping.")
+    ]
+    random_sentences = random.choice(example_sentences)
+    sent0 = mid_st.text_input("Type your first sentence here", value=random_sentences[0], key="sent0")
+    sent1 = mid_st.text_input("Type your second sentence here", value=random_sentences[1], key="sent1")
+else:
+    # Allow the user to input two sentences
+    sent0 = mid_st.text_input("Type your first sentence here", value="De fait, mon mari ne parlait jamais de ses affaires avec moi.", key="sent0")
+    sent1 = mid_st.text_input("Type your second sentence here", value="M'n man had het met mij nooit over z'n zaken, inderdaad.", key="sent1")
+
+
+# Display the mapping between the two sentences
 DEBUG = False
 if DEBUG:
+
     # Use some dummy data
     tokens = [
         ["[0]\u2581De", "[0]\u2581fait", "[0],", "[0]\u2581mon", "[0]\u2581mari", "[0]\u2581ne", "[0]\u2581parlait", "[0]\u2581jamais", "[0]\u2581de", "[0]\u2581ses", "[0]\u2581affaires", "[0]\u2581avec", "[0]\u2581moi", "[0]."],
@@ -30,13 +80,8 @@ else:
 
     model, tokenizer = load_model_and_tokenizer()
 
-    _, mid_st, _ = st.columns([1, 2, 1])
-    sent0 = mid_st.text_input("Type your first sentence here", value="De fait, mon mari ne parlait jamais de ses affaires avec moi.", key="sent0")
-    sent1 = mid_st.text_input("Type your second sentence here", value="M'n man had het met mij nooit over z'n zaken, inderdaad.", key="sent1")
-
-    sentences = [
-        sent0, sent1
-    ]
+    # Encode the sentences
+    sentences = [sent0, sent1]
     tokens = []
     embeddings = []
     for sentence in sentences:
@@ -46,25 +91,17 @@ else:
             tokens.append(tokenizer.tokenize(sentence))
             embeddings.append(embedded_sentence)
 
+    # Calculate the cross-token similarity
     token_similarities = F.normalize(embeddings[0], dim=1) @ F.normalize(embeddings[1], dim=1).T
 
+    # Calculate the overall sentence similarity
     sentence_similarity = F.normalize(torch.mean(embeddings[0], dim=0), dim=-1) @ F.normalize(torch.mean(embeddings[1], dim=0), dim=-1)
 
-    #print("="*60)
-    #print("Mapping sentence1 to sentence2...")
-    #print("="*60)
+    # Map sentence1 to sentence2
     token_probabilities_12 = F.softmax(20*token_similarities, dim=1)
-    for i in range(len(tokens[0])):
-        j = torch.argmax(token_probabilities_12[i])
-        #print(tokens[0][i].ljust(15), tokens[1][j].ljust(15), round(token_probabilities_12[i][j].item(), 2))
 
-    #print("="*60)
-    #print("Mapping sentence2 to sentence1...")
-    #print("="*60)
+    # Map sentence2 to sentence1
     token_probabilities_21 = F.softmax(20*token_similarities.T, dim=1)
-    for j in range(len(tokens[1])):
-        i = torch.argmax(token_probabilities_21[j])
-        #print(tokens[1][j].ljust(15), tokens[0][i].ljust(15), round(token_probabilities_21[j][i].item(), 2))
 
     # Convert to naive python objects
     sentence_similarity = max(0, round(sentence_similarity.item(), 2))
@@ -72,12 +109,12 @@ else:
     token_probabilities_21 = token_probabilities_21.numpy().tolist()
 
 # Simplify the tokens for display
-tokens = [[token[3:].replace("\u2581", " ") for token in sentence] for sentence in tokens]
+tokens = [[token[3:].replace("\u2581", " ").replace("Ġ", " ") for token in sentence] for sentence in tokens]
 
 html = ''
 html += """
 <article>
-    <div>"""
+    <div style="color: gray">"""
 html += f"""{("✅ Congrats!" if sentence_similarity >= 0.65 else "❌ Sorry!")} These sentences have {100*sentence_similarity}% similarity."""
 html += """
     </div>
@@ -99,16 +136,13 @@ html += """
     article {
         font-family: sans-serif;
         text-align: center;
-    }
-    button:hover {
-        background-color: #0056b3;
+        margin-top: 2em;
     }
     p {
         margin: 0.5em;
         font-size: 2em;
         text-wrap: balance;
     }
-
     span {
         animation-name: rotate_bg;
         animation-duration: 15s;
@@ -120,7 +154,6 @@ html += """
         color: rgba(0, 0, 0, calc((50% + 50% * var(--p))));
         text-decoration-color: hsla(161, 100%, 43%, var(--p));
         background-color: hsla(161, 100%, 43%, calc(var(--p) * 0.2));
-
         --p: var(--p0); """
 for i in range(len(tokens[0])):
     html += f"""--p{i}: 0; """
@@ -161,7 +194,6 @@ for i in range(len(tokens[0])):
 html += """
     }
 </style>
-
 """
 
-st.html(html)
+st.html(html)