added explanations

app.py

@@ -3,7 +3,10 @@ import transformers as tf
 import pandas as pd
 from datetime import datetime
 from plotly import graph_objects as go
-
+from transformers_interpret import SequenceClassificationExplainer
+from annotated_text import annotated_text
+from palettable.scientific.sequential import Devon_10_r
+from palettable.colorbrewer.diverging import RdYlGn_10, PuOr_10, BrBG_10
 from overview import NQDOverview
 
 import torch
@@ -94,10 +97,79 @@ with st.form('comment_form'):
         st.experimental_rerun()
     
 results = run_models(models_to_load, models, st.session_state['comment'])
-# Modify results to sum the QuAL score and to ignore Q3 if Q2 no suggestion
-# if results['q2i']['label'] == 1:
-    # results['q3i']['label'] = 1 # can't have connection if no suggestion
-# results['qual']['label'] = results['q1']['label'] + (not results['q2i']['label']) + (not results['q3i']['label'])
+#Modify results to sum the QuAL score and to ignore Q3 if Q2 no suggestion
+if results['q2i']['label'] == 1:
+    results['q3i']['label'] = 1 # can't have connection if no suggestion
+results['qual']['label'] = results['q1']['label'] + (not results['q2i']['label']) + (not results['q3i']['label'])
 
 overview = NQDOverview(st, results)
-overview.draw()
+overview.draw()
+
+def rescale(x):
+    return (x + 1.0) / 2.0
+
+def get_explained_words(comment, pipe, label, cmap):
+    cls_explainer = SequenceClassificationExplainer(
+        pipe.model,
+        pipe.tokenizer)
+    word_attributions = cls_explainer(comment, class_name=label)[1:-1]  
+
+    # Get rid of "##"
+    to_disp = [
+        (word, '', f'rgba{tuple([int(c*255) for c in cmap.mpl_colormap(rescale(word_score))])}')
+        for word, word_score in word_attributions
+    ]
+    return to_disp
+
+qual_map = {
+    0: 'minimal',
+    1: 'very low',
+    2: 'low',
+    3: 'average',
+    4: 'above average',
+    5: 'excellent'
+}
+
+q1_map = {
+    0: "minimal",
+    1: "low",
+    2: "moderate",
+    3: "high"
+}
+
+q2i_map = {
+    0: "did",
+    1: "did not"
+}
+            
+with st.expander('Expand to explore further'):
+    st.write(f'Your comment was rated as a QuAL score of **{results["qual"]["label"]}**, indicating **{qual_map[results["qual"]["label"]]}** quality feedback.')
+    
+    st.markdown('### Level of Detail')
+    st.write(f"The model identified a **{q1_map[results['q1']['label']]}** level of detail in your comment.")
+    st.write("Below are words that pointed the model toward (green) or against (red) identifying a high level of detail:")
+    annotated_text(get_explained_words(st.session_state['comment'], models['q1'], 'LABEL_3', RdYlGn_10))
+
+    st.markdown('### Suggestion for Improvement')
+    st.write(f"The model **{q2i_map[results['q2i']['label']]}** predict that you provided a suggestion for improvement in your comment.")
+    st.write(f"Below are words that pointed the model toward (green) or against (red) identifying a suggestion for improvement:")
+    annotated_text(get_explained_words(st.session_state['comment'], models['q2i'], 'LABEL_0', RdYlGn_10))
+
+    if results['q2i']['label'] == 0:
+        st.markdown('### Suggestion Linking')
+        st.write(f"The model **{q2i_map[results['q3i']['label']]}** predict that you linked your suggestion")
+        st.write(f"Below are words that pointed the model toward (green) or against (red) identifying a linked suggestion:")
+        annotated_text(get_explained_words(st.session_state['comment'], models['q3i'], 'LABEL_0', RdYlGn_10))
+
+
+# annotated_text(to_disp)
+# cls_explainer = SequenceClassificationExplainer(
+#     models['q1'].model,
+#     models['q1'].tokenizer)
+# word_attributions = cls_explainer(st.session_state['comment'], class_name='LABEL_3')[1:-1]
+# to_disp = [
+#     (word, f'{word_score:.2f}', f'rgba{tuple([int(c*255) for c in Devon_10_r.mpl_colormap(word_score)])}')
+#     for word, word_score in word_attributions
+# ]
+# print(to_disp)
+# annotated_text(to_disp)

requirements.txt

@@ -3,9 +3,12 @@ torch
 torchvision 
 torchaudio
 transformers
+transformers-interpret
 plotly==5.11.0
 pandas
 spacy
 altair
 hydralit_components
-matplotlib
+matplotlib
+st-annotated-text
+palettable