added scatter plot

.gitignore

@@ -0,0 +1,2 @@
+tmp.ipynb
+__pycache__

.streamlit/config.toml

@@ -0,0 +1,6 @@
+[theme]
+primaryColor="#4361ee"
+backgroundColor="#FFFFFF"
+secondaryBackgroundColor="#F0F2F6"
+textColor="#262730"
+font="sans serif"

app.py

@@ -1,10 +1,13 @@
 import os
 import torch
+import dash
 import streamlit as st
+import pandas as pd
 from transformers import AutoTokenizer, AutoModelForSequenceClassification
 from transformers import pipeline
 from plotly.subplots import make_subplots
 import plotly.graph_objects as go
+import plotly.express as px
 
 
 def z_score(y, mean=.04853076, sd=.9409466):
@@ -42,103 +45,192 @@ def indicator_plot(value, title, value_range, domain):
 
     return plot
 
-body = """
+def scatter_plot(df, group_var):
+
+    colors = ['#36def1', '#4361ee'] if group_var else ['#4361ee']
+
+    plot = px.scatter(
+        df, 
+        x='Machine-ratings', 
+        y='Human-ratings',
+        color=group_var,
+        facet_col='x_group', 
+        facet_col_wrap=2,
+        trendline='ols',
+        trendline_scope='trace',
+        hover_data={
+            'Text': df.text,
+            'Language': False,
+            'x_group': False,            
+            'Human-ratings': ':.2f',
+            'Machine-ratings': ':.2f',
+            'Study': df.study,
+            'Instrument': df.instrument,
+        },
+        width=400,
+        height=400,
+        color_discrete_sequence=colors
+    )
+    
+    plot.for_each_annotation(lambda a: a.update(text=a.text.split('=')[-1]))
+    plot.update_layout(
+        legend={
+            'orientation':'h',
+            'yanchor': 'bottom',
+            'y': -.30
+        })
+    plot.update_xaxes(title_standoff = 0)
+
+    return plot
+
+# data import and wrangling
+covariate_columns = {
+        'content_domain': 'Content Domain',
+        'language': 'Language',
+        'rater_group': 'Rater Group',
+    }
+
+df = (
+    pd
+    .read_feather(path='data.feather').query('partition == "test" | partition == "dev"')
+    .melt(
+        value_vars=['sentiment_model', 'desirability_model'],
+        var_name='x_group',
+        value_name='x',
+        id_vars=['mean_z', 'text', 'content_domain', 'language', 'rater_group', 'study', 'instrument']
+        )
+    .replace(
+        to_replace={
+            'en': 'English',
+            'de': 'German',
+            'other': 'Other',
+            'personality': 'Personality',
+            'laypeople': 'Laypeople',
+            'students': 'Students',
+            'sentiment_model': 'Sentiment Model',
+            'desirability_model': 'Desirability Model'
+        }    
+    )
+    .rename(columns=covariate_columns)
+    .rename(
+        columns={
+            'mean_z': 'Human-ratings',
+            'x': 'Machine-ratings',
+        }
+    )
+)
+
+st.markdown("""
     # NLP for Item Desirability Ratings
-    This web application accompanies the paper *Leveraging Natural Language Processing for Item Desirability Ratings: 
-    A Machine-Based Alternative to Human Judges* submitted to the Journal *Personality and Individual Differences*.
+    This web application accompanies the paper "*Expanding the Methodological Toolbox: Machine-Based Item Desirability Ratings as an Alternative to Human-Based Ratings*".
 
     ## What is this research about?
     Researchers use personality scales to measure people's traits and behaviors, but biases can affect the accuracy of these scales. 
     Socially desirable responding is a common bias that can skew results. To overcome this, researchers gather item desirability ratings, e.g., to ensure that questions are neutral. 
     Recently, advancements in natural language processing have made it possible to use machines to estimate social desirability ratings, 
-    which can provide a viable alternative to human ratings and help researchers, scale developers, and practitioners improve the accuracy of personality scales.
+    which can provide a viable alternative to human ratings and help researchers, scale developers, and practitioners improve the accuracy of personality scales.  
+""")
 
+# demo
+st.markdown("""
     ## Try it yourself!
-    Use the text field below to enter a statement that might be part of a psychological questionnaire (e.g., "I love a good fight.").
-    The left dial will indicate how socially desirable it might be to endorse this item. 
-    The right dial indicates sentiment (i.e., valence) as estimated by regular sentiment analysis (using the `cardiffnlp/twitter-xlm-roberta-base-sentiment` model).    
-"""
-
-st.markdown(body)
-
-input_text = st.text_input(
-    label='Estimate item desirability:',
-    value='I love a good fight.',
-    placeholder='Enter item'
-)
-
-# desirability model
-# remote or local?
-if os.environ.get("item-desirability"):
-    model_path = 'magnolia-psychometrics/item-desirability'
-else:
-    model_path = '/nlp/nlp/models/finetuned/twitter-xlm-roberta-base-regressive-desirability-ft-4'   
-
-auth_token = os.environ.get("item-desirability") or True
+    Use the text field below to enter a statement that might be part of a psychological questionnaire (e.g., "I love a good fight.").    
+    The left dial indicates how socially desirable it might be to endorse this item. 
+    The right dial indicates sentiment (i.e., valence) as estimated by regular sentiment analysis (using the `cardiffnlp/twitter-xlm-roberta-base-sentiment` model).
+""")
 
-if 'tokenizer' not in globals():
-    tokenizer = AutoTokenizer.from_pretrained(
-        pretrained_model_name_or_path=model_path,
-        use_fast=True,
-        use_auth_token=auth_token
-    )
-
-if 'model' not in globals():
-    model = AutoModelForSequenceClassification.from_pretrained(
-        pretrained_model_name_or_path=model_path, 
-        num_labels=1, 
-        ignore_mismatched_sizes=True,
-        use_auth_token=auth_token
-    )
-
-# sentiment classifier
-if 'classifier' not in globals():
-    sentiment_model = 'cardiffnlp/twitter-xlm-roberta-base-sentiment'
-    classifier = pipeline("sentiment-analysis", model=sentiment_model, tokenizer=sentiment_model, use_fast=False, top_k=3)
-
-classifier_output = classifier(input_text)
-classifier_output_dict = {x['label']: x['score'] for x in classifier_output[0]}
-classifier_score = classifier_output_dict['positive'] - classifier_output_dict['negative']
-
-if input_text:    
+## desirability model
+with st.spinner('Processing...'):
     
-    inputs = tokenizer(input_text, padding=True, return_tensors='pt')
-
-    with torch.no_grad():
-        score = model(**inputs).logits.squeeze().tolist()
-        z = z_score(score)
-
-    p1 = indicator_plot(
-        value=z,
-        title=f"Item Desirability",
-        value_range=[-4, 4],
-        domain={'x': [0, .45], 'y': [0, 1]},    
+    if os.environ.get('item-desirability'):
+        model_path = 'magnolia-psychometrics/item-desirability'
+    else:
+        model_path = '/nlp/nlp/models/finetuned/twitter-xlm-roberta-base-regressive-desirability-ft-4'   
+
+    auth_token = os.environ.get('item-desirability') or True
+
+    if 'tokenizer' not in globals():
+        tokenizer = AutoTokenizer.from_pretrained(
+            pretrained_model_name_or_path=model_path,
+            use_fast=True,
+            use_auth_token=auth_token
+        )
+
+    if 'model' not in globals():
+        model = AutoModelForSequenceClassification.from_pretrained(
+            pretrained_model_name_or_path=model_path, 
+            num_labels=1, 
+            ignore_mismatched_sizes=True,
+            use_auth_token=auth_token
+        )
+
+    ## sentiment model
+    if 'classifier' not in globals():
+        sentiment_model = 'cardiffnlp/twitter-xlm-roberta-base-sentiment'
+        classifier = pipeline("sentiment-analysis", model=sentiment_model, tokenizer=sentiment_model, use_fast=False, top_k=3)
+
+    input_text = st.text_input(
+        label='Estimate item desirability:',
+        value='I love a good fight.',
+        placeholder='Enter item text'
     )
 
-    p2 = indicator_plot(
-        value=classifier_score,
-        title=f"Item Sentiment",
-        value_range=[-1, 1],
-        domain={'x': [.55, 1], 'y': [0, 1]}
+    if input_text:    
         
-    )
-
-    fig = go.Figure()    
-    fig.add_trace(p1)
-    fig.add_trace(p2)
-
-    fig.update_layout(
-        title=dict(text=f'"{input_text}"', font=dict(size=36),yref='paper'),
-        paper_bgcolor = "white", 
-        font = {'color': "black", 'family': "Arial"})
-        
-    st.plotly_chart(fig, theme=None, use_container_width=True)
-    
-
-    notes = """
-    Item desirability: z-transformed values, 0 indicated "neutral".
+        classifier_output = classifier(input_text)
+        classifier_output_dict = {x['label']: x['score'] for x in classifier_output[0]}
+        classifier_score = classifier_output_dict['positive'] - classifier_output_dict['negative']
+
+        inputs = tokenizer(input_text, padding=True, return_tensors='pt')
+
+        with torch.no_grad():
+            score = model(**inputs).logits.squeeze().tolist()
+            z = z_score(score)
+
+        p1 = indicator_plot(
+            value=classifier_score,
+            title=f'Item Sentiment',
+            value_range=[-1, 1],
+            domain={'x': [.55, 1], 'y': [0, 1]}        
+        )
+
+        p2 = indicator_plot(
+            value=z,
+            title=f'Item Desirability',
+            value_range=[-4, 4],
+            domain={'x': [0, .45], 'y': [0, 1]},    
+        )
+
+        fig = go.Figure()
+        fig.add_trace(p1)
+        fig.add_trace(p2)
+
+        fig.update_layout(
+            title=dict(text=f'"{input_text}"', font=dict(size=36),yref='paper'),
+            paper_bgcolor = "white", 
+            font = {'color': "black", 'family': "Arial"})
+            
+        st.plotly_chart(fig, theme=None, use_container_width=True)
+                
+        st.markdown("""
+            Item sentiment: Absolute differences between positive and negative sentiment.
+            Item desirability: z-transformed values, 0 indicated "neutral".        
+        """)
+
+## plot
+st.markdown("""
+    ## Explore the data
+    Figures show the accuarcy in precitions of human-rated item desirability by the sentiment model (left) and the desirability model (right), using `test`-partition data only.
+""")
+
+
+show_covariates = st.checkbox('Show covariates', value=True)
+
+if show_covariates:
+    option = st.selectbox('Group by', options=list(covariate_columns.values()))
+else:
+    option = None
 
-    Item sentiment: Absolute differences between positive and negative sentiment.
-    """
+plot = scatter_plot(df, option)
 
-    st.markdown(notes)
+st.plotly_chart(plot, theme=None, use_container_width=True)

requirements.txt

@@ -1,4 +1,5 @@
 torch
 transformers
 plotly
-sentencepiece
+dash
+statsmodels