Model - Multiple Vectorizers

Imports

import os
import pandas as pd
from iqual import iqualnlp, evaluation, crossval, vectorizers

Load annotated (human-coded) and unannotated datasets

data_dir         = "../../data"
human_coded_df   = pd.read_csv(os.path.join(data_dir,"annotated.csv"))
uncoded_df       = pd.read_csv(os.path.join(data_dir,"unannotated.csv"))

Load spacy or sentence-transformers using precomputed dictionaries

dict_dir        = "../../dictionaries"

sentence_transformer_models = ["all-mpnet-base-v2", "all-roberta-large-v1","distiluse-base-multilingual-cased-v2"]
spacy_models                = ["en_core_web_sm","en_core_web_md","en_core_web_lg"]
model_paths  = [os.path.join(dict_dir,m+'.pkl') for m in [*sentence_transformer_models,*spacy_models]]

Split the data into training and test sets

from sklearn.model_selection import train_test_split
train_df, test_df = train_test_split(human_coded_df,test_size=0.25)
print(f"Train Size: {len(train_df)}\nTest Size: {len(test_df)}")

Train Size: 7470
Test Size: 2490

Configure training data

### Select Question and Answer Columns
question_col = 'Q_en'
answer_col   = 'A_en'

### Select a code
code_variable = 'marriage'

### Create X and y
X = train_df[[question_col,answer_col]]
y = train_df[code_variable]

Initiate model

# Step 1: Initiate the model class
iqual_model = iqualnlp.Model()

# Step 2: Add layers to the model
#  Add text columns, and choose a feature extraction model (Available options: scikit-learn, spacy, sentence-transformers, saved-dictionary (picklized dictionary))
iqual_model.add_text_features(question_col,answer_col,model='../dictionaries/all-mpnet-base-v2.pkl',env='saved-dictionary')

# Step 3: Add a feature transforming layer (optional)
# A. Choose a feature-scaler. Available options: 
# any scikit-learn scaler from `sklearn.preprocessing`
### iqual_model.add_feature_transformer(name='StandardScaler', transformation="FeatureScaler")
# OR
# B. Choose a dimensionality reduction model. Available options:
# - Any scikit-learn dimensionality reduction model from `sklearn.decomposition`
# - Uniform Manifold Approximation and Projection (UMAP) using umap.UMAP (https://umap-learn.readthedocs.io/en/latest/)

### iqual_model.add_feature_transformer(name='PCA', transformation="DimensionalityReduction")

# Step 4: Add a classifier layer
# Choose a primary classifier model  (Available options: any scikit-learn classifier)
iqual_model.add_classifier(name="LogisticRegression")

# Step 5: Add a threshold layer. This is optional, but recommended for binary classification
iqual_model.add_threshold(scoring_metric='f1')

# Step 6: Compile the model
iqual_model.compile()

Pipeline(steps=[('Input',
                 FeatureUnion(transformer_list=[('question',
                                                 Pipeline(steps=[('selector',
                                                                  FunctionTransformer(func=<function column_selector at 0x0000023B71C74280>,
                                                                                      kw_args={'column_name': 'Q_en'})),
                                                                 ('vectorizer',
                                                                  Vectorizer(env='saved-dictionary',
                                                                             model='../dictionaries/all-mpnet-base-v2.pkl'))])),
                                                ('answer',
                                                 Pipeline(steps=[('selector',
                                                                  FunctionTrans...
                                                                             model='../dictionaries/all-mpnet-base-v2.pkl'))]))])),
                ('Classifier',
                 Classifier(C=1.0, class_weight=None, dual=False,
                            fit_intercept=True, intercept_scaling=1,
                            l1_ratio=None, max_iter=100,
                            model='LogisticRegression', multi_class='auto',
                            n_jobs=None, penalty='l2', random_state=None,
                            solver='lbfgs', tol=0.0001, verbose=0,
                            warm_start=False)),
                ('Threshold', BinaryThresholder())])

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

model_paths = [os.path.join(dict_dir,model+'.pkl') for model in [*spacy_models,*sentence_transformer_models]]

# Saved-dictionary (precomputed vectors using spacy/sentence-transformers)
params_saved = {
    "Input":{
        "question":{
            "vectorizer":{
                        "model":model_paths,
                        "env":["saved-dictionary"],               
                         },
        },
        "answer":{
            "vectorizer":{
                        "model":model_paths,
                        "env":["saved-dictionary"],                
                         },                        
        },
    },
        
    "Classifier":{
            "model":["LogisticRegression"],
            "C":[0.01,0.1],
        },
}

# Scikit-learn vectors (TfidfVectorizer/CountVectorizer)

params_sklearn = {
    "Input":{
        "question":{
            "vectorizer":{
                        "model":['TfidfVectorizer','CountVectorizer'],
                        "max_features":[500,1000,1500,2500,],
                        "env":["scikit-learn"],               
                         },
        },
        "answer":{
            "vectorizer":{
                        "model":['TfidfVectorizer','CountVectorizer'],
                        "max_features":[1500,2500,4000,],
                        "env":["scikit-learn"],                
                         },                        
        },
    },
        
    "Classifier":{
            "model":["LogisticRegression"],
            "C":[0.01,0.1],
        },
}

CV_SEARCH_PARAMS = [
    crossval.convert_nested_params(params_saved),
    crossval.convert_nested_params(params_sklearn)
]
print(CV_SEARCH_PARAMS)

[{'Input__question__vectorizer__model': ['../../dictionaries\\en_core_web_sm.pkl', '../../dictionaries\\en_core_web_md.pkl', '../../dictionaries\\en_core_web_lg.pkl', '../../dictionaries\\all-mpnet-base-v2.pkl', '../../dictionaries\\all-roberta-large-v1.pkl', '../../dictionaries\\distiluse-base-multilingual-cased-v2.pkl'], 'Input__question__vectorizer__env': ['saved-dictionary'], 'Input__answer__vectorizer__model': ['../../dictionaries\\en_core_web_sm.pkl', '../../dictionaries\\en_core_web_md.pkl', '../../dictionaries\\en_core_web_lg.pkl', '../../dictionaries\\all-mpnet-base-v2.pkl', '../../dictionaries\\all-roberta-large-v1.pkl', '../../dictionaries\\distiluse-base-multilingual-cased-v2.pkl'], 'Input__answer__vectorizer__env': ['saved-dictionary'], 'Classifier__model': ['LogisticRegression'], 'Classifier__C': [0.01, 0.1]}, {'Input__question__vectorizer__model': ['TfidfVectorizer', 'CountVectorizer'], 'Input__question__vectorizer__max_features': [500, 1000, 1500, 2500], 'Input__question__vectorizer__env': ['scikit-learn'], 'Input__answer__vectorizer__model': ['TfidfVectorizer', 'CountVectorizer'], 'Input__answer__vectorizer__max_features': [1500, 2500, 4000], 'Input__answer__vectorizer__env': ['scikit-learn'], 'Classifier__model': ['LogisticRegression'], 'Classifier__C': [0.01, 0.1]}]

Configure a Hyperparameter Grid for cross-validation + fitting

Model training:

Cross-validate over hyperparameters and select the best model

# Scoring Dict for evaluation
scoring_dict = {'f1':evaluation.get_scorer('f1')}

cv_dict = iqual_model.cross_validate_fit(
    X,y,                                # X: Pandas DataFrame of features, y: Pandas Series of labels
    search_parameters=CV_SEARCH_PARAMS, # search_parameters: Dictionary of parameters to use for cross-validation
    cv_method='RandomizedSearchCV',     # cv_method: Cross-validation method to use, options: GridSearchCV, RandomizedSearchCV
    n_iter=50,                          # n_iter: Number of iterations for RandomizedSearchCV
    scoring=scoring_dict,               # scoring: Scoring metric to use for cross-validation    
    refit='f1',                         # refit: Metric to use for refitting the model
    n_jobs=-1,                          # n_jobs: Number of parallel threads to use  
    cv_splits=3,                        # cv_splits: Number of cross-validation splits
)
print()
print("Average F1 score: {:.3f}".format(cv_dict['avg_test_score']))

.......168 hyperparameters configurations possible.....
Average F1 score: 0.835

Check best model

params = iqual_model.get_params()

print(f"Q vectorizer: {params.get('Input__question__vectorizer__model','Dropped')}")
print(f"A vectorizer: {params.get('Input__answer__vectorizer__model')}")
print(f"Classifier: {params.get('Classifier__model')}")
print(f"Threshold for classification: {params.get('Threshold__threshold'):.3f}")

iqual_model.model

Q vectorizer: ../../dictionaries\distiluse-base-multilingual-cased-v2.pkl
A vectorizer: ../../dictionaries\distiluse-base-multilingual-cased-v2.pkl
Classifier: LogisticRegression
Threshold for classification: 0.192

Pipeline(steps=[('Input',
                 FeatureUnion(transformer_list=[('question',
                                                 Pipeline(steps=[('selector',
                                                                  FunctionTransformer(func=<function column_selector at 0x0000023B71C74280>,
                                                                                      kw_args={'column_name': 'Q_en'})),
                                                                 ('vectorizer',
                                                                  Vectorizer(env='saved-dictionary',
                                                                             model='../../dictionaries\\distiluse-base-multilingual-cased-v2.pkl'))])),
                                                ('answer',
                                                 Pipeline(steps=[('...
                 Classifier(C=0.1, class_weight=None, dual=False,
                            fit_intercept=True, intercept_scaling=1,
                            l1_ratio=None, max_iter=100,
                            model='LogisticRegression', multi_class='auto',
                            n_jobs=None, penalty='l2', random_state=None,
                            solver='lbfgs', tol=0.0001, verbose=0,
                            warm_start=False)),
                ('Threshold',
                 BinaryThresholder(threshold=0.1919019469710422,
                                   threshold_range=(0.004475209108360682,
                                                    0.8112250807781638)))])

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

Evaluate model using out sample data (Held out human-coded data)

test_X = test_df[['Q_en','A_en']]
test_y = test_df[code_variable]

f1_score  = iqual_model.score(test_X,test_y,
                              scoring_function=evaluation.get_metric('f1_score'))
print(f"Out-sample F1-score: {f1_score:.3f}")

accuracy  = iqual_model.score(test_X,test_y,
                              scoring_function=evaluation.get_metric('accuracy_score'))
print(f"Out-sample accuracy-score: {accuracy:.3f}")

Out-sample F1-score: 0.795
Out-sample accuracy-score: 0.981

Predict labels for unannotated data

uncoded_df[code_variable+'_pred'] = iqual_model.predict(uncoded_df[['Q_en','A_en']])

uncoded_df[code_variable+"_pred"].hist(figsize=(3,3),bins=3)

<AxesSubplot:>

Examples for positive predictions

for idx, row in uncoded_df.loc[(uncoded_df[code_variable+"_pred"]==1),['Q_en','A_en']].sample(3).iterrows():
    print("Q: ",row['Q_en'],"\n","A: ", row['A_en'],sep='')
    print()

Q: Well, currently studying in the sixth standard. How long will you teach?
A: If it is God's command, I will marry now. If there is no order to stay until next year, then I will teach until next year.

Q: What are your hopes for your children?
A: Now I am tensed, my daughter has grown up, I am not able to get married. Son wants to read, but can't read. School is not open, nowhere is open, can't teach. Can't teach Arabic, Mugh (Burmese).

Q: You said it. What career will you be happy after completing your studies?
A: I will be happy if I can get married after my studies.