highlighter.py

from lime.lime_text import LimeTextExplainer
from nltk.tokenize import sent_tokenize
from predictors import predict_for_explainanility


def explainer(text, model_type):
    def predictor_wrapper(text):
        return predict_for_explainanility(text=text, model_type=model_type)

    class_names = ["negative", "positive"]
    explainer_ = LimeTextExplainer(
        class_names=class_names, split_expression=sent_tokenize
    )
    sentences = [sent for sent in sent_tokenize(text)]
    num_sentences = len(sentences)
    exp = explainer_.explain_instance(
        text, predictor_wrapper, num_features=num_sentences, num_samples=500
    )
    weights_mapping = exp.as_map()[1]
    sentences_weights = {sentence: 0 for sentence in sentences}
    for idx, weight in weights_mapping:
        if 0 <= idx < len(sentences):
            sentences_weights[sentences[idx]] = weight
    print(sentences_weights, model_type)
    return sentences_weights, exp


def analyze_and_highlight(text, model_type):

    highlighted_text = ""
    sentences_weights, _ = explainer(text, model_type)
    positive_weights = [weight for weight in sentences_weights.values() if weight >= 0]
    negative_weights = [weight for weight in sentences_weights.values() if weight < 0]

    smoothing_factor = 0.001  # we do this cos to avoid all white colors
    min_positive_weight = min(positive_weights) if positive_weights else 0
    max_positive_weight = max(positive_weights) if positive_weights else 0
    min_negative_weight = min(negative_weights) if negative_weights else 0
    max_negative_weight = max(negative_weights) if negative_weights else 0

    max_positive_weight += smoothing_factor
    min_negative_weight -= smoothing_factor

    for sentence, weight in sentences_weights.items():
        sentence = sentence.strip()
        if not sentence:
            continue

        if weight >= 0 and max_positive_weight != min_positive_weight:
            normalized_weight = (weight - min_positive_weight + smoothing_factor) / (
                max_positive_weight - min_positive_weight
            )
            color = f"rgb(255, {int(255 * (1 - normalized_weight))}, {int(255 * (1 - normalized_weight))})"
        elif weight < 0 and min_negative_weight != max_negative_weight:
            normalized_weight = (weight - max_negative_weight - smoothing_factor) / (
                min_negative_weight - max_negative_weight
            )
            color = f"rgb({int(255 * (1 - normalized_weight))}, 255, {int(255 * (1 - normalized_weight))})"
        else:
            color = "rgb(255, 255, 255)"  # when no range

        highlighted_sentence = (
            f'<span style="background-color: {color}; color: black;">{sentence}</span> '
        )
        highlighted_text += highlighted_sentence

    if model_type == "bc":
        gradient_labels = ["HUMAN", "AI"]
    elif model_type == "quillbot":
        gradient_labels = ["ORIGINAL", "HUMANIZED"]
    else:
        raise ValueError(f"Invalid model type: {model_type}")

    highlighted_text = (
        "<div>"
        + highlighted_text
        + "<div style='margin-top: 20px; text-align: center;'>"
        + "<div style='position: relative; display: inline-block; width: 60%; height: 20px; background: linear-gradient(to right, #00FF00, #FFFFFF, #FF0000); font-family: \"Segoe UI\", Tahoma, Geneva, Verdana, sans-serif; font-size: 10px; font-weight: 600; color: #222; border-radius: 10px; box-shadow: 0px 2px 5px rgba(0, 0, 0, 0.1);'>"
        + f"<span style='position: absolute; left: 5px; top: 50%; transform: translateY(-50%); color: #000; font-weight: 600;'>{gradient_labels[0]}</span>"
        + f"<span style='position: absolute; right: 5px; top: 50%; transform: translateY(-50%); color: #000; font-weight: 600;'>{gradient_labels[1]}</span>"
        + "</div>"
        + "</div>"
        + "</div>"
    )

    return highlighted_text