topic-classification.html

<!DOCTYPE html>
<html lang="en">
<head>
    <meta charset="UTF-8">
    <meta name="viewport" content="width=device-width, initial-scale=1.0">
    <!--==================== UNICONS ====================-->
    <link rel="stylesheet"
        href="https://unicons.iconscout.com/release/v4.0.0/css/line.css">
    <!--==================== SWIPER CSS ====================-->
    <link rel="stylesheet" href="assets/css/swiper-bundle.min.css">
    <title>Topic Classification</title>
    <!--==================== CSS ====================-->
    <link rel="stylesheet" href="assets/css/tc.css">
    <link rel="stylesheet" href="assets/css/styles.css">

</head>
<body>
    
    <h1>Topic Classification</h1>
    <div>
        We first train a classifier on the existing labeled dataset. This gives us a base model. We use this base model to generate predictions of a portion of the unlabeled data. These predictions are also called pseudo-labels. We stack the pseudo labels and existing labeled data to generate a bigger dataset and train an improved classifier. This classifier is then used to further generate predictions and the process repeats till we get an acceptable result. 
        We have implemented this approach in 2 ways:
    </div>
</br>
    <div class="container">
        <div>
            <div class="method_title">
                Approach 1: Graph Neural Networks
            </div>
            <div class="method_desc">
                This method can be best explained using the following flowchart:
                <img src="/assets/img/topic_class/approach1.png" alt="">
            </div>
            <div class="method_scores">
                The following graphs describe the loss and F1 score achieved by this approach:
                <img src="/assets/img/topic_class/loss1.png" alt="">
                <img src="/assets/img/topic_class/f1score1.png" alt="">
            </div>
            <div class="method_test">
                Test
            </div>
        </div>
        <div>
            <div class="method_title">
                Approach 2: BERT embeddings
            </div>
            <div class="method_desc">
                This method can be best explained using the following flowchart:
                <img src="/assets/img/topic_class/approach2.png" alt="">
    
            </div>
            <div class="method_scores">
                The following graphs describe the accuracy and F1 score achieved by this approach:
                <img src="/assets/img/topic_class/loss2.png" alt="">
                <img src="/assets/img/topic_class/accuracy2.png" alt="">
            </div>
            <div  class="method_test">
                Test
            </div>
        </div>
    </div>

</body>
</html>