Methods and systems are presented for automatically detecting positions of various webpage elements within a webpage when the webpage is rendered, based on analyzing the programming code of the webpage using graph-based and NLP-based techniques. A position detection system obtains the programming code of the webpage and generates a DOM tree based on the programming code. A group of leaf tags may be identified based on traversing the DOM tree. The position detection system may use two different, but coordinated, prediction models to make the classification prediction for the leaf tags, where the first prediction model focuses on recall by identifying as many positive leaf tags as possible at the cost of low precision and the second prediction model refines the results from the first prediction model by improving the precision of the classification.