Automatic Title Generator for news articles.
This work studies the task of automatically generating titles for a given text. A summarization approach is taken to develop a method that utilize TF-IDF scores, sentence extraction and parse trees. The method is evaluated on the Harvard Dataverse News Articles corpus, containing 3824 news articles. Four judges were presented 20 generated titles and asked to rate them based on grammar, conciseness, relevancy and overall fit. The results show that the presented method can generate perfect titles with a success rate of 20%. A new approach for finding candidate sentences utilizing word similarity is also explored and evaluated and is shown to outperform the baseline approach.