Automated natural language headline generation using discriminative machine learning models

Headline or short summary generation is an important problem in Text Summarization and has several practical applications. We present a discriminative learning framework and a rich feature set for the headline generation task. Secondly, we present a novel Bleu measure based scheme for evaluation of headline generation models, which does not require human produced references. We achieve this by building a test corpus using the Google news service. We propose two stacked log-linear models for both headline word selection (Content Selection) and for ordering words into a grammatical and coherent headline (Headline Synthesis). For decoding a beam search algorithm is used that combines the two log-linear models to produce a list of k-best human readable headlines from a news story. Systematic training and experimental results on the Google-news test dataset demonstrate the success and effectiveness of our approach.