Zusammenfassung
Online content providers process massive streams of texts to supply topics and entities of interest to their customers. In this process, they face several information overload problems. Apart from identifying topically relevant articles, this includes identifying duplicates as well as filtering summary articles that comprise of disparate topical sections. Such summary articles would be treated as ...
Zusammenfassung
Online content providers process massive streams of texts to supply topics and entities of interest to their customers. In this process, they face several information overload problems. Apart from identifying topically relevant articles, this includes identifying duplicates as well as filtering summary articles that comprise of disparate topical sections. Such summary articles would be treated as noise from a media monitoring perspective, an end user might however be interested in just those articles. In this paper, we introduce the recognition of summary articles as a novel task and present theoretical and experimental work towards addressing the problem. Rather than treating this as a single-step binary classification task, we propose a framework to tackle it as a two-step approach of boundary detection followed by classification. Boundary detection is achieved with a bi-directional LSTM sequence learner. Structural features are then extracted using the boundaries and clusters devised with the output of this LSTM. A range of classifiers are applied for ensuing summary recognition including a convolutional neural network (CNN) where we treat articles as 1-dimensional structural `images'. A corpus of natural summary articles is collected for evaluation using the Signal 1M news dataset. To assess the generalisation properties of our framework, we also investigate its performance on synthetic summaries. We show that our structural features sustain their performance on generalisation in comparison to baseline bag-of-words and word2vec classifiers.
Altmetric