Deep neural models to represent news events

Abstract

The thesis is dedicated to the background linking tasks for news articles, utilizing the deep neural network models. The goal is to retrieve similar articles based on the news story currently viewed. We examined neural and non-neural representations for raw text and discussed notions of similarity a good model should identify and retrieve. We covered various deep neural network models and highlighted their advantages and disadvantages. Inspired by deep neural architectures in the area of Information Retrieval we adjusted the Deep Semantic Similarity model to the background linking task. Our refactored DSSM architecture employs a convolutional neural network with multiple filters and regularization techniques. This convolutional network acts as an auto-encoder and learns the compressed representations of news articles and news stories. Cosine similarity is used as the proximity metric to retrieve related news articles. Experimental results prove that our adjusted DSSM model is applicable for the background linking task, and overperforms the baseline SVM model. We discovered that corpora distributions affect the performance of the model. A model trained on news corpus containing mostly political and social news will perform poorly on news corpus about sport and entertainment news. Grid search and hyperparameter tuning are also important. Deep neural network architectures are powerful tools which can be used to solve complicated tasks and approximate nearly any function. Having a good quality dataset is half of the success. The DSSM model is planned to be adjusted to various news corpora and applied to different tasks; such as automatic linking of news articles to Wikipedia pages and linking news articles to news events. We assume this model can be extended to learn representations of a sequence of events for the task of linking background events.