Abstract
Traditional extractive summarization models primarily rely on the main body of a document to identify salient sentences. This paper introduces SIDENET, a neural framework designed to leverage "side information"—such as titles and image captions—to guide the extraction process. By incorporating an attention mechanism over this metadata, SIDENET significantly outperforms standard baselines in both informativeness and fluency, as demonstrated through automatic ROUGE metrics and human evaluation.
Introduction
Extractive summarization involves selecting a subset of sentences from a source document to form a concise summary. While human readers intuitively use contextual cues like titles and captions to grasp an article's core message, most automated systems ignore this metadata. The main body of an article can be dense and noisy, making the identification of salient points difficult without hierarchical understanding. Previous attempts to utilize side information relied on manual feature engineering. SIDENET addresses this by using a neural architecture that automatically learns the relationship between side information and sentence importance.
Methodology
The SIDENET architecture is a hierarchical encoder-decoder model comprising three primary components:
- Sentence Encoder: A Convolutional Neural Network (CNN) transforms sentences from both the main body and side information into continuous vectors. CNNs are particularly effective at identifying local patterns like named entities.
- Document Encoder: A Recurrent Neural Network (RNN) with Long Short-Term Memory (LSTM) cells processes sentence vectors in reverse order to build a global document representation.
- Sentence Extractor: An LSTM-based decoder labels sentences for extraction. Unlike standard models, it employs an attention mechanism over the side information (titles and captions) to calculate a "context vector" for each extraction decision.
The model was trained on an augmented version of the CNN news highlights dataset, which includes titles and image captions scraped from the original website. On average, each article had about three image captions, and 40% of the articles contained at least one caption.
Results and Discussion
Evaluation was conducted using ROUGE scores and human rankings. SIDENET consistently outperformed the LEAD baseline and POINTERNET, a state-of-the-art model without side information.
Key findings include:
- Side Information as a Filter: Titles and captions serve as a proxy for author intent, helping the model filter out non-essential information.
- Mitigating Lead Bias: While news summarization often suffers from "lead bias," SIDENET uses titles to identify key events that may appear later in the text.
- Ablation Study: As shown in the results table, combining titles and captions yielded the highest performance, confirming that different types of side information provide complementary signals.
Human evaluation further supported these findings. In a study where participants ranked summaries from 1st to 4th place, SIDENET was ranked 1st or 2nd significantly more often than the other automated systems. Humans found the SIDENET summaries to be more informative and closer to the manually written "gold" highlights.
Conclusion
SIDENET demonstrates the value of incorporating metadata into neural summarization. Future work may explore multi-modal side information, such as raw image features, and the application of this framework to domains where headings or user tags can serve as contextual anchors.
Summary updated