We first introduce the problem and examine several domains where it has been examined. We then address the two primary research questions that arise when comparing different methods: how best to model and evaluate timelines. In each of these two areas we organise the key papers thematically. The contributions of each paper are synthesised into a general discussion about the pros and cons of each approach.
Timeline generation (TLG) is a way of representing a large amount of temporally dependent information concisely. It is query driven; we retrieve a corpus of text linked to some entity, event or other term. The canonical TLG model clusters these articles into topics or stories, selects the most important of these clusters and returns timestamped summaries. It can be seen as a generalisation of the multi-document summarisation task, where we have introduced temporal dependency and structure.
Clustering: Our model handles both current and historical articles. Several domains where the TLG model is desirable (e.g. news) are characterised by a large historical catalog and high frequency. Our clustering model has to be scaleable and ideally functional in a streaming context.
Topic models: Incorporating a topic structure into our model gives us an intuitive and understandable document representation. Looking at the topic distribution also lets us prioritise certain kinds of sentences or stories.
Timelines: TLG differs from regular multi-document summarisation in its temporal dependence. As such, we value diversity in both the kind of content selected as well as when it was published.
Summarisation: While clustering and topic modeling are useful structuring tools, we must still compress the data for human readability.
Timeline generation has been applied in a range of domains. Several approaches focus on specific sub-problems in TLG. Althoff et al. timelines entities in the Freebase repository (Althoff 2015). This knowledge base is comprised of subject and objects linked by predicates; 'Robert Downey Junior' might be linked by 'Starred In' to 'The Avengers. The TLG model seeks to group and select a subset of these subject-predicates that summarise our query entity. Working from knowledge-base instead of a corpus of text provides useful structure. We are spared the task of preprocessing, entity linking, consolidation and data validation on our text corpora. Of course, this relies on the existence of a knowledge base of timestamped facts about our query entity. This is an issue if we want to be able to perform queries on lesser known entities. A similar approach was used by Ahmed et al. on a corpus of academic papers (Ahmed 2012).
Single-document variations of the TLG model have been applied to Wikipedia and Wikinews articles (Bauer 2015, Minard 2015). The single-document nature of the task is characterised by a focus on the summarisation and selection tasks over clustering. Timelines took the form of timestamped subsets of the document sentences. A similar approach has been applied to Twitter feeds; Hong et al. applied a single-document version of the TLG model to a user's tweet-history (Hong 2011).
The most frequent application of TLG models is to a corpus of news articles (Chieu 2004, Hong 2011, Yan 2011, Yan 2011a, Allan 2001, Swan 2000, Ahmed 2011). A document retrieval service is given a query term; a person, event, company or generic term. The body of documents retrieved form the input for the TLG task. The task then is to group and summarise a potentially large body of articles and provide a compact representation of the query term.