NTCIR Temporal Information Access (Temporalia) Task

Comparison with other tasks

Q: What is the difference between TREC Temporal Summarization Task (TempSum) with Temporalia?

A: TREC Temporal Summarization task (http://www.trec-ts.org/) is composed of two subtasks: Sequential Update Summarization and Value Tracking. The task of Sequential Update Summarization is to find timely, sentence-level, reliable, relevant and non-redundant updates about developing event. Value tracking aims at tracking values of event-related attributes that are of high importance to the event. Examples are the number of fatalities or financial impact. Both subtasks have clear temporal character since the updates have to be timely and relevant. Their outcome is then related to the subtask of ranking documents for the "recency" class of queries (Temporal Information Retrieval subtask). However, in the case of TempSum the scope is limited to the information about a concrete past event or to a particular type of attribute-like information such as a numerical value, etc. Also, returned information needs to be short and non-redundant.

Temporal Information Retrieval subtask in Temporalia, on the other hand, can take any category of temporal query as input. The results are in the form of ranked list of documents for which neither redundancy nor text length play any special role.

Q: What is the relation of TREC KBA to Temporalia?

A: TREC Knowledge Base Acceleration (KBA) ( http://trec-kba.org/trec-kba-2013.shtml ) is a task for filtering a large stream of text to find documents that can help update knowledge bases like Wikipedia, FaceBook, Crunchbase. It is composed of two subtasks: Cumulative Citation Recommendation and Streaming Slot Filling. The former is about filtering documents that are worth citing in a profile of selected entities (profile could be a Wikipedia page of the entity). It does not put any requirement on temporality neither novelty of the content. On the other hand, Streaming Slot Filling tracks the attributes and relations of a selected entity over time. As the organizers state, it virtually “allows the target entity to "move" as accumulated content implies modifications to the attributes, relations, and free text associated with” an entity. Thus, again, like in the case of TempSum, the temporal information need is in the form of the recency requirement put on documents related to a particular entity such as a person or organization.

Q: Are there any similarities of Temporalia with GeoCLEF task?

A: Space and time are commonly treated as orthogonal or parallel dimensions. GeoCLEF 2008 task (http://www.uni-hildesheim.de/geoclef/) required to categorize queries into geographic queries and non geographic queries. Similarly, in Temporalia the query categorization task consists of deciding whether query is temporal or non-temporal. GeoCLEF also looked deeper into the subparts of geographic queries differentiating typical three parts: "where", "geo-relation" and "what" components. The "what" component is of a map type (e.g., river, beach, mountain, monuments), yellow page type (e.g., businesses or organizations, like hotels, restaurants, hospitals) or information type. In Temporalia we propose deeper, hierarchy of query categorization proposing diverse query types such as recency, past-related or future-related queries. We then set up the second task consisting of ranking documents for diverse temporal query types.

Q: Then how about Geotemporal Information Retrieval GeoCLEF task (GEOTIME)?

A: GeoTime task ( http://metadata.berkeley.edu/NTCIR-GeoTime/) aims at answering mixed geo-temporal information needs represented by questions such as "When and where did George Kennan die?" or "When and where were the last three Winter Olympics held?". The temporal answer is in the form of date or period/interval type variable. In Temporalia we focus on more diverse types of temporal search needs (e.g., ones for which freshness and timeliness of answers play key role and so on).