III: Small: Using Location for Retrieving Text and Images in News And Social Media Posts

点击翻译按钮获取中文摘要

The goal of this project is to enhance systems which make use of a map interface to access documents such as news and social media postings. Most geographic search systems require the user to enter the latitude and longitude of two corners of the bounding box that encloses the region of interest. Some allow locations to be specified by name, but the region associated with the name is fixed in advance and is often different from the region that the user intends to search. For example, a user might enter the location name "Los Angeles", hoping the search for social media posts from the city. However, the system might be designed to interpret "Los Angeles" to mean the greater Los Angeles area (too large) or the geographic point center of the city (too small). This project will create a system that overcomes these problems with text geographic queries. The underlying textual data is accessed via a map query interface using direct manipulation actions such as pan and zoom to navigate the data. The advantage of these actions is that the act of pointing at a location (e.g., by the appropriate positioning of a pointing device) and making the interpretation of the precision of this positioning specification dependent on the zoom level is equivalent to permitting spatial synonyms. This means that when searching for a "rock concert in Manhattan", responses such as Harlem, New York City, and Brooklyn are all acceptable when no concert is found in Harlem. The search for Harlem is expanded by using its spatial synonyms since Harlem is contained in Manhattan, New York City includes Manhattan, and Brooklyn is a sibling (neighbor) of Manhattan. The location-oriented map-based provides a new paradigm to reading and analyzing news and social media on a local level or global level. This project addresses challenges in: (1) Detecting tweets or other social media posts about local events. This is difficult as only a few people may be posting related tweets in contrast to global events where many people post tweets thereby making it easier to detect global events. (2) Improving the resolution of ambiguous location names when retrieving documents using textually-specified locations by developing more appropriate precision and recall evaluation metrics. (3) Enabling domain-specific tracking of mentions of events such as crimes and diseases in news and social media such as Twitter over time with the aid of heat maps which may have an impact on public safety and health. (4) Allowing users to specify the desired domain as well as infer it by use of exemplars. (5) Improving the method the NewsStand system, a prior contribution of the investigator, uses clustering by using word2vec which makes better use of semantics than the currently used TF-IDF. This clustering is used for the actual documents and their associated images and videos, and enables the detection of similar images based on semantics rather than local features such as color and texture.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.