TL;DR
This paper tackles the challenge of understanding what users are looking for when they search for geo-referenced information, going beyond just where. They propose a novel user interest model with three dimensions (Subjectivity, Search Target, and Restriction) to categorize these nuanced search intents, which can help improve the relevance of geo-located search results.
What Problem Does This Paper Address?
As mobile search becomes ubiquitous, understanding user intent in geo-referenced queries is crucial. While existing work often focuses on disambiguating location names or identifying simple spatial relationships (e.g., "restaurant near Eiffel Tower"), it falls short of capturing the specific aspects of a location users are interested in. For instance, a query like "jogging safely in New York" has a different information need than "restaurants in New York," even though both are informational and involve the same location. Current search systems struggle to differentiate these finer-grained intents, leading to potentially irrelevant results. This paper aims to bridge this gap by developing a model that can classify these diverse user interests associated with geo-referenced entities.
Methods
The authors propose a data-driven approach to build a user interest model. They collected a large corpus of questions from a Korean Q&A service (Naver "Knowledge-In") because these questions express user information needs more explicitly than typical web search queries.
- Aspect Identification: Three graduate students manually annotated these questions, identifying geographic entities and the aspects of information sought. They iteratively refined annotation tags to arrive at 16 distinct aspectual categories (e.g., Direction, Price, Quality, Landmark, Time, Location, Reason, Opinion, Event, Accessibility, Mood, List, Experience, Activity, Number of Participants, Product).
- Model Construction: These 16 aspects were then conceptualized into a three-dimensional user interest model:
- Subjectivity: Characterizes the degree of subjective information desired (Fact, Interpreted Fact (IF), Opinion).
- Search Target: Defines what the user wants to know about the location (Object, Activity, About there, Transportation).
- Restriction: Indicates whether the user is looking for information that might limit their access or decision-making (Restriction, No Restriction).
- Validation: The model's usefulness was validated using two datasets: the original Naver KIN data and a dataset from Yahoo! Answers. They measured coverage (how frequently categories are used) and inter-annotator agreement (how consistently annotators assign categories).
Datasets
- Naver "Knowledge-In" Service: 900 popular questions were initially collected, filtered down to 590 relevant geo-referenced questions. This dataset was used for aspect identification and model construction.
- Yahoo! Answers: 300 questions related to "NYC" were collected and annotated. This dataset was used for validation to assess the model's applicability across different platforms and potentially cultures.
Key Insights and Ideas
- Beyond "Where": The core insight is that geo-referenced searches are not just about finding a place, but about understanding specific aspects related to that place (e.g., its history, accessibility, associated activities, or even the mood it evokes).
- Q&A as a Rich Source: Leveraging Q&A platforms provides a more explicit and nuanced view of user information needs compared to short, ambiguous web search queries.
- Multi-dimensional Intent: User intent in geo-referenced search is complex and can be effectively captured by a multi-dimensional model rather than a single category. The proposed three dimensions (Subjectivity, Search Target, Restriction) offer a structured way to represent this complexity.
- Aspects are Not Mutually Exclusive: A single query or document can encompass multiple aspects, highlighting the need for flexible classification.
Key Results
- The analysis of the Naver KIN dataset identified 16 distinct aspectual categories, with "Landmark," "Time," "Quality," and "Price" being among the most frequent.
- The constructed three-dimensional model (Subjectivity, Search Target, Restriction) showed that categorical values under these dimensions were generally well-represented in both the Naver KIN and Yahoo! Answers datasets.
- Validation indicated that the "Search Target" dimension was particularly easy for annotators to identify consistently (high kappa coefficients), while "Subjectivity" and "Restriction" showed moderate agreement.
- Comparison between Naver KIN and Yahoo! Answers showed some cultural differences in search focus, with English queries on Yahoo! Answers showing a higher interest in "Restriction," "Activity," and "Opinion" compared to Naver KIN.
Limitations and Future Work
- Inter-Annotator Agreement: While "Search Target" showed good agreement, "Subjectivity" and "Restriction" had moderate agreement, suggesting these dimensions might require further refinement or clearer definitions for consistent annotation.
- Mobile Search Application: The current model was developed using Q&A questions. The authors acknowledge the need to apply and potentially adapt this model to shorter, more ambiguous mobile search queries, which may require user studies to understand context.
- Text Classifier Development: A significant future direction is to develop an automatic text classifier that can categorize web pages and queries based on this user interest model, which is noted as a challenging task.
- Model Adaptability: The authors suggest that the model might need to be adjusted as they explore its application to different contexts, particularly mobile search.
Summary updated