Abstract
Phoebus is a system designed to transform unstructured, ungrammatical online classifieds into structured, searchable databases. By combining record linkage against a trusted reference set with targeted information extraction, it standardizes messy listings (e.g., "93 civic") into clean, integrated data points (e.g., "1993 Honda Civic"). This approach enables advanced search, sorting, and seamless integration with external data sources.
Introduction: The Challenge of Unstructured Data
Online marketplaces like Craigslist and eBay are rich data sources that are notoriously difficult to search and organize. Sellers often write descriptions without regard for structure or standard naming conventions. A buyer might search for "Honda" but miss a listing for a "93 civic" because the seller omitted the make. Misspellings (e.g., "Civc") and the use of sub-models without the primary model name (e.g., "Del Sol" instead of "Civic Del Sol") further complicate keyword-based searching.
Traditional Natural Language Processing (NLP) often fails in these contexts because posts lack proper grammar and sentence structure. Standard data cleaning techniques typically assume data is already in a field-based format. Phoebus addresses this gap by extracting and standardizing meaningful attributes from raw, "noisy" text to enable better search, sorting, and integration with external data sources.
Methodology
The Phoebus system employs a two-step approach leveraging "reference sets"—external collections of known entities, such as the Edmunds car buying guide.
- Record Linkage: Instead of parsing text in isolation, Phoebus first attempts to match a post to the most likely member of a reference set. For example, a post stating "93 civic" is linked to the "1993 Honda Civic" entry. This process automatically fills in missing information (like the make "Honda") and corrects misspellings.
- Information Extraction: Once a match is established, the system performs a second pass to extract attributes not found in the reference set, such as price or posting date. Anchoring the extraction to a known entity allows the system to accurately distinguish between similar-looking data (e.g., identifying "$50" as a price rather than a model number).
Finally, the system uses these standardized attributes to dynamically generate queries for external sites, such as MSN Autos, facilitating seamless data integration.
Datasets
The demonstration system utilizes the following data:
- Input Data: Real-world used car listings scraped from Craigslist across six cities, including Los Angeles and Boston.
- Reference Set: The Edmunds car buying guide, providing "ground truth" for car makes, models, trims, and years.
- Integration Target: MSN Autos, used to demonstrate how extracted data can link to external quality ratings and reviews.
Key Insights
- Reference Sets as Anchors: The core innovation is that understanding ungrammatical text is unnecessary if it can be mapped to a known, clean entity. The reference set provides a scaffold for messy data.
- Standardization Enables Integration: Extraction allows unstructured text to be treated as a database. Once "93 civic" is standardized to "1993 Honda Civic," the system can programmatically build URLs to link the post to other car-related databases.
- Beyond Keywords: Semantic annotation solves the "recall" problem in search. A user searching for "Honda" can find posts that never explicitly mention the word "Honda" through the record linkage step.
Results
The Phoebus system successfully demonstrates:
- Handling Noise: Correct identification and categorization of posts with misspellings ("Civc") and missing attributes.
- Dynamic UI: A web interface allowing users to search and sort Craigslist posts by attributes (Make, Model, Trim, Year, Price) that were never explicitly structured by sellers.
- Real-time Integration: Successful construction of functional URLs for MSN Autos at query time, proving extracted attributes are accurate enough for external production systems.
Limitations and Future Work
The system's efficacy is tied to the quality and coverage of the reference set; rare or custom items may result in incorrect matches. Future work includes adapting the system to other domains (e.g., electronics) where reference sets may be less comprehensive and exploring "deep integration" to scrape and display external data directly within the Phoebus interface.
Summary updated