TL;DR
This paper introduces INFOS, an intelligent news filtering system that combines a simple keyword-based "Global Hill Climbing" (GHC) method with a more sophisticated "Case-Based Reasoning" (CBR) approach. By integrating these techniques, INFOS aims to effectively combat information overload on the internet by learning user preferences and filtering out irrelevant news articles with improved accuracy.
What Problem Does This Paper Address?
The internet, particularly Usenet newsgroups, has become a massive source of information, leading to an "information overload" for users. Manually sifting through hundreds of daily articles to find relevant content is time-consuming and inefficient. Existing filtering methods often struggle with the inherent inconsistency of user interests and the nuances of language. This paper tackles the challenge of building an intelligent agent that can automatically learn a user's preferences and accurately filter news articles, thereby reducing the user's search burden.
Methods
The INFOS system employs a hybrid approach, combining two main techniques:
- Global Hill Climbing (GHC): This is a simple, keyword-based classification scheme. It builds a user model by learning from user feedback (accepting or rejecting articles). The system maintains a table of features (words, authors, collaborative data) and their associated "accepted" and "rejected" counts. A new article is classified based on the aggregated weights of its features. This method is computationally efficient and easy for users to understand and modify.
- Case-Based Reasoning (CBR): To address the limitations of GHC, particularly its assumption of word independence and difficulty with polysemy (words with multiple meanings), INFOS incorporates CBR. This approach uses a knowledge-based method derived from WordNet, a lexical database. Instead of raw keywords, INFOS maps words to concepts using WordNet's sense definitions. It then builds a memory hierarchy of "cases" (articles) indexed by these concepts. When a new article arrives, INFOS searches this hierarchy for similar cases and uses their classification to predict the new article's relevance. This allows for a deeper understanding of article content and concept comparison.
The paper also discusses Index Extraction, which involves using WordNet to identify relevant noun and verb senses. It employs Paice's algorithm to identify frequently occurring words within a "topic neighborhood" of sentences to extract relevant concepts.
Datasets
The primary dataset used for experiments consisted of 144 sequentially posted news articles from the ucd.life newsgroup. This newsgroup was chosen because it covers a variety of topics likely to be of interest to undergraduate students and receives a moderate number of messages daily, making filtering applicable.
Key Insights and Ideas
- Hybridization is Key: The paper strongly suggests that combining simple, scalable methods (like GHC) with more sophisticated, knowledge-rich approaches (like CBR) yields better results than either method alone.
- User Model Simplicity Matters: For a filtering system to be adopted, its user model must be easy to understand and modify. GHC excels here, but CBR offers deeper understanding.
- Word Sense Disambiguation is Crucial: The paper highlights that words having multiple meanings (polysemy) pose a significant challenge for filtering systems. INFOS attempts to address this by using WordNet to map words to concepts and by using techniques like Paice's algorithm to identify relevant senses within a topic neighborhood.
- Collaborative Filtering Adds Value: Incorporating feedback from other users (collaborative data) in the user model can significantly improve filtering accuracy.
- Feature Weighting is Important: Not all features are equally predictive. Experiments showed that subject-based features were more predictive than text body or author features, suggesting the need for differential weighting in classification.
Key Results
- GHC alone: Achieved a correct classification rate of 51.5% with a low error rate of 7.3%.
- CBR alone: Performed worse than GHC, with a correct classification rate of 39.8% and a higher error rate of 9.5%. This is attributed to potential irrelevant cases being retrieved due to word sense disambiguation issues.
- Hybrid GHC + CBR: The combined approach yielded the best correct classification rate at 58.0%, although it had a slightly higher error rate (12%) compared to GHC alone. This indicates the potential of the hybrid method, but also points to the need for more robust word sense disambiguation.
Limitations and Future Work
- Word Sense Disambiguation: The paper acknowledges that the current word sense disambiguation techniques are not perfect and can lead to irrelevant cases being retrieved in the CBR method. This is identified as a key area for improvement.
- Scalability of CBR: While GHC is computationally efficient, the depth-first search in the memory hierarchy for CBR might become computationally expensive with a very large number of cases.
- Future Work: The authors propose several directions:
- Incorporating genetic algorithms to explore the news space and disambiguate input text more accurately.
- Developing offline filtering capabilities.
- Creating a graphical user interface.
- Implementing self-modifying parameters.
- Expanding knowledge bases.
- Applying the system to the World Wide Web and intelligent tutoring systems.
Summary updated