Automating Readers’ Advisory to Make BookRecommendations for K-12 Readers
by Alicia Wood
Problem
• Existing book recommenders failed to offer adequate choices for K-12 readers
• Important to provide good reading material among K-12 students
• Not easy to find the right books for the right audience
Who cares?
• Parents – 32% of American 4th graders proficient in reading
• Children
Previous Work
• Previous book recommenders • Extract features, opinion, feature/opinion pairs
– Bootstrapping, NLP, ML, extraction rules, latent semantic analysis, statistical analysis, and information retrieval
• Information extraction approaches on product reviews
• Amazon• Require historical data• Require an ontology• Don’t consider readability level of users
Proposed Solution
• Rabbit• Multi-dimensional approach• No feedback from users• ABET (Appeal-based extraction tool)
Readers’ Advisory
• Offers materials of potential interest with “the help of knowledgeable and non-judgmental library staff”
• Based on:– Reasons behind preferences– Topical areas– Content descriptions– Appeal factors (pacing, description of characters,
tone, etc.)
Appeal Factors & Terms
ABET
• Extracts appeal-term descriptions of books from reviews available
• Imperative to properly associate appeal terms and appeal factors– pairs can be correctly extracted to generate
accurate appeal-term description for the book
Extraction Rules for ABET
Example
• SA = “The narrative of the book is dramatic”– Subject: narrative
• SB = “He creates believable characters”– Object: character (AF)
• If subject/object is an appeal factor, the word semantically linked to that subject/object is often an appeal term
• Rules 1 + 2
Example
• “The characters are not simple– Rule 4 – negation
ABET
• Creates the appeal-term description for a book applying the rules
• Frequency of occurrence– degree of significance
Rabbit
1. Analyze profile of a reader2. Determine readability level 3. Select books4. Compute ranking score
Candidate Books
• CB – candidate book available at a book repository Rep• PB each book in R’s profile• |P| - # of books in R’s profile• TRoLL(CB), TRoLL(PB) – grade level of CB/PB
determined by TRoLL
Topical Similarity Measure
• CB – vector of weights of CB if subject heading is of CB
• P – vector of weights of Pi (proportion between number of books in P that have been assigned Pi)
Content Similarity
• Enhanced version of cosine similarity
• CB = vector of Wcb1…Wcbn • P = vector of Wp1…Wpn• Wpi, Wcbi = weights of keywords Pi and Cb
Appeal Term Similarity
• F = set of appeal factors in appeal term descriptions
• CBf and Pf = n dimensional vector representation of appeal term distribution of an appeal factor (f)
Ranking Candidate Books
• Multiple linear regression
• Train using Ordinary Least Squares method– T set dataset
Experimental Results
1. Compute precision and recall of appeal factor-appeal term pairs extracted from book reviews
2. Analyze correctness of appeal-term descriptions created by ABET
3. Compare appeal-term descriptions generated by ABET with respect to ones extracted from Novelist
Experimental Results – 1
• 100 books • Manually annotated and compared with ABET• Precision: 0.85• Recall: 0.82• F-measure: 0.83• High accuracy for ABET in generating appeal
factor-appeal term pairs
Experimental Results - 2
• Surveys to determine correctness• Overall 94% accuracy
Experimental Results - 3
• Surveys to determine comparison• Appeal-term descriptions provided by ABET
favored over Novelist
Validation
• Computed empirical studies• Assessed performance of Rabbit using Eset in
terms of Normalized Discounted Cumulative Gain
• Rabbit locates more relevant books• Rabbit outperforms GoodReads and Novelist