Generative and Discriminative Models in Text Classification David D. Lewis Independent Consultant Chicago, IL, USA Workshop on Challenges in Information Retrieval and Language Modeling, U Mass, CIIR, Amherst, MA, 11 Sept 2002
Text Classification Given a document, decide which of several classes it belongs to: –TREC filtering –TDT tracking task –Text categorization! Automated indexing, content filtering, alerting,... More LM papers here than any other IR problem –Others: parts of IE, author identification,...
Lang. Models are Generative Model predicts probability document d will be generated by a source c e.g. Unigram language model: Parameters, i.e. P(w|c)’s, are fit to optimally predict generation of d
Classify Text w/ Gen. Model One source model for each class c Choose class c with largest value of: For 2 classes, unigram P(d|c), we have: aka Naive Bayes (NB), Roberston/KSJ
The Discriminative Alternative Directly model probability of generating class conditional on words: P(c|w) Logistic regression: Tune parameters to optimize conditional likelihood (class probability predictions)
LR & NB: Same Parameters!
Observations LR & NB have same parameterization for 2- or k-class, binary or raw TF weighting LR outperforms NB in text categorization and batch filtering studies –NB optimizes parameters to predict words, LR optimizes to predict class
False Hopes for LM? Leveraging unlabeled data (e.g. EM)? –Initial results show only small impact (same story as syntactic class tagging) Non-unigram models –More accurately predict the wrong thing? Cross-lingual TC –Any more than MT followed by TC?
True LM Hopes 1: Small Data? Number training examples to reach maximum effectiveness (Ng & Jordan ‘01): –NB: O(log # features) –LR: O(# features) LR and NB not compared yet (?) in low data (TREC adaptive, TDT tracking) case –Priors/smoothing likely to prove critical
True LM Hopes 2: Facets? MeSH category assignments: Anti-Inflammatory Agents, Non- Steroidal/*therapeutic use Tumor Necrosis Factor/antagonists & inhibitors/immunology Most combinations have zero training data Berger & Lafferty MT approach?
Non-LM TC Challenges? Integration of prior knowledge Choosing documents to label (TREC adaptive, active learning, sampling) Combining text and nontext predictors Knowing how well a classifier will/can do Evolving category systems, switching vocabularies