Understanding User Intents in Online Health Forums Thomas Zhang, Jason H.D. Cho, Chengxiang Zhai Department of Computer Science University of Illinois at Urbana-Champaign 5th ACM Conference on Bioinformatics, Computational Biology, and Health Informatics Newport Beach, California 22nd September 2014 1 Online Health Forums • Purpose: To provide a convenient platform to facilitate discussion among patients and professionals • Huge user base, and still growing! • In 2011, 80% of all web users searched for health information online, of which 6% participated in health related discussions • Forums contain valuable information – Contain rich, often first hand experiences 2 Deficiencies of Forums • Threads are scattered • Similar questions are asked again and again • Keyword search is inadequate – Finding several keyword matches in a thread does not necessarily mean that the thread is relevant 3 Post about cholinergic urticaria in April 2004 Received 3rd and final reply a week later Post from March 2012 No replies as of July 2014 4 Applications of Intents • Improving thread retrieval – e.g. A thread whose original post matches both keywords and intent specified by the user are more likely to be helpful • Filtering threads – e.g. To treat a condition, only look at posts asking about treatment • Understanding user behavior in forums – i.e. users of different forums have different intents 5 This Paper • Introduces problem of identifying user intents in health forums as a classification problem • Derives the first taxonomy of user intents • Designs a set of novel features for use with machine learning to solve the problem • Create the first dataset for evaluation, and conducted experiments to make empirical findings 6 Roadmap 1. Problem formulation 2. Intent taxonomy derivation 3. Methodology – – – Support vector machines Hierarchical classification Feature design 4. Evaluation – – – Dataset Experiments Results 5. Intents in MedHelp forums 6. Wrap-up 7 Problem Formulation Given π, an original thread post from our dataset π· with intent ππ from a taxonomy of user intents πΆ = π1 , … , ππ . Denote π = {π 1 , … , π π } as the sentence representation of π. Classify π as some ππ ∈ πΆ using π as evidence. π is correctly classified if and only if π = π 8 Taxonomy Derivation • No taxonomy exists for health forum intents • Solution: Create our own! • First reduce top ten most commonly asked generic questions by doctors (Ely et al, 2000) into three intent classes – Classes match the intents of users who search for health information online (Choudhury et al, 2014) • Next introduce two additional intent classes that are specific to health forum posts 9 Taxonomy • Manage: How should I manage or treat condition X? • Cause: What is the cause of symptom/physical/test finding X? • Adverse: Can drug or treatment X cause adverse finding Y? • Combo: Combination (at least two of first three) • Story: Story telling, news, sharing or asking about experience, soliciting support, or others 10 Where are we? 1. Problem formulation 2. Intent taxonomy derivation 3. Methodology – – – Support vector machines Feature Selection Hierarchical classification 4. Evaluation – – – Dataset Experiments Results 5. Intents in MedHelp forums 6. Wrap-up 11 Support Vector Machines (SVM) • Main idea: Learn a hyperplane from examples to separate them into two classes • Use learned hyperplane to classify unseen examples • Capable of non-linear and multiclass classification • Shown to have good performance on high dimensional data 12 Post Representation • How should we represent posts? – SVMs require examples to be represented as a vector of features • What are features? – Some measurable property of the observed data • How should we select them? 13 Feature Selection A good feature should be: 1. Generic enough to be found in many posts 2. Sufficiently discriminative for different intents 14 Solution: Patterns! • Sequence of (possibly non-contiguous) tokens that represent recurring text patterns in sentences • Very generic – Lowercasing, stemming – POS tagging – UMLS semantic group tagging • Very discriminative – “What could X be…?” signifies Cause intent, but “What does X do…?” signifies Manage intent 15 Pattern Types Each pattern falls under one of four types: • LSP: Lowercased + stemmed tokens only – E.g. “…what can caus…” • POSP: LSP + POS tags – E.g. “…how to <VERB>…” • SGP: LSP + semantic group tags – E.g. “…if <CHEM> works…” • ALL: All types of tokens and tags – E.g. “…<CHEM> make <PRP> feel…” 16 UMLS Semantic Groups • MetaMap labels text phrases with semantic group labels from the UMLS Metathesaurus 17 Caveat • Patterns possess limitations – Difficult to achieve good coverage without sacrificing discriminative properties – Impossible to extract for posts with large content variations (e.g. Story posts) • However, we still want complete coverage of our dataset! 18 Solution: Hierarchical Classification! • Two cascading SVM classifiers – The first uses binary pattern features (Pattern SVM) – The second uses unigram features with TF-IDF weighting (Word SVM) • Complete coverage allows comparison with unigram baseline Input Post Match ≥ 1 pattern? Yes No Pattern SVM Word SVM Output Class 19 Where are we? 1. Problem formulation 2. Intent taxonomy derivation 3. Methodology – – – Support vector machines Hierarchical classification Feature design 4. Evaluation – – – Dataset Experiments Results 5. Intents in MedHelp forums 6. Wrap-up 20 Dataset • No labeled dataset exists, since this is a new problem • So we create our own! – 1,192 original HealthBoards posts, evenly divided among four topics: allergies, breast cancer, depression, and heart disease • Ideally want more posts, but labeling is expensive • Why the four topics? 21 Dataset Labeling • Labeling done by two CS students – Substantial* agreement with medical students (π = 0.67) – Substantial* agreement between themselves (π = 0.665, 74.67% labels match) • Combo posts labeled by a third CS student according to their underlying classes – A Combo post is predicted correctly if a classifier outputs one of its class labels *Per Landis and Koch, 1977 22 Experiments • What is the best performing set of patterns? – Try different type combinations of patterns • How does hierarchical compare with baseline? – Five-fold cross validation (CV) • Does performance suffer if we train on posts from three topics and test on the fourth? – Four-fold forum CV 23 Selecting a Pattern Set πΆππ. πΆππ. 2ππ π= ,π = , πΉ1 = πππ‘. π + πΆ + |π΄| π+π 24 CV Takeaways • Overall Patterns reach labeling agreement upper bound Patternsimprovement give high precision butforum low recall generalize well across topics is underwhelming, why? – Why is this acceptable? Hierarchical Classification Performance Word Classifier (Baseline) Performance 25 Intents in MedHelp Forums We applied our Pattern SVM to 61,225 MedHelp posts split across allergies, breast cancer, depression, and heart disease 26 Concluding Remarks • Introduced the new problem of forum post intent analysis • Designed the first taxonomy and dataset for classification • Proposed a novel set of pattern features for SVMs • Proved that patterns give high classification precision while generalizing well across forums 27 Future Work • Administer study of health forum user intents • Expand pattern feature set to improve recall • Handle classification of Story posts • Identify all intents from Combo posts • Further evaluation with larger datasets 28 Thank you! Questions? Comments? 29