SMBM Talks NLP for Biomedical Text Mining SMBM

SMBM Talks NLP for Biomedical Text Mining SMBM, Cambridge, April 11 -13 (Edinburgh May 2)

Resources and Tools for Biomedical Text Mining Junichi Tsujii (U of Tokyo) Keywords: GENIA corpus; annotation Main point: progress in text mining depends on the integration of growing GENIA annotation (coreference, eg) with lexical resources for domain knowledge (ontologies) and software development. Take home message: see main point above

• annotated corpus • POS • NER • coreference (670 abstracts, Singapore) • interaction (biological events; cooperation with CNRS) • parse trees (1. 5 million GENIA abstracts parsed in 10 days using a 100 PC cluster) • ontology • top nodes: substance; source; other • software development • POS tagger • NER tagger • parser • IR system (Medusa) • IE (event extraction: relation gene/disease) system

• POS tagger • Max. Ent model (Kazama and Tsujii 2003, 2005) • Trained on WSJ (>39, 000 sent. ) and GENIA (18, 500 sent. ) train test WSJ GENIA WSJ+GENIA WSJ 97. 0 75. 2 96. 9 GENIA 84. 3 98. 1 • NER tagger • combines a rule-based and statistical approach • on Bio. NLP: 70. 8% (? ) -- our system got 70. 1%

• HPSG-based parser (Enju) • • • see Miyao et al. ACL 05 available on website XML output dependency relations predicate-argument accuracy: • PTB: prec=88. 3% rec=87. 2 • GENIA: lower. . . • gene/disease relation extraction • pred/arg works better than bag of words or local context (gives best precision)

Recognising noun phrases in biomedical text: an evaluation of lab prototypes and a commercial chunker J. Wermter, J. Fluck, J. Stroetgen, S. Geissler, U. Hahn (U. Jena, Temis) Keywords: chunking, portability Main point: take several existing chunkers trained on (or developed for) newspaper text and evaluate their performance on biomedical data (beta version of GENIA syntactic annotation). Take home messages: • overall performance drop (~3 -6 points) for ML systems when shifting to bio domain • no significant difference between statistical and rule-based systems

Three statistical chunkers: • Yam. Cha (support vector machine) • Tbl (transformation-based error-driven learning) • Bo. SS (boundaries predictor by combining observed probabilities of NP boundaries and POS patterns in trainset) One rule-based commercial system • Temis 1. Uses words rather than GENIA POS tags 2. Computes morphological information (Xe. LDA toolkit) 3. HMM POS tagger disambiguates chain of POS tags • hand-coded grammar had to be modified (on PTB) • tagset had to be translated (not straightforward)

Training and Test Sets Train • sections 15 -18 of Penn Treebank for training (over 200, 000 POS-tagged tokens and IOB-chunked) Test • GENIA treebank (beta version) (200 Med. Line abstracts with syntactic annotation) the GENIA treebank was automatically converted into the IOB format • just under 45, 000 tokens • ~11, 000 = devtest for settting Temis’ IOB output • ~34, 000 = actual test set

Results and Errors PTB Corpus Rec Yam. Cha Prec GENIA Corpus F Rec Prec F 89. 00 89. 30 89. 15 90. 10 90. 01 86. 46 86. 84 86. 65 91. 80 92. 03 86. 31 85. 49 85. 90 86. 94 Temis 94. 22 92. 27 Tbl 94. 15 89. 92 Bo. SS 94. 29 86. 61 87. 14 85. 34 86. 23 After domain adaptations Temis Bo. SS 91. 24 90. 59 90. 91 87. 25 89. 19 88. 21 Errors • Coordination • bracketed elements • . . .

Automatic Term List Generation for Entity Tagging Ted Sandler, Andrew Schein, and Lyle Ungar (CS, UPenn) Keywords: NER, automatic gazetteer creation Main point: term lists can be obtained automatically, and when integrated in a NER (gene)tagger (CRF) boost its performance to a level comparable with hand-modelled lists Take home messages: • unsupervised gazetteer creation is feasible and useful • supervised methods for obtaining terms outperform unsupervised methods

Overall Approach • choose set of vocabulary items (nouns) to partition into classes • choose set of useful syntactic relations • frequent • informative • relatively noise-free • parse corpus to extract relations and collect statistics • use clustering algorithm to partition the vocabulary • resulting partitions are term lists 4 related methods for generating term lists; they differ wrt: (see table) • word representation • clustering algorithms to partition the words • choice of feature weighting

Corpus • 15, 000 sentences from Bio. Creative + 1, 800, 547 Medline abs • parsed using Minipar; vocabulary=7782 single token nouns Representation of the base vocabulary • vector space where each item is represented by set of syn configurations it occurs in • affinity matrix where each item is represented as its similarities to other items in the vocabulary Weighting Schemes • Pearson’s chi-square test • Generalized Likelihood Ratio (G-square; Dunning 1993; better with sparse data) • first better at “common sense” generalisations; second better at domain-specific generalisations Clustering Algorithms • kmeans clustering for words in vector space (high recall) • agglomerative clustering for data in affinity matrix (high prec)

NER (Gene) Tagging • Mc. Donald and Pereira’s CRF tagger • automatically generated 2, 164 overlapping term lists incorporated as features in the model • binary feature (0/1) for each term list (in=1; not=0) • baseline tagger without lists • tagger augmented with hand-compiled lists of genes (57, 563) • tagger augmented with large list of genes obtained via supervised learning (Tanabe and Wilbur Gene. Lexicon: 1, 145, 913) TRAIN/TEST: 5 -fold Xvalidation on 394, 661 words of Bio. Creative (1/5 for training and 4/5 for testing) Baseline Unsupervised Supervised Manual prec 0. 698 0. 705 0. 709 0. 716 rec 0. 613 0. 622 0. 621 0. 631 f-score 0. 653 0. 661 0. 662 0. 671

Protein-Protein Interaction Extraction: A Supervised Learning Approach J. Xiao, J. Su, G. Zhou, C. Tan (Inst. For Infocomm Research, Singapore) Keywords: relation extraction Main point: a Max. Ent approach to protein-protein relation extraction that exploits simple local features performs better than co-occurrence and rule-based approaches, achieving nearly 94% recall and 88% precision on 303 Med. Line abstracts. Take home message: • supervised learning with shallow features work well for protein interaction extraction

Task: extract couple of interacting proteins • no direction • perfect NER (manual annotation) Procedure • • • tokenisation and morphological analysis POS tagging NER sentence analysis (parsing) coreference resolution (including abbreviations and aliases) Max. Ent classifier

Features • Words • all words that appear in two protein names • words in between two protein names • previous/next words in a n-words window (unordered) • Overlap • number of protein names in between 2 protein names • Keywords • occurrence of word from keyword list in surroundings • Chunks • all heads of base phrases in between 2 protein names • all heads surrounding the protein name pair • all phrase types between 2 protein names • Parse Tree • Dependency Tree • dependency between two proteins • Pair of heads of protein names • Pair of abbreviations of two proteins

Experiment and Results • corpus: IEPA (Iowa University) • 303 Medline abstracts • 633 positive instances • 1080 negative instances • POS tagger trained on GENIA using an HMM model • Collin’s parser • 10 -fold Xvalidation • best result: rec=93. 9%; prec=88%; f=90. 9 GOOD Features - words (esp. surrounding) - chunks - pairs of protein heads - pairs of abbreviations - keywords (so and so) NOTSOGOOD Features - overlap - parse trees - dependency relations

Challenges of Information Mining in a Pharmaceutical Environment Philippe Sanseau (Glaxo-Smith-Kline, UK) Main point: Q: How do you see the role of NLP in your field? A: Excuse me, could someone explain what NLP is, please. Take home question: are NLP and pharmaceutical communities on the same track?