Beyond Genes, Proteins, and Abstracts: A Framework to Capture Scientific Claims Catherine Blake School of Information and Library Science University of North Carolina at Chapel Hill http://www.ils.unc.edu/~cablake cablake@email.unc.edu Motivation • Relentless increase in electronically available text – Life Sciences • The NLM added the 17 millionth entry to PubMed in April 2007 • 5,200 journals indexed • 12,000 new articles each week ! – Chemistry – more than 110,000 articles in 1 year alone • Consequences: – Hundreds of thousands of relevant articles – Implicit connections between literature go unnoticed Shift from Retrieval to Synthesis 2 Entity Extraction • Newspaper genre – People, places, and organizations – Message Understanding Conference (MUC) • Biomedical genre – – – – Genes and proteins Diseases and treatments Chemical compounds Challenges: BioCreative , GENIA, JNLPBA 3 Relationship Extraction • Newspaper genre – Person moving from one company to another • Biomedicine genre – – – – – genes and proteins e.g. binds, inhibits ARBITER (Rindflesch, Rajan, & Hunter, 2000) Geneways (Rzhetsky, et al, 2004) relEx (Fundel, Kuffner, & Zimmer, 2007) GENIA www-tsujii.is.s.u-tokyo.ac.jp/GENIA 4 Causal Relationships • Newspaper genre – Causal relationships (Khoo, Chan, & Niu, 1998) • Biomedical genre – Causes and treats (Price & Delcambre, 2005) – Causal knowledge (Khoo, Chan, Niu, 2000) • Universal Grammar – Causatives (Comrie, 1974, 1981) – Action verbs (Thomson, 1987) 5 Claim Definition • “To assert in the face of possible contradiction” • Example sentence reporting a claim – “This study showed that Tamoxifen reduces the breast cancer risk” • Example Claim Framework – Tamoxifenagent – reduceschange – [breast cancer risk] object 6 Goals • Create a Framework that reflects how claims made in biomedical literature • The Framework should – generalize beyond biomedicine – differentiate between different levels of confidence in the claim – consider claims made in the full text • Populate the Framework automatically 7 The Claim Framework • Information facets – concepts – change – basis of the claim • Each information facet may have – modifiers – directionality 8 The Claim Framework 1. Explicit Claim Agent Object Nature of change Required 2. Implicit Claim Agent Object Optional Optional Required Required N/A Required Required Required Required Required Required Optional Required Optional Category 3. Correlation 4. Comparison 5. Observation Concept A Concept B Claim Basis Optional 9 Explicit Claims Indeed, glycine prevented Wy-14643-stimulated superoxide production by Kupffer cells. Claim 1 – glycineagent – preventedchange – [Wy-14643-stimulated superoxide production]object Claim 2 – [Kupffer cells]agent – produceschange – [Wy-14643-stimulated superoxide]object. 10 Implicit Claims In liver the number of peroxisomes increases from about 500-600/cell to > 5000/cell after exposure to peroxisome proliferators. Claim 1 – – – – [Peroxisomes proliferators] agent increaseschangeDirection Peroxisomesobject [In the liver]agentModifier – [number]agentModifier 11 Correlations A weak but statistically significant correlation was observed between the plasma nm23-H1 level and the WBC count (Figure 1, n=102, r=0.437, P<0.0001) – – – – [plasma nm23-H1 level] agent [WBC count] object correlation change [statistically significant] changeModifier 12 Comparisons The plasma concentration of nm23-H1 was higher in patients with AML than in normal controls (P = .0001) Claim 1 – – – – [plasma concentration of nm23-H1] basis of claim [Patients with AML]agent higher changeDirection [normal controls]object 13 Observations However, the plasma nm21-H1 protein level was increased in SML-M3 patients (P=.0002) Claim 1 – [nm21-H1 protein level]object – IncreasedchangeDirection – [SML-M3 patients]objectModifier 14 Working Hypothesis 1 The Claim Framework reflects how a scientist communicates her findings – Full text documents randomly selected from biomedical literature will report findings using constructs within the Claim Framework – Human annotators will agree on facets within the Claim Framework – The Claim Framework will generalize to a variety of scientific literatures 15 Working Hypothesis 2 Facets within the Claim Framework can be populated automatically – The system will detect all claims identified by the human annotators (i.e. recall) – The system will only identify claims that were identified by the human annotators (i.e. precision) – The system design will generalize to new literatures by avoiding domain specific constructs 16 Validating the Claim Framework • Draft Claim Framework given to two annotators • Pilot Study: Identify every claim – Include claims that don’t conform to the framework – Don’t consider how this will be automated 17 Validating the Claim Framework • Main study – 25 articles • Verification – Random set of sentences annotated twice – Feedback provided daily 18 Results • All documents – – – – – Total number of sentences: 5535 Sentences with >=1 claim: 1250 (22.6%) Total number of claims: 3228 Average claims per sentence: 2.51 Claims that did not fit in the Framework: 31 • Per document – Average number of sentences: 191 – Average number of sentences with >=1 claim:43 19 Distribution of Claim Categories Category Total (%) Pilot(%) Explicit 2489 Implicit 87 2.70 3 0.75 84 2.98 Observation 298 9.23 24 6.03 274 9.73 Correlation 174 5.39 12 3.02 162 5.75 Comparison 165 5.11 27 6.85 138 4.9 100 398 100 2830 100 Total 3228 77.11 332 83.42 Main(%) 2157 76.63 20 Annotation Agent Agent Direction Agent Modifier Object Object Direction Object Modifier Change Change Direction Change Modifier Claim Basis Claim Basis Dir. Claim Basis Mod. Total All Documents Total (%) Words (Avg) 2894 89.65 5221 1.80 285 8.83 291 1.02 1246 38.60 4448 3.57 3197 99.04 6849 2.14 271 8.40 283 1.04 1561 48.36 5383 3.44 1897 58.77 1953 1.03 1337 41.42 1358 1.02 1147 35.53 1618 1.41 165 5.11 394 2.39 42 1.30 43 1.02 86 2.66 266 3.09 21 3228 28107 8.70 Inter Annotator Agreement Information Facet Agent Object Change Change+ChangeDir Kappa 0.71 0.77 0.57 0.88 Agreement substantial substantial moderate almost perfect 22 Location of Claims Section Abstract Introduction Method Result Discussion Total Total Sentences With % % Claim Total section claim 98 309 31.72 7.84 357 979 36.47 28.56 6 1121 0.54 0.48 293 1829 16.02 23.44 539 1406 38.34 43.12 1250 5535 22.58 100.00 23 Findings thus far • 99% of the claims made in these articles could be captured in the Claim Framework • 22% of sentences report at least 1 claim • 77% of the claims identified were explicit • 8% of claims are made in the abstract • Agreement – substantial between agents and objects – almost perfect for change and change direction 24 Acknowledgements – This project supported in part by – Renaissance Computing Institute (RENCI) Faculty Fellowship Program – NSF Center for Environmentally Responsible Solvents and Processes (CERSP CHE-9876674) – This project used resources provided by – the OSG, which is supported by the NSF & the U.S. Department of Energy's Office of Science • The speaker thanks • Nassib Nassar and Mats Rynge (RENCI) • Amol Bapat and Ryan Jones (SILS) Questions and Comments Welcome Catherine Blake cablake@email.unc.edu http://www.ils.unc.edu/~cablake School of Information and Library Science University of North Carolina at Chapel Hill