Isw3.naist.jp

by automatically produced patterns by BACT – The aim of clinical trial is to investigate the efficacy and safety of the new treatment comparing with current – The US National Library of Medicine's bibliographic – The most popular information source for finding evidence – To what the new treatment shows greatness? – To construct the information extraction (IE) system from – To whom the new treatment shows greatness? • To construct the information extraction (IE) system from MEDLINE clinical trial abstracts We compared drug X and drug Y by investigating blood pressure and survival time in patients with high blood •TITLE: Peginterferon Alfa-2a plus ribavirin
versus interferon alfa-2a plus ribavirin for
chronic hepatitis C in HIV-coinfected persons.
(1) Compared Treatment
•BACKGROUND: Chronic hepatitis C virus peginterferon alfa-2a plus
(HCV) infection is a cause of major … interferon ribavirin
plus ribavirin for the treatment of chronic hepatitis interferon alfa-2a plus ribavirin
C in persons coinfected with HIV.
(2) Endpoint
randomly assigned to receive … either a virologic sustained virologic response
(3) Patient Population
Take notice that each clinical trail has sometimes more •RESULTS: Treatment with peginterferon and ribavirin was associated with a significantly persons coinfected with HIV
higher rate of sustained virologic response than
was treatment with interferon and ribavirin.
INPUT sentence: from a clinical trial MEDLINE abstract
“We conducted a multi-center, randomized trial comparing peginterferon plus ribavirin with interferon plus ribavirin for the treatment of chronic hepatitis C NP chunking: by YamCha (Kudo and Matsumoto, 2001)
“[We] conducted [a multi-center, randomized trial] comparing [peginterferon plus ribavirin] with [interferon plus ribavirin] for [the treatment] of [chronic hepatitis C] in [persons] co-infected with [HIV].” by automatically produced patterns by BACT NP tagging: manually by domain specific knowledge
“[We] conducted [STUDY] comparing [DRUG] with [DRUG] for [THERAPY] of [DISEASE] in [PATIENT] co-infected NP tagging: manually by domain specific knowledge
“[We] conducted [STUDY] comparing [DRUG] with [DRUG] for DISEASE: disease, symptom, virus
[THERAPY] of [DISEASE] in [PATIENT] co-infected with DRUG: drug, chemical compound interferon
STUDY: clinical trial
Sentence
Classification:
by BACT (a
THERAPY: treatment, regimen
Boosting Algorithm for
PATIENT: participants in the trial
Classification of Trees; Kudo
and Matsumoto, 2004)
TARGET: endpoints
SCHEDULE: time schedule of the trial
VALUE: value of TARGET
Information
Compared Treatment : “peginterferon plus Extraction:
NUMBER: numeral expression
Compared Treatment : “interferon plus ribavirin” Patient Population : “persons co-infected with NP tagging: manually by domain specific knowledge
“[We] conducted [STUDY] comparing [DRUG] with [DRUG] for
[THERAPY] of [DISEASE] in [PATIENT] co-infected with [DISEASE].”
Information Extraction: by manually written regular expressions
” [DRUG] “with” [DRUG]
[PATIENT]
with [DISEASE]
by automatically produced patterns by BACT Compared Treatment : “peginterferon plus ribavirin”
Compared Treatment : “interferon plus ribavirin”
Patient Population : “persons co-infected with HIV”
“[We] conducted [STUDY] comparing [DRUG] with [DRUG] for NP tagging: manually by domain specific knowledge
[THERAPY] of [DISEASE] in [PATIENT] co-infected with [DISEASE].” “[We] conducted [STUDY] comparing [DRUG] with [DRUG] for [THERAPY] of [DISEASE] in [PATIENT] co-infected with [DISEASE].” Sentence Classification
Matsumoto, 2004) , that automatically produce optimal patterns The accuracy of IE may improve if we apply by automatically produced patterns by BACT This table shows that not all NPs tagged by “DRUG” correspond to IE target of “Compared Treatment”.
We downloaded the 50 most recent abstracts of • Precision, Recall 5-fold cross validation clinical trials from the MEDLINE database: http://www.ncbi.nlm.nih.gov/entrez/query.fcgi "hepatitis"[MeSH Terms] AND hasabstract[text] AND To simplify the experiment, abstracts were selected • dep is outperformed by Ngram with respect to “Patient Population”. We can guess the reason here: parse errors occurred in many of the dependency trees caused by PP attachment ambiguity (“PATIENT with DISEASE”). – In this experiment, manually tagged by domain – In this experiment, only 50 abstracts.
by automatically produced patterns by BACT – How Sentence Classification contribute to the

Source: http://isw3.naist.jp/IS/21COE/pd-sk-seminar/2005/200503/200503_hara.pdf