Learning what to read: Focused machine reading

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2905–2910 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Learning what to read: Focused machine reading Enrique Noriega-Atala Marco A. Valenzuela-Esc´arcega Clayton T. Morrison Mihai Surdeanu University of Arizona Tucson, Arizona, USA {enoriega,marcov,claytonm,msurdeanu}@email.arizona.edu Abstract Recent efforts in bioinformatics have achieved tremendous progress in the ma- chine reading of biomedical literature, and the assembly of the extracted biochem- ical interactions into large-scale models such as protein signaling pathways. How- ever, batch machine reading of literature at today’s scale (PubMed alone indexes over 1 million papers per year) is unfea- sible due to both cost and processing over- head. In this work, we introduce a focused reading approach to guide the machine reading of biomedical literature towards what literature should be read to answer a biomedical query as efficiently as pos- sible. We introduce a family of algorithms for focused reading, including an intuitive, strong baseline, and a second approach which uses a reinforcement learning (RL) framework that learns when to explore (widen the search) or exploit (narrow it). We demonstrate that the RL approach is capable of answering more queries than the baseline, while being more efficient, i.e., reading fewer documents. 1 Introduction The millions of academic papers in the biomedi- cal domain contain a vast amount of information that may lead to new hypotheses for disease treat- ment. However, scientists are faced with a prob- lem of “undiscovered public knowledge,” as they struggle to read and assimilate all of this informa- tion (Swanson, 1986). Furthermore, the literature is growing at an exponential rate (Pautasso, 2012); PubMed1 has been adding more than a million pa- pers per year since 2011. We have surpassed our 1http://www.ncbi.nlm.nih.gov/pubmed ability to keep up with and integrate these findings through manual reading alone. Large ongoing efforts, such as the BioNLP task community (N´edellec et al., 2013; Kim et al., 2012, 2009) and the DARPA Big Mechanism Pro- gram (Cohen, 2015), are making progress in ad- vancing methods for machine reading and as- sembly of extracted biochemical interactions into large-scale models. However, to date, these meth- ods rely either on the manual selection of relevant documents, or on the processing of large batches of documents that may or may not be relevant to the model being constructed. Batch machine reading of literature at this scale poses a new, growing set of problems. First, access to some documents is costly. The PubMedCentral (PMC) Open Access Subset2 (OA) is estimated3 to comprise 20%4 of the total literature; the remain- ing full-text documents are only available through paid access. Second, while there have been great advances in quality, machine reading is still not solved. Updates to our readers requires reprocess- ing the documents. For large document corpora, this quickly becomes the chief bottleneck in infor- mation extraction for model construction and anal- ysis. Finally, even if we could cache all reading results, the search for connections between con- cepts within the extracted results should not …

Content truncated to 500 words