Séminaires RALI-OLST

deepBioWSD: A One-size-fits-all Deep Bidirectional LSTM for Word Sense Disambiguation of Biomedical Text Data

Ahmad Pesaranghader (ahmadpgh <at> gmail (point) com)

Dalhousie University

Wednesday 13 December 2017 at 11:30 AM

P-217 Roger-Gaudry — !!! attention : salle inhabituelle !!!

With the recent advances in biomedicine, we have a wealth of information hidden in unstructured narratives such as research articles and clinical documents. A high accuracy Word Sense Disambiguation (WSD) algorithm will avoid a myriad of downstream difficulties in the natural language processing (NLP) applications pipeline dealing with such data. Due to large feature spaces, traditionally, discriminative approaches such as logistic regression and support vector machines with n-grams and syntactic and semantic features (e.g., part of speech and named entities) have been used for WSD where additional performance gains were typically made through feature selection and ensemble approaches. In these supervised algorithms, which, by and large, outperform unsupervised and knowledge-based methods, generally per each ambiguous term in the given domain one classifier will be trained. Taking into account the large number of ambiguous terms in the biomedical domain, and consequently, the need for an extensive number of expressive data to train the classifier of each term, we face two important bottle-necks in proper deployment of such models.

In this paper, by introducing our one-size-fits-all model of deepBioWSD, we demonstrate a more direct WSD approach that leverages a Bidirectional Long Short-Term Memory (BLSTM) network. In deepBioWSD framework, first, the Unified Medical Language System (UMLS) concept embeddings will be computed using their textual definitions; and then, a single WSD network will be initialized with these embeddings prior to being trained on the ambiguous terms collectively. Having this single network enables the model to share statistical strength and to scale well with the vocabulary size. Evaluating on MSH-WSD dataset, we show deepBioWSD outperforms traditional approaches in biomedical text WSD. We also represent deepBioWSD, by demanding less number of training data, and also by drawing on the learned sequential information of the surrounding contexts, works better than the recent deep learning-based one-classifier-per-one-term WSD algorithms. Some important factors that offer deepBioWSD to be accurately deployable in real-time environments.

To receive weekly talk announcements, please send an e-mail to majordomo@iro.umontreal.ca. Simply write a message containing the single line 'subscribe ralli' (without the quotes, with a double 'l' in 'ralli').

See all the weekly talks for the year:

1991 1992 1993 1994 1995 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022