I describe my participation on the 2015 edition of the BioASQ challenge in which I submitted results for the concept matching, document retrieval, passage retrieval, exact answer and ideal answer subtasks. My approach relies on a in-memory based database (IMDB) and its built-in text analysis features, as well as on PubMed for retrieving relevant citations, and on prede ned ontologies and terminologies necessary for matching concepts to the questions. Although results are far below the ones obtained by other groups, I present an novel approach for answer extraction based on sentiment analysis.
1 http://bioasq.org/
I participated with a system developed on top of an in-memory database (IMDB)
[
Similar to previous QA systems [
In this step, the system processes the questions using the Standford CoreNLP
[
The second query aims on more precision and less recall and lters tokens further based on a list of the 5,000 most popular words of English3 and uses the \AND" operator for connecting words. Only the document retrieval step used this high precision query for ranking relevant documents from PubMed. For instance, for the question \What disease is mirtazapine predominantly used for?", \disease OR mirtazapine OR predominantly OR used" is the resulting high recall query and \mirtazapine AND predominantly" is a higher precision query. 2.3
The approach is the same that I used in the 2014 edition of the challenge [
The approach for retrieving relevant PubMed documents for each question is
similar to the one described in my recently submitted paper [
I retrieved passages using on the built-in information retrieval features available in the IMDB, which is based in approximated string similarity to match terms from the query to the words in the documents. The system proceeds ranks the passages (sentences) based on the TF-IDF metrics and I retrieve the top 10 sentences and corresponding documents as answers for the passage and document retrieval sub-tasks, respectively. 2.5
I extracted both exact and ideal answers based on the gold-standard snippets that the organizers made available for phase B of task 3b. The process consisted in inserting the snippets into the IMDB database and I utilized built-in text analysis features for the extracting the answers, as described in details below for each question type.
Yes/No: Decision on either the answers \yes" or \no" was based on the sentiment analysis predictions provided by the IMDB. The assumption was that all snippets are somehow related to the question and that detection of sentiments in these passages could be used to distinguish between the two possible answers. Figure 3 shows the sentiments which were detected for a certain question.
The IMDB returns 10 types of sentiments, namely \StrongPositiveSentiment"", \StrongPositiveEmoticon", \WeakPositiveSentiment", \WeakPositiveEmoticon", \StrongNegativeSentiment", \StrongNegativeEmoticon", \MajorProblem", \WeakNegativeSentiment", \WeakNegativeEmoticon" and \MinorProblem". I merged some of these sentiment types into coarser categories according to simples rules (Table 1). The sentiments were rst grouped into four coarse categories, i.e., \positiveStrong", \positiveWeak", \negativeStrong", \negativeWeak", and then into the three main sentiments \positive" or \negative". For the rules shown in Table 1, I consider that the \positiveStrong" sentiment is stronger than the \negativeStrong" one, and therefore I assign the \positive" sentiment for such cases. Similarly, I consider \positiveWeak" weaker than \negativeWeak" when both are returned for the same question. Cases which did not match none of the rules for \positive" or \negative" sentiments are classi ed as \neutral". Final decision for the the answers \yes" or \no" was based on these three coarse sentiments. By default, I return the answer \no", unless I get \positive" or \neutral' as output from the above rules. coarse sentiment Rule positiveStrong StrongPositiveSentiment OR StrongPositiveEmoticon positiveWeak WeakPositiveSentiment OR WeakPositiveEmoticon negativeStrong StrongNegativeSentiment OR StrongNegativeEmoticon OR MajorProblem negativeWeak WeakNegativeSentiment OR WeakNegativeEmoticon OR MinorProblem positive (positiveStrong OR positiveWeak) AND (NOT(negativeStrong)
AND NOT(negativeWeak)) positive positiveStrong AND negativeStrong positive positiveStrong AND negativeWeak negative (negativeStrong OR negativeWeak) AND (NOT(positiveStrong)
AND NOT(positiveWeak)) negative positiveWeak AND negativeStrong negative positiveWeak AND negativeWeak Factoid and list: I extracted factoid and list answers based also on built-in predictions provided by our IMDB, more speci cally, on the annotations of noun phrases and topics, as presented in Figure 4. Given that no semantic processing was performed neither for the question nor for the snippets, in oder to tag named entities and to identify the entity type of expected answer, I choose the ve top answers based on the order returned by the IMDB.
Summary: I also built summaries for the ideal answers based on the phrases which contain sentiments, as shown in Table 6. The assumption was that such phrases are more informative and relevant than the ones in which no sentiments were found. My approach consisted in concatenating the sentences up to a limit of 200 words, as speci ed in the challenge's guidelines. 3
I submitted results for all ve batches of test questions for task 3b: (a) phase A, i.e., concept mapping and document and passage retrieval, and (b) phase B, i.e., exact and ideal answers. Di erent from previous editions of the BioASQ challenge, when participants were allowed to submit up to 100 entries per question for each of the required sub-tasks, whether documents, concepts or exact answers, this year's edition limited concepts, documents and passages up to 10 per question and factoid answers up to 5. I present below the results I obtained as published by the organizers in the BioASQ Web site 4. I do not show results for concept matching because the organizers seem not to have made them available yet.
Table 3 shows my results for document retrieval for each of the ve test batches. As discussed in the methods section, I did not implement any speci c approach for this task and documents were ranked based on the relevancy of the query to the passages and not to the documents (abstracts) themselves. In 4 http://participants-area.bioasq.org/results/3b/phaseA/;http:// participants-area.bioasq.org/results/3b/phaseB/
second-generation antidepressants (selective serotonin reuptake inhibitors, nefazodone, venlafaxine, and mirtazapine) in participants younger than 19 years with MDD, OCD, or non-OCD anxiety disorders. patients 65 years or older with major depression.
A case report involving linezolid with citalopram and mirtazepine in the precipitation of serotonin syndrome in a critically ill bone marrow transplant patient is described in this article.
In 26 patients with FMS who completed a 6-week open study with mirtazapine, 10 (38%) responded with a reduction of at least 40% of the initial levels of pain, fatigue and sleep disturbances (Samborski et al 2004).
In general, drugs lacking strong cholinergic activity should be preferred. Drugs blocking serotonin 5-HT2A or 5-HT2C receptors should be preferred over those whose sedative property is caused by histamine receptor blockade only. this year's edition of the challenge, organizers required participants to submit up to 10 document, which is a hard assignment, given the millions of citations in PubMed. Indeed, results have been lower than the ones obtained by participants last year and it is unclear whether we (teams) performed better than the baseline systems as the organizers did not publish results for these systems yet.
Table 4 shows my results for passage retrieval for each of the ve test batches. Few groups participated in this task, in comparison to the number of submissions for the document retrieval task. A task which is already very complex has been made even more di cult this year by the limitation of providing up to only 10 top passages.
Finally, tables 5 and 6 shows the results I obtained for the exact and ideal answers in phase B of task 3b. Acknowledgements MN would like to acknowledge funding from the HPI Research School.