Research On Joint Recognition Of Trigger Words And Attributes In Biomedical Events

ObjectiveThis article conducted joint recognition research of trigger words and attributes in biomedical events.By constructing a bidirectional long short-term memory-conditional random field(BiLSTM-CRF)model,the recognition of event trigger words and attributes was transformed into sequence labeling problem,which combined word vectors and dependence characteristics.Then taked pharmacovigilance events as empirical objects to evaluate their results.This research helps people to do fine-grained extraction of pharmacovigilance events,also supports the mining of relevant information.MethodBy reading the relevant literature about extracting biomedical events at home and abroad,focused on the identification of event trigger words and attributes,to learn the methods used to determine the identification method in this article.Firstly,the training set and test set of the biomedical events corpus were divided,and the words in the corpus were labeled by the IBO system,and different types of event trigger words and attributes words were labeled.Then,learn word vectors based on large-scale unlabeled Pubmed abstracts,and extract word vector features in sentences.After that,Stanford CoreNLP,an NLP analysis tool developed by Stanford University,was used for syntactic analysis to extract the characteristics of dependencies in sentences.Then,for the training set,the word vector features and dependency relationship features were input into the BiLSTM-CRF model,and appropriate parameters were selected to perform supervised deep machine learning through labeled event trigger words and attributes categories.Finally,the test results were used to obtain the experimental results,The accuracy,recall,and F values were used for evaluation and analysis.For the evaluation of the method,as no joint recognition of event trigger words and attributes has been found,this article choosed the trigger word recognition experiment based on the EventMine from original article about corpus for comparison.Firstly,the BiLSTM-CRF model is used to identify the trigger words.to prove the effectiveness of the method,and then evaluate and analyze the effectiveness of joint identification.Results(1)The trigger words in pharmacovigilance events was identified by the BiLSTM-CRF method.Its F value was 74.2%.Compared with the effect of EventMine in the original article,the F value was 61.6%.The superiority of this method was proved.(2)Through joint recognition of trigger words and attributes for pharmacovigilance events,the overall F value is 72.5%,and the F value of each category exceeds 70.0%.Among them,the Manner attribute recognition effect is the best,the accuracy rate,100.0%,is the highest of all recognition results,followed by the Negated negative attribute,the accuracy rate is 91.7%,which has a certain relationship with the number of these two categories.The F value of the combination recognition result is the lowest,being 70.1%,and its accuracy rate is 92.5%,which is not bad,but the recall rate is only 56.5%,which is relatively low,indicating that it has poor sensitivity and cannot be easily identified.The highest recall rate is the trigger word for Adverse_effect side effects,which reached 77.7%,which has a certain relationship with the number of Adverse_effect,and its sensitivity is good.ConclusionThe extraction of biomedical events is a hotspot in today's biomedical natural language processing research.Fine-grained extraction is an inevitable trend of development in the future.The joint recognition of trigger words and attributes is conducive to refined selection and improves information use efficiency.Relevant personnel are better able to make advance judgments on current information.(1)In this article,the trigger words and attributes in pharmacovigilance events were jointly identified.The combination of word vectors and dependencies was used to perform a supervised deep learning task based on the BiLSTM-CRF model.(2)Through the method proposed in this research,the event trigger words and attributes in pharmacovigilance events were jointly identified,and the recognition results performed well.