Research On Detection Of The Causes Of Suicidal Tendency For Weibo Text

Suicide has always been a mental health issue of concern and to be solved all over the world,and it is related to personal health and social harmony and stability.Therefore,finding the cause of suicidal tendency before suicide will play a key role in preventing suicide.The traditional cause analysis of suicidal tendency is mainly carried out by psychologists or experts through interviews,psychological evaluation reports and clinical diagnosis.This method has many shortcomings.The first is that the investigation and diagnosis of intrusive methods have a certain impact on their current mental state.Secondly,due to the current psychological state and privacy considerations,suicides will resist such intrusive investigations or interviews.The most serious is that the traditional methods have limitations of timeliness and poor pertinence.Therefore,the inadequacies and limitations of the traditional research on the causes of suicidal tendencies bring interference and delay to the diagnosis and prevention of suicide.At present,social network platforms such as Twitter,Sina Weibo,Wechat,and so on have developed rapidly in recent years.Users often use social networking platform to publish some information with personal emotions such as their true inner thoughts and mental states,including many active suicidal users,which makes social networks provide good support for studying mental health problems such as depression and suicide.However,there are few researches on the cause detection of suicidal tendency based on social media,and the related researches mainly focus on the mental health assessment and the detection of suicidal tendency.This paper proposes a method based on conditional random field CRF and a Long Short-Term Memory EAP-Bi-LSTM-CRF method to detect the cause of suicidal tendency from suicidal social media microblog texts.CRF model relies on manual construction features,which include: word,Pos,dependency relationship and dominant word distance in dependency syntactic features,suicide dictionary words and dictionary dependency path in suicide dictionary features.The EAP-Bi-LSTM-CRF model based on the benchmark model Bi-LSTM-CRF,replaces the Word2 vec word vector representation with the word vector model ELMo trained on the suicide dataset,and adding Pos representation and attention mechanism.Thisimprovement is based on the following three points: first,Elmo word vector can better express and analyze the context and semantic understanding of words;second,proposing part-of-speech representations through the performance of Pos in CRF experiment detect effect and the idea of using char vector in LSTM network model;finally,combined with the idea of adding attention from coding to decoding in machine translation,this paper uses the attention mechanism in the middle of Bi-LSTM output to CRF of the benchmark model Bi-LSTM-CRF,The output of the LSTM network is used as the input of the CRF after being emphasized.The evaluation index uses accuracy rate,recall rate,and F value,and with the application scenario after detecting the cause of suicidal tendency to use four evaluation index,namely EM,FM,PRM,and WRL.After using all candidate features of CRF,the F values of EM,FM,PRM and WRL were 0.586,0.629,0.683 and 0.757 respectively.The F values of the benchmark model Bi-LSTM-CRF in four evaluation index are 0.593,0.652,0.681 and 0.792 respectively.After integrating the word vector model ELMo,Pos representation and attention mechanism,the F values of EAP-Bi-LSTM-CRF model in four evaluation index are 0.621,0.677,0.700 and0.801 respectively,which are respectively 2.8,2.5,1.9 and 0.9 percentage points higher than that of the benchmark model,and 3.5,4.8,1.7 and 4.4 percentage points higher than the CRF with all features.Counting detected causes of suicidal tendency by the length of the cause,93.3%of the cause of suicidal tendency were less than 4.Through the analysis of the recognition results of the two methods,CRF and EAP-Bi-LSTM-CRF models have better detection effect at the cause of shorter suicidal tendency;meanwhile,EAP-Bi-LSTM-CRF model is significantly better than the CRF model in detecting causes of suicidal tendency of various lengths,The longer the length of causes of suicidal tendency,the more significant the EAP-Bi-LSTM-CRF detection effect.It can be known that EAP-Bi-LSTM-CRF model is better than CRF in both the overall detection effect and the classification detection effect according to the length of the cause of suicidal tendency.