Suicidal Tendencies: The Automatic Classification of Suicidal and Non-Suicidal Lyricists Using NLP

International Joint Conference on Natural Language Processing, pages 680–684, Nagoya, Japan, 14-18 October 2013. Suicidal Tendencies: The Automatic Classification of Suicidal and Non-Suicidal Lyricists Using NLP Matthew Mulholland Montclair State University Montclair, NJ, USA mulhollandm2@mail.montclair.edu Joanne Quinn Montclair State University Montclair, NJ, USA quinnj11@mail.montclair.edu Abstract Can natural language processing be used to predict the likelihood that a musician will commit or has committed suicide? In order to explore this idea, we built a cor- pus of songs that includes a development set, a training set, and a test set, all con- sisting of different lyricists. Various vo- cabulary and syntactic features were then calculated in order to create a suicide/non- suicide song classifier. The features were input into the Weka machine learning suite and tested with an array of algorithms. We were able to achieve up to a 70.6% classifi- cation rate with the SimpleCart algorithm, a 12.8% increase over the majority-class baseline. Our findings suggest that syntac- tic and vocabulary features are useful indi- cators of the likelihood that a lyricist will commit or has committed suicide. 1 Introduction Recently, research into the application of NLP to the detection of health illnesses has proved fruit- ful. For instance, in their study of the effects of dementia on writers, Le et al. (2011), guided by previous research, explored various hypotheses in the novels of three British writers. Their research found a decline in the type-token ratio of the nov- elists suffering from dementia. The use of passive constructions was also explored since it is gener- ally believed that it represents a syntactic structure that is particularly complex and likely would be used less often by writers suffering from demen- tia. Indeed, they found that authors with dementia use less passives than their healthy peers. Their results indicate the potential for natural language processing in the language of mental illness. Sim- ilarly, much research into the application of NLP to depression and suicide prediction has been con- ducted in recent years (Pestian et al. 2012; Sohn et al. 2012). While one might not expect depression or a suicidal tendency to affect language in the same way as an illness such as dementia, it is reason- able to assume that there will be textual indica- tions of these mental illnesses also. In Stirman and Pennebaker (2001), word use is treated as an indicator of the mental states of suicidal and non- suicidal poets. Stirman & Pennebaker developed the Linguistic Inquiry and Word Count program to analyze over 70 language dimensions, includ- ing: polarity, affect states, death, sexuality, tense, etc. Their research found a correlation between the likelihood of a poet committing suicide and his/her disengagement from society based on the suicidal poets’ heavy use of first-person singular pronouns and decreased use of the first-person plural pro- nouns. Interestingly, they also noted that poets who had committed suicide generally used more sexual words and references to death than their non-suicidal counterparts. Additionally, latent se- mantic analysis has been used to detect depression in …

Content truncated to 500 words