Personality Trait Detection
Using Bagged SVM over BERT Word Embedding Ensembles

An individual’s personality has a great impact on their lives, affecting their life choices, well-being, health and even preferences and desires. Hence, there is a huge interest to develop models which can automatically identify an individual’s personality with important practical applications such as in recommendation systems [Yin et al., 2018], job screening [Liem et al., 2018], social network analysis [Maria Balmaceda et al., 2014], etc. Our model makes binary predictions of the author’s personality based on the famous Big-Five [Digman, 1990] personality measure, which are the following five traits: Extraversion (EXT), Neuroticism (NEU), Agreeableness (AGR), Conscientiousness (CON) and Openness (OPN).

Common author personality detection techniques usually involve extracting psycholinguistic features from text, such as Linguistic Inquiry and Word Count (LIWC) [Pennebaker et al., 2001], Mairesse features [Mairesse et al., 2007], and SenticNet [Cambria et al., 2018], which are then fed into traditional machine learning classifiers such as support vector machine (SVM) [Hearst et al., 1998], Naïve Bayes, etc. More recent work leverage deep learning and make use of pre-trained word embeddings like Word2Vec [Mikolov et al., 2013] and Glove [Pennington et al., 2014]. Recently,  [Mehta et al., 2020] reviewed the latest advances in deep learning-based automated personality detection from the viewpoint of different input modalities along with recent techniques for effective multimodal personality prediction.

The previous state of the art [Majumder et al., 2017] on the Essays dataset also make use of a deep learning based approach with a convolutional network on top of word embeddings extracted from Word2Vec. They also incorporate other inputs such as the Mairesse features, word count, average sentence length, etc. for their final prediction. Their model outperformed the previous best  [Mohammad and Kiritchenko, 2015] by 0.55%, whereas our model outperforms [Majumder et al., 2017] by 1.04% and at the same time being significantly more computationally efficient to train.

Each of our inputs is an essay with a mean size of around 650 words. The maximum number of tokens BERT can process at a time is 512. Hence, to extract maximum information from the input, we break the essay into multiple chunks (sub-documents), with the maximum length of a chunk being 200 tokens. All these sub-documents of a particular essay are then annotated with the same personality label as that essay. We experimented with various methods of pre-processing the essays text before feeding it to the BERT tokenizer and use the best performing one.

We split the text into a sequence of sentences at the period and question mark characters and remove all characters other than ASCII letters, digits, quotations and exclamation marks. We expand all shortenings (e.g., “you’re” becomes “you are”) which increases the maximum length of a sub-document from 200 to 250 tokens. After this initial pre-processing step, the sub-documents are then fed into the pre-trained BERT_BASE model. For each layer of BERT, we average the contextual token representations of that layer. Then, we concatenate the last four layer representations and concatenate this with the corresponding 84 Mairesse features for the essay. This is then considered as the feature vector for the document, which is of the dimension ${\rm I\!R}^{3156}$.

In the classification phase, we feed the document feature vector to a SVM which predicts a binary label corresponding to a particular personality trait. To enhance the performance further, we use ten SVM classifiers to perform the prediction in parallel like the bagging classifier [Breiman, 1996]. The estimator trains on all the features on the total number of the sub-document stack with replacement and the final predicted model for a document is done by majority voting.

Our model achieves a 1.04% increase in performance in comparison to the previous state of the art along with being significantly more computationally efficient. To put it in perspective,  [Majumder et al., 2017] train their model on an Intel Core i7-4720 HQ CPU with the optimal configuration and it takes about 50 hours to complete. Our fine-tuning model only takes about 7 minutes to train. Table 2 gives a comparison of our model, BB-SVM, with others. We modify various parts of the BB-SVM model and discuss their effect on performance in the following section.

In this paper, we presented a computationally efficient deep learning-based model which outperformed the state of the art on the famous stream of consciousness Essays dataset. We hope that our model can be useful for research teams which do not have access to large computational resources. We believe a promising direction of future research would be to make more interpretable deep learning models which can provide valuable insights into the main psychological features driving these predictions and in turn also help advance psychological studies. Currently, the availability of quality personality datasets is quite limited. If an individual’s personality can predicted with a little more reliability, there is scope for integrating automated personality detection in almost all human-machine interaction agents such as voice assistants, robots, cars, etc.