Fig 1.
Taxonomy of counseling methods along with examples.
Here, OP (original poster) is a common Internet terminology for the person who creates posts on peer-to-peer platforms. In peer-to-peer therapy, we inspect the level of engagement in three different categories based on the abundance of interaction with the help-seeker—(a) interactive: if there are back-and-forth conversations between the OP and peers, (b) non-interactive: if the post engages peers, but the OP does not reply to peers, and (c) isolated: if the post does not have any comment, but one-to-one therapy involves the continuous exchange of dialogues between therapist and client (help-seeker).
Table 1.
Statistics of the BeCOPE dataset.
We collected a total of ∼10K posts and ∼50K comments. We annotated all the posts using three core labels—(i) intent, (ii) criticism, and (iii) readability (Clear: Excellent, Good, and Average; Non-clear: Mediocre and Poor). IAA (κ) represents the inter-annotator agreement using Cohen’s kappa score.
Table 2.
Example of posts and their corresponding labels in BeCOPE.
Intent:Help-seeking, Rant, Chit-chat, and Survey; Criticism:Self-criticism (SC), Criticism with reasoning (CR), Criticism with no reasoning (CNR), and No-criticism (NC); Readability:Excellent (5), Good (4); Average (3), Mediocre (2), and Poor (1); Emotion:Admiration, Amusement, Anger, Annoyance, Approval, Caring, Confusion, Curiosity, Desire, Disappointment, Disapproval, Disgust, Embarrassment, Excitement, Fear, Gratitude, Grief, Joy, Love, Nervousness, Optimism, Pride, Realization, Relief, Remorse, Sadness, Surprise, and Neutral; Engagement:Interactive, Non-interactive, and Isolated.
Fig 2.
(a) Confusion matrix to represent the performance of pseudo labeling of criticism, intent, and readability labels. We exploit BERT to fine-tune on ~5K manually annotated posts to predict criticism, intent, and readability on the remaining posts. (b) Distribution of behavioral signals (criticism and intent) along with readability in the complete BeCOPE dataset.
Fig 3.
Distribution of behavioral signals and readability in BeCOPE across all engagement categories.
(a) The intent distribution indicates that a majority (45.35%) of posts show explicit intentions (seek-help) through queries or the articulation of pressing needs on OMHC platforms, yielding productive responses as opposed to merely airing surveys or rants. (b) The criticism distribution shows that help-seekers are more likely to engage in self-criticism (43.32%), and those who criticise openly on others with proper reasoning are more likely to receive assistance. (c) The readability statistics of posts in BeCOPE state that well-written posts receive 2.2× more support (responses) as compared to poorly written posts.
Fig 4.
(a) Distribution of emotion labels in the BeCOPE dataset. For brevity, we show plots for the top 10 emotion labels only. Each post is tagged with primary and secondary emotion labels. We further analyze the emotion label distribution across three engagement categories. (b) Topical analysis on the BeCOPE dataset. We perform Latent Dirichlet Allocation (LDA) [23] to form 8 clusters of topics. To analyze the topics on which peers respond, we club interactive and non-interactive posts, where peers respond and compare them with topics from isolated posts.