Answers are grounded in this paper and cite their sources.
Digital journaling is a low-cost channel for emotional self-monitoring, but most deployed systems reduce affect to coarse positive/negative/neutral sentiment, run inference in the cloud, and expose no reasoning for their outputs — three properties that are poorly suited to sensitive reflective text. This paper presents Mind Mirror, a locally hosted emotional journaling prototype that combines fine-grained emotion classification, token-level attribution, and local-first data handling in a single deployable system. Classification is performed by a pre-trained DistilBERT GoEmotions student checkpoint that natively emits 28 labels; The application combines those outputs after the fact into six target categories—joy, sadness, anger, fear, disgust, and surprise—by applying the 27-to-6 Ekman taxonomy mapping as stated in the official GoEmotions repository, with the neutral label omitted and the distribution renormalised; since the checkpoint does not have a six-class classification head, the aggregation process is also not included. Inference is performed on the user’s own machine through a Python FastAPI service, along with a React and TypeScript front end, a Node.js and Express back end, and local JSON storage that is encrypted with AES-256-GCM. The token attributions are calculated using a leave-one-out perturbation method and are shown in the interface. Three separate evaluations are carried out and kept as distinct. The version of the checkpoint that includes the production 28-to-6 aggregation is assessed at the six-class argmax stage before the confidence and entropy presentation gate, yielding 70.80% accuracy, a macro F1 of 0.5818, and a weighted F1 of 0.7032 on a deterministic subset consisting of the first 1000 eligible examples from the GoEmotions test split, eligibility being defined as the annotations corresponding to exactly one target class; this subset is not randomly chosen and is not necessarily representative of the entire filtered split. Six separately trained six-class models are compared under the same filtering conditions: a DistilBERT baseline with constraints reaches 46.50% accuracy but collapses to the majority class, while a class-weighted configuration trained on a stratified subset of 5090 examples taken from the single-target training pool using seed = 42 achieves an accuracy of 55.83% and a macro F1 score of 0.4765 on the full 3539-example single-target test split. The average observed prototype response time was about 2.2 s over ten scripted requests on a machine that had only a CPU. Furthermore, a descriptive three-day observation involving N = 60 college students recorded their self-reported emotion choices, and in this study 90.0% of the participants selected a different emotion on at least one day; the observation is purely descriptive and does not validate anything about the classifier. The findings show that it is possible to carry out locally hosted, privacy-oriented emotion analysis with token-level attribution using off-the-shelf hardware. Here, privacy refers to a feature of the architecture and not to an actual security evaluation, and there is no evaluation of the system on journal text.
No takes yet. Share an insight, caveat, or question.
D et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: