HELPER: How We Built a Corpus of Psychologically Charged Texts and Fine-Tuned Qwen for Local Emotion Analysis
Students and researchers from HSE University and VK have created a corpus of psychologically charged Russian-language texts with local annotation of emotions and psychological characteristics. The HELPER annotation tool was developed, and based on the corpus, Qwen-3 8B and 14B models were fine-tuned for the tasks of extracting emotional fragments and assessing the author's state.
A team from the HSE University and VK Applied AI Workshop (Alexander Sabko, Victoria Belyavskaya, Sergey Pavlukhin) has presented the HELPER project, aimed at analyzing the emotional state of a text's author, taking into account local emotional manifestations and their dynamics. Standard sentiment analysis approaches (binary or multiclass classification) are unsuitable for psychologically loaded texts, as emotions appear fragmentarily rather than as a general background. Existing Russian-language datasets have different annotation schemes and cannot be combined. The team conducted 10 in-depth interviews with practicing psychologists and, based on the analysis, developed a comprehensive annotation scheme consisting of four components: emotional segmentation (14 emotions based on Ekman and OCC), evaluation of psychologist responses by four criteria, thematic tags, and a final assessment of the client's state on five scales. The corpus was collected from three sources: counseling dialogues from B17.ru, reviews from Yandex.Maps, and VKontakte posts with comments. Currently, the corpus includes 786 dialogues annotated for 7 emotions (about 9 million character labels) and 156 dialogues under the extended scheme of 14 emotions, as well as reviews and posts. For annotation, the HELPER application was created using the .NET 10, ASP.NET Core, Blazor, and PostgreSQL stack with authentication via Keycloak. The application supports depth-first prioritization of dialogues, a scoring system for evaluating annotators' work, and a suspicious user detector. To demonstrate the data's effectiveness, the team fine-tuned the Qwen-3 8B and Qwen-3 14B models using LoRA on their corpus (computations on HSE's cHARISMa). Dialogues are normalized before input (removing BB codes, URLs), but speech style is preserved, as speech features are important for analysis. The project showed that the collected corpus yields better results for psycholinguistic analysis tasks compared to generalized approaches.
Source: Habr — хаб ML —
original