task_categories:
- text-classification
language:
- ru
tags:
- tl
- tg
- mirea
pretty_name: tl-mirea-topic-classification
size_categories:
- 1K<n<10K
viewer: true
configs:
- config_name: default
data_files: data/*.parquet
RTU MIREA Telegram Channel Messages Topic Classification
The dataset has been created to fine-tune mDeBERTa-v3-base-mnli-xnli model for text classification tasks on Telegram messages with 17 different topics specific to the target of classification.
Overview
This dataset has been parsed from the official Telegram channel of RTU MIREA, using aiogram, and manually labelled based on the topic of a given message.
Preprocessing
Preprocessing of the data included text trimming, emoji and hashtags removal, NA values processing and data labeling.
Labelling
EDA was performed to check clustering abilities of the messages and find common topics. For embeddings has been used paraphrase-multilingual-MiniLM-L12-v2 from sentence-transformers and HDBSCAN algorithm for clustering.
As it can be inferred, a very small portion of messages has clustering abilities (small light grey dots - noise). 90% of all data is referred by the algorithm as noise, the messages are clustered into 64 groups. Checking these groups gave a valuable insight into some topics of the channel (e.g. weather_forecast, rating, digest, cybersecurity).
After data exploration finished, the following 18 topics has been proposed:
- admission
- achievement
- social
- digest
- science
- career
- patriotism
- sport
- volunteering
- competition
- education
- announcement
- international_relations
- rating
- scholarship
- weather_forecast
- cybersecurity
The following barplot shows the amount of messages for each topic, the overall is 1003 messages out of the current total ~ 10k:
Clearly, topics have a significant class disbalance, but since a pretrained classification model for fine-tuning has been used, and some topics are not so well-distinguished as others, it has been decided not to pay too much attention to this problem.
Contact
If you have any suggestions, please contact me through issues.