Datasets:

bulatovv
/

utmn-study-feedbacks-absa

license: mit
task_categories:
  - text-classification
language:
  - ru
tags:
  - education
pretty_name: UTMN Study Feedbacks ABSA
size_categories:
  - 1K<n<10K

Dataset Summary

A dataset for training and evaluating models in aspect-based sentiment analysis. It contains student reviews of academic courses written in Russian.

Dataset Structure

Data Fields

Input

text: review text

Output — sentiment labels for each aspect, the aspect is described in the column name:

лекции
доклады
проекты
презентации
фильмы
видео-уроки
задания__задачи
онлайн-курс
баллы__оценки
практики__семинары
тесты
домашняя работа
эссе
выступления
зачет__экзамен
материал__информация__темы
литература__учебники
игры__интерактивность
преподаватель

Values in each aspect column are sentiment class labels:

0: absent
1: neutral
2: positive
3: negative

Data Splits

The dataset is split into three parts:

train
validation
test

Random split in proportions: 0.8 / 0.1 / 0.1

	train	validation	test
Input Sentences	1020	127	127

Neutral occurrences

Aspect	Train	Validation	Test
`баллы__оценки`	43	5	7
`видео-уроки`	23	4	1
`выступления`	27	0	5
`доклады`	39	3	6
`домашняя работа`	64	11	5
`задания__задачи`	91	6	6
`зачет__экзамен`	33	4	2
`игры__интерактивность`	19	1	0
`лекции`	104	14	10
`литература__учебники`	29	1	5
`материал__информация__темы`	71	8	4
`онлайн-курс`	12	3	1
`практики__семинары`	108	10	14
`презентации`	87	5	8
`преподаватель`	85	8	11
`проекты`	69	6	7
`тесты`	35	6	3
`фильмы`	33	5	3
`эссе`	14	1	3

Positive occurrences

Aspect	Train	Validation	Test
`баллы__оценки`	188	34	22
`видео-уроки`	13	2	2
`выступления`	20	0	3
`доклады`	6	2	3
`домашняя работа`	47	8	4
`задания__задачи`	106	10	17
`зачет__экзамен`	164	29	20
`игры__интерактивность`	34	2	9
`лекции`	79	10	8
`литература__учебники`	16	4	2
`материал__информация__темы`	199	36	33
`онлайн-курс`	16	2	1
`практики__семинары`	81	10	8
`презентации`	27	5	4
`преподаватель`	480	69	64
`проекты`	20	4	3
`тесты`	25	2	4
`фильмы`	11	1	1
`эссе`	0	0	1

Negative occurrences

Aspect	Train	Validation	Test
`баллы__оценки`	45	6	6
`видео-уроки`	5	2	0
`выступления`	6	1	2
`доклады`	12	1	0
`домашняя работа`	16	2	2
`задания__задачи`	24	2	2
`зачет__экзамен`	31	2	4
`игры__интерактивность`	1	0	0
`лекции`	44	2	5
`литература__учебники`	12	1	0
`материал__информация__темы`	50	4	7
`онлайн-курс`	2	2	0
`практики__семинары`	13	0	2
`презентации`	17	0	1
`преподаватель`	79	8	11
`проекты`	7	2	3
`тесты`	15	2	0
`фильмы`	3	0	0
`эссе`	4	0	0

Summary

Split	neutral	positive	negative
Train	986	1,532	386
Validation	101	230	37
Test	101	209	45

Dataset Creation

Curation Rationale

This dataset was created for training aspect-based sentiment analysis models on course review data. Traditional sentiment analysis lacked the granularity needed for educational analytics.

Source Data

Initial Data Collection and Normalization

Reviews were collected from the website: Otzyvus (Отзывус) On this platform, students from Tyumen State University leave feedback on elective courses they have taken. All reviews available as of May 6, 2024 were collected. Nonsensical or off-topic reviews were excluded.

Who are the source language producers?

The reviews were written by undergraduate students (typically aged 18–21, though not strictly limited to this range).

Annotations

Annotation process

Aspects were identified using keyphrase extraction across the review corpus. Relevant keywords were grouped into unified aspects. Each aspect was assigned one of the sentiment classes based on the following:

Positive: a positive opinion expressed
Neutral: the aspect is mentioned, but no clear opinion or a mixed one is expressed
Negative: a negative opinion is expressed
Absent: the aspect is not mentioned

Annotation was done by two annotators, each labeling half of the dataset. There was no cross-annotation, but difficult cases were discussed jointly and resolved with a final verdict.

Who are the annotators?

The annotators were two students from Tyumen State University.

Personal and Sensitive Information

No personal information (such as name, faculty, or program) is included in the dataset.

Considerations for Using the Data

Social Impact of Dataset

A model trained on this dataset can automatically extract insights into student opinions on the learning process, enabling data-driven decision-making in educational contexts.

Other Known Limitations

The dataset may not be large enough for robust training.
Some aspects are underrepresented.
Additional data or augmentation may be required.
It may be better to use stratified splits instead of random ones.

Additional Information

Dataset Curators

Licensing Information

MIT License

Citation Information

@misc{fazlyev2024studyfeedbackabsa,
  author    = {Albert Fazlyev and Danil Krivorogov},
  title     = {A Dataset for Aspect-Based Sentiment Analysis of Russian Student Course Reviews},
  year      = {2024},
  howpublished       = {\url{https://huggingface.co/datasets/bulatovv/aspect-sentiment-student-reviews}}
}

Contributions

Thanks to the students who actively left reviews — without you, this dataset would not exist. You are changing the future of education!