UIT-ViLipBTA

UIT-ViLipBTA is a Vietnamese dataset for Brand-Targeted Aspect-Based Sentiment Analysis (BT-ABSA), built from public TikTok comments about lip-care products (lip balms). Each comment is annotated with one or more quadruples:

(CommentType, Brand, Aspect, Sentiment)

Brand is an open-vocabulary element: users refer to brands through abbreviations (ea → Elizabeth Arden), product-line names (lip monster → Kate), misspellings (vasaline → Vaseline), or only through the parent comment they reply to. Replies are released together with their parent comment as conversational context.

• Paper: UIT-ViLipBTA: A Vietnamese TikTok Dataset for Brand-Targeted ABSA in Lip-Care Product Discussions, RIVF 2026
• Authors: Anh Thi-Hoang Nguyen, Dung Ha Nguyen (University of Information Technology, VNU-HCM)
• Language: Vietnamese (informal social-media register: teencode, abbreviations, missing diacritics, emoji)
• License: CC BY-NC-SA 4.0

Quick start

from datasets import load_dataset

ds = load_dataset("visolex/UIT-ViLipBTA")
ex = ds["train"][0]
print(ex["context"], "->", ex["text"])
print(ex["quadruples"])

Dataset statistics

Split	#Videos	#Comments	#Quadruples	% Replies	Labels
train	107	3,403	4,748	39.3	silver
validation	16	487	677	34.7	silver
test	12	446	518	39.0	gold
Total	135	4,336	5,943	38.7

Comments span June 2022 – September 2025. The split is by video: no video contributes comments to more than one split, so every reply sits in the same split as its parent. The partition was selected from 800 random permutations to minimise divergence of the aspect, sentiment and reply distributions from the full corpus.

CommentType

CommentType	#Quadruples	%
ProductReview	5,000	84.1
OutOfDomain	652	11.0
ContentCreator	171	2.9
VideoContent	120	2.0

62 comments carry more than one CommentType, so CommentType classification is multi-label.

Aspect × Sentiment (ProductReview quadruples)

Aspect	Positive	Negative	Neutral	Total
GeneralExperience	1,352	262	124	1,738
Benefits	813	269	74	1,156
Color	442	166	134	742
SideEffects	18	311	6	335
Price	125	109	23	257
Texture	76	125	35	236
Longevity	79	104	21	204
Scent	115	57	23	195
Packaging	62	26	18	106
Ingredients	24	3	4	31

The distribution is long-tailed (the top three aspects cover 72.7% of quadruples) and polarity is strongly tied to aspect (SideEffects is 92.8% negative). Report per-class or macro scores alongside micro-F1.

Brand frequency is Zipfian: 2,565 ProductReview quadruples carry General (no specific brand), and the remainder spread over a long tail of brands, many appearing only once or twice.

Data fields

Field	Type	Description
id	string	Anonymised comment identifier (keyed hash of the TikTok comment ID)
video_id	string	Anonymised identifier of the source video. Use it to group comments by thread; never split a video across train and test
parent_id	string	Anonymised identifier of the parent comment for replies, empty otherwise. The parent is not necessarily a row in the dataset
is_reply	bool	Whether the comment is a reply
context	string	Text of the parent comment for replies, empty otherwise
text	string	The comment text
quadruples	list	One or more {comment_type, brand, aspect, sentiment} records
label_source	string	silver (LLM-assisted with human review of disagreements) or gold (blind triple annotation + adjudication)

Label space

Element	Values
comment_type	ProductReview, ContentCreator, VideoContent, OutOfDomain
brand	Canonical brand name (open vocabulary), General (lip-care product without a specific brand), or None (non-ProductReview)
aspect	GeneralExperience, Benefits, Color, SideEffects, Price, Texture, Longevity, Scent, Packaging, Ingredients, or None (non-ProductReview)
sentiment	Positive, Negative, Neutral, or None (OutOfDomain only)

Structural constraints:

• ProductReview → (ProductReview, <brand|General>, <aspect>, <Positive|Negative|Neutral>)
• ContentCreator / VideoContent → (type, None, None, <Positive|Negative|Neutral>)
• OutOfDomain → (OutOfDomain, None, None, None)

None is stored as the string "None", not as null.

Example

{
  "id": "d0d148df59f2a0f4",
  "video_id": "addd6272a1313d72",
  "parent_id": "2971f53869e900b1",
  "is_reply": true,
  "context": "Đỏ gạch thì màu số bnhiu vậy ạ",
  "text": "Tui mua Roman 029 thấy đánh OK lắm 😉😉😉",
  "quadruples": [
    {"comment_type": "ProductReview", "brand": "Roman", "aspect": "Color", "sentiment": "Positive"}
  ],
  "label_source": "gold"
}

Additional files

File	Description
brand_dict.json	Brand normalisation dictionary: surface form (lower-cased) → canonical brand. 117 canonical brands. Required for the normalized match metric
guidelines/btabsa_guidelines_vi.docx	Annotation guideline (Vietnamese)

Tasks and evaluation

Given a comment c and, for replies, its parent p:

• Task 1 — CommentType classification: predict the set of comment types (multi-label). Micro/macro F1.
• Task 2 — Aspect–Sentiment pair extraction: predict the set of (aspect, sentiment) pairs of ProductReview quadruples. Micro/macro F1.
• Task 3 — Quadruple extraction: predict the full set of quadruples. F1 at three match levels:
  • Exact: raw string equality on all four elements.
  • Normalized: gold and predicted brands are both mapped through brand_dict.json before comparison.
  • Triplet: brand dropped; (comment_type, aspect, sentiment) must match.

Baseline results (test set, mean over 3 seeds)

Task	Metric	Best model	Score
Task 1	Micro-F1	XLM-R (majority baseline: 0.705)	0.748
Task 2	Micro-F1	ViSoBERT	0.611
Task 3	Normalized F1	ViSoBERT	0.518
Task 3	Exact F1	BARTpho-syllable	0.471
Human	Pairwise quadruple F1	3 annotators	0.649

Adding the parent comment as context improves ViSoBERT's Task 3 normalized F1 by 0.080 overall and by 0.154 on replies. See the paper for the full tables.

Dataset creation

Collection

Public Vietnamese comments were collected from 135 TikTok videos about lip-care products, found through product-related search queries. For each video, up to 200 top-level comments and their replies with language tag vi were kept. Five copy-pasted spam comments were removed, and texts duplicated across splits were removed from the lower-priority split (priority test > validation > train). Duplicates within a split were kept because they come from different users or videos.

Annotation

1. A guideline was refined on a pilot of 100 comments annotated by two members of the research team.
2. The remaining comments were annotated independently by two to three large language models each. High-agreement cases were accepted; disagreements were resolved by human annotators. These are the silver labels in train and validation.
3. Brand labels were normalised with a 126-entry dictionary (326 observed ProductReview surface forms → 117 canonical brands).
4. Test set (gold): three trained annotators re-annotated all 446 test comments from scratch, blind to existing labels. Disagreements were resolved by majority vote, then expert adjudication.
5. Systematic errors found during adjudication (e.g., product-line names missing from the dictionary) were fixed as dictionary entries and guideline rules and re-applied to all splits. No adjudicated test label was copied into training data.

Annotation quality

Measure	Value
Fleiss κ — CommentType (macro over labels)	0.513
Fleiss κ — Aspect	0.697
Fleiss κ — Sentiment	0.634
Pairwise quadruple F1 between annotators	0.649
Silver vs. gold, full quadruple F1 (test, n = 446)	0.604
Silver vs. gold, full quadruple F1 (train sample, n = 554)	0.583

The low CommentType κ is driven by ContentCreator (κ = 0.175, assigned only 11 times across three annotators); the main substantive disagreement is the ProductReview/OutOfDomain boundary, e.g., resale offers that mention a product without evaluating it.

Limitations and known issues

• Train and validation labels are silver. They agree with gold at about 0.58–0.60 quadruple F1. Treat the test set as the only gold-standard evaluation.
• Annotation provenance was not preserved per sample: it is not recorded which silver labels were accepted from model agreement and which were decided by a human.
• Four gold test brands are surface forms rather than canonical names: 029 (a shade code), Sài (the verb xài, "to use"), toboco (likely Tocobo), and wrinkle plus alpha cx. They were kept unchanged so that the paper's results stay reproducible.
• Single domain and platform. The data covers lip-care products on TikTok only; transfer to other categories or platforms has not been evaluated.
• Brand names appear as part of users' opinions. They do not reflect any assessment by the authors.

Ethical considerations and privacy

• Only comments that were public at collection time are included. No user profiles, usernames, avatars or video content are released.
• User IDs, timestamps and engagement counts were removed. Comment, parent and video IDs were replaced by keyed hashes (HMAC-SHA256 with a secret key), so they cannot be mapped back to TikTok IDs. Reply structure is still preserved.
• Comment text is released verbatim, including any @mentions users typed, and may still be searchable on TikTok. If you find your own comment in the dataset and want it removed, contact the authors (see below) and it will be removed in the next version.
• The dataset is for non-commercial research only (CC BY-NC-SA 4.0).

Citation

@inproceedings{nguyen2026uitvilipbta,
  title     = {{UIT-ViLipBTA}: A Vietnamese {TikTok} Dataset for Brand-Targeted {ABSA} in Lip-Care Product Discussions},
  author    = {Nguyen, Anh Thi-Hoang and Nguyen, Dung Ha},
  booktitle = {Proceedings of the 2026 RIVF International Conference on Computing and Communication Technologies (RIVF)},
  year      = {2026},
  publisher = {IEEE}
}

Contact

Anh Thi-Hoang Nguyen — anhnth@uit.edu.vn Dung Ha Nguyen — dungngh@uit.edu.vn

Acknowledgement

This research was supported by The VNUHCM-University of Information Technology's Scientific Research Support Fund.