SemEval-STM — Dataset Description

SemEval-STM is a dataset built on SemEval data to demonstrate the superiority of Segment Topic Modeling (STM). It contains annotations for two topic allocation paradigms:

• Document Based Topic Allocation (DBTA): assigns a single topic to an entire document
• Segment Based Topic Allocation (SBTA): assigns topics at the segment level within a document

---

Last updated: 2026-04-19 Rows marked bold in the summary table are the primary evaluation datasets (23 laptop / 11 restaurant topics).

---

Summary Table — Data Generation Pipeline

Step	File Pattern	Domain	# Topics	# Entries	Granularity
[1]	raw_input/*_json.json	L/R	—	395 / 335	review → topic list
[2]	raw_input/*_segmentation_generation.json	L/R	—	395 / 335	review → JSON str
[3]	raw_input/*_segmentation_document_wise_topic.json	L/R	—	395 / 335	review (structured)
[4]	raw_input/*_dbta.json	L/R	76 / 12	1,982 / 1,435	full review / topic
[5]	raw_input/*_sbta.json	L/R	76 / 12	2,287 / 1,837	segment / topic
[6]	dbta/*_dbta_preprocessed.json	L/R	32 / 12	1,825 / 1,435	full review / topic
[7]	sbta/*_sbta_preprocessed.json	L/R	33 / 12	2,133 / 1,837	segment / topic
[8]	dbta/*_dbta_curated.json	L/R	23 / 11	1,611 / 1,344	full review / topic
[8]	dbta/shuffle_test/*_dbta_curated_shuffled_{0-4}.json	L/R	23 / 11	1,611 / 1,344	full review / topic
[9]	sbta/*_sbta_curated.json	L/R	23 / 11	2,123 / 1,767	segment / topic
[9]	sbta/shuffle_test/*_sbta_curated_shuffled_{0-4}.json	L/R	23 / 11	2,123 / 1,767	segment / topic
[+]	SIT_TASK_DATA/*_sit_dataset_250517.json	both	—	800 / domain	sentence (task)

Domain [L = laptop, R = restaurant] shuf{} = 5 shuffled seeds (shuffled_0 through shuffled_4)

---

Data Source

All raw reviews originate from SemEval 2016 Task 5: Aspect-Based Sentiment Analysis (ABSA).

Domain	Source	Load method	# Reviews	License
Laptop	HuggingFace scholl99/semeval-2016-absa-laptop-processed — train split	load_dataset(..., split='train')	395	CC BY-NC-SA 3.0
Restaurant	SemEval 2016 Task 5 (alt.qcri.org) — ABSA16_Restaurants_Train_English_SB2.xml	xml.etree.ElementTree parsing	335	CC BY-NC-SA 3.0

Common preprocessing applied to both domains:

• MISCELLANEOUS category label → GENERAL
• Laptop only: sentiment polarity suffix (_positive, _negative, _neutral, _conflict) stripped from topic labels; entries with : in sentence ID excluded

---

License

This dataset is licensed under CC BY-NC-SA 4.0, and is subject to the Terms of Use of the model (OpenAI o3-mini) used in its creation.

---

Overview

This dataset supports research on topic-based opinion mining using SemEval review data (laptop & restaurant domains). It contains two paradigms:

• DBTA (Document-Based Topic Assignment): Full review sentences/paragraphs grouped by aspect topic.
• SBTA (Segment-Based Topic Assignment): Fine-grained text segments (sub-sentences) extracted per topic.

The pipeline reduces 76 (laptop) / 12 (restaurant) raw topics down to a curated set of 23 / 11 topics shared by both DBTA and SBTA, enabling a fair head-to-head comparison of the two paradigms.

---

Directory Structure

SBTA_DATASET/
├── SIT_TASK_DATA/                         # Sentence Intruder Task data
│   ├── laptop_sit_dataset_250517.json
│   └── restaurant_sit_dataset_250517.json
│
└── semeval_stm/
    ├── laptop/
    │   ├── raw_input/                     # Steps [1]–[5]
    │   │   ├── laptop_json.json
    │   │   ├── laptop_segmentation_generation.json
    │   │   ├── laptop_segmentation_document_wise_topic.json
    │   │   ├── laptop_dbta.json
    │   │   └── laptop_sbta.json
    │   ├── dbta/                          # Steps [6], [8]
    │   │   ├── laptop_dbta_preprocessed.json
    │   │   ├── laptop_dbta_curated.json                    ← ★ 23 topics
    │   │   └── shuffle_test/                               ← ★ 23 topics
    │   │       └── laptop_dbta_curated_shuffled_{0-4}.json
    │   └── sbta/                          # Steps [7], [9]
    │       ├── laptop_sbta_preprocessed.json
    │       ├── laptop_sbta_curated.json                    ← ★ 23 topics
    │       └── shuffle_test/                               ← ★ 23 topics
    │           └── laptop_sbta_curated_shuffled_{0-4}.json
    │
    └── restaurant/                        # Same structure as laptop/
        ├── raw_input/
        ├── dbta/
        └── sbta/

---

Step [1] — raw_input/{domain}_json.json

Base annotation: review text → aspect topic list

Files

semeval_stm/laptop/raw_input/laptop_json.json
semeval_stm/restaurant/raw_input/restaurant_json.json

Description
Starting point of the pipeline. Maps each full review text to the list of aspect topic labels contained in that review, derived from the original SemEval annotation. This index is used to build all downstream datasets.

Data Type: dict

Keys & Values

Field	Type	Description
key	str	Full review text
value	list[str]	Aspect topic labels present in the review (e.g. ["LAPTOP#CONNECTIVITY", "DISPLAY#QUALITY"])

Stats

Domain	# Reviews	Avg topics/review	Min	Max
laptop	395	~4.7	1	13
restaurant	335	~4.3	2	9

Example

{
  "Purchased as a gift for a friend. The service I received from Toshiba went above and beyond the call of duty. My friend reports the notebook is astonishing in performance, picture quality, and ease of use. It is extremely portable and easily connects to WIFI at the library and elsewhere. Just what the doctor ordered.": [
    "LAPTOP#CONNECTIVITY",
    "DISPLAY#QUALITY",
    "LAPTOP#USABILITY",
    "LAPTOP#GENERAL",
    "LAPTOP#OPERATION_PERFORMANCE",
    "LAPTOP#PORTABILITY",
    "SUPPORT#QUALITY"
  ]
}

---

Step [2] — raw_input/{domain}_segmentation_generation.json

LLM raw output: review text → JSON-encoded segmentation string

Files

semeval_stm/laptop/raw_input/laptop_segmentation_generation.json
semeval_stm/restaurant/raw_input/restaurant_segmentation_generation.json

Description
Raw output from an LLM that segments each review into topic-specific spans. The value is stored as a JSON-encoded string (not a parsed dict), preserving the model output before downstream parsing.

Data Type: dict

Keys & Values

Field	Type	Description
key	str	Full review text
value	str	JSON-encoded string of {topic_label: [segment, ...], ...}

Stats

Domain	# Entries
laptop	395
restaurant	335

Example (value after json.loads)

{
  "Purchased as a gift for a friend. The service I received from Toshiba went above and beyond the call of duty. My friend reports the notebook is astonishing in performance, picture quality, and ease of use. It is extremely portable and easily connects to WIFI at the library and elsewhere. Just what the doctor ordered.": "{\"LAPTOP#CONNECTIVITY\": [\"easily connects to WIFI at the library and elsewhere\"], \"DISPLAY#QUALITY\": [\"notebook is astonishing in performance, picture quality\"], \"LAPTOP#USABILITY\": [\"ease of use\"], \"LAPTOP#PORTABILITY\": [\"It is extremely portable\"]}"
}

After json.loads(value):

{
  "LAPTOP#CONNECTIVITY": ["easily connects to WIFI at the library and elsewhere"],
  "DISPLAY#QUALITY":     ["notebook is astonishing in performance, picture quality"],
  "LAPTOP#USABILITY":    ["ease of use"],
  "LAPTOP#PORTABILITY":  ["It is extremely portable"]
}

---

Step [3] — raw_input/{domain}_segmentation_document_wise_topic.json

Structured segmentation: one entry per review document

Files

semeval_stm/laptop/raw_input/laptop_segmentation_document_wise_topic.json
semeval_stm/restaurant/raw_input/restaurant_segmentation_document_wise_topic.json

Description
Structured (parsed) version of segmentation_generation.json. Same information, but the JSON-string value has been parsed into a native dict, and each entry is a self-contained object with base_text + quotations.

Data Type: list[dict]

Keys & Values

Field	Type	Description
base_text	str	Full original review text
quotations	dict[str, list[str]]	topic label → list of extracted text segments

Stats

Domain	# Documents	Avg topics/doc	Avg segs/doc
laptop	395	~5	~6
restaurant	335	~4	~5

Example

[
  {
    "base_text": "Purchased as a gift for a friend. The service I received from Toshiba went above and beyond the call of duty. My friend reports the notebook is astonishing in performance, picture quality, and ease of use. It is extremely portable and easily connects to WIFI at the library and elsewhere. Just what the doctor ordered.",
    "quotations": {
      "LAPTOP#CONNECTIVITY": ["easily connects to WIFI at the library and elsewhere"],
      "SUPPORT#QUALITY":     ["The service I received from Toshiba went above and beyond the call of duty"],
      "DISPLAY#QUALITY":     ["notebook is astonishing in performance, picture quality"],
      "LAPTOP#USABILITY":    ["ease of use"],
      "LAPTOP#PORTABILITY":  ["It is extremely portable"]
    }
  }
]

---

Step [4] — raw_input/{domain}_dbta.json

Raw DBTA: topic → full review documents (76 / 12 topics)

Files

semeval_stm/laptop/raw_input/laptop_dbta.json
semeval_stm/restaurant/raw_input/restaurant_dbta.json

Description
DBTA (Document-Based Topic Assignment) format. Each topic is mapped to all full review documents that mention it. A single review may appear under multiple topics. All topics from the original SemEval annotation are kept, including long-tail topics with very few documents.

Data Type: dict

Keys & Values

Field	Type	Description
key	str	Aspect topic label (e.g. "LAPTOP#CONNECTIVITY")
value	list[str]	Full review texts that contain this topic

Stats

Domain	# Topics	Total docs	Min docs/topic	Max docs/topic
laptop	76	1,982	1	395
restaurant	12	1,435	20	335

Example

{
  "LAPTOP#CONNECTIVITY": [
    "Purchased as a gift for a friend. The service I received from Toshiba went above and beyond the call of duty. My friend reports the notebook is astonishing in performance, picture quality, and ease of use. It is extremely portable and easily connects to WIFI at the library and elsewhere. Just what the doctor ordered.",
    "I enjoy very much my new Toshiba, purchased specifially for attending online schooling. I have found it very easy to use, very imformative, wonder alerts and tutorials making it very easy for someone like me who is not exactly technologically advanced to learn to use the various features and programs. I have found also, it is very easy to be able to access wireless internet access; without having to have extra equipment and/or complicated routes to take to be able to do so. I also love the small, convenient size of my laptop, making it a perfect tool for my academic studies. I am very pleased with my purchase!"
  ]
}

The same review text can appear under multiple topic keys (e.g. the first review also appears under DISPLAY#QUALITY, LAPTOP#USABILITY, etc.).

Laptop topics (76)

LAPTOP#CONNECTIVITY, DISPLAY#QUALITY, LAPTOP#USABILITY, LAPTOP#GENERAL, LAPTOP#OPERATION_PERFORMANCE, LAPTOP#PORTABILITY, SUPPORT#QUALITY, SOFTWARE#GENERAL, MULTIMEDIA_DEVICES#GENERAL, LAPTOP#DESIGN_FEATURES, SOFTWARE#USABILITY, SOFTWARE#DESIGN_FEATURES, DISPLAY#DESIGN_FEATURES, OS#USABILITY, COMPANY#GENERAL, SHIPPING#QUALITY, KEYBOARD#DESIGN_FEATURES, BATTERY#OPERATION_PERFORMANCE, LAPTOP#PRICE, HARD_DISC#QUALITY, LAPTOP#QUALITY, SUPPORT#PRICE, KEYBOARD#QUALITY, KEYBOARD#GENERAL, OS#GENERAL, MOUSE#OPERATION_PERFORMANCE, MOUSE#USABILITY, MEMORY#DESIGN_FEATURES, PORTS#DESIGN_FEATURES, OS#QUALITY, OS#OPERATION_PERFORMANCE, GRAPHICS#GENERAL, PORTS#QUALITY, HARDWARE#QUALITY, CPU#GENERAL, CPU#OPERATION_PERFORMANCE, HARDWARE#GENERAL, KEYBOARD#USABILITY, MOUSE#DESIGN_FEATURES, MOTHERBOARD#QUALITY, POWER_SUPPLY#QUALITY, KEYBOARD#OPERATION_PERFORMANCE, POWER_SUPPLY#DESIGN_FEATURES, SOFTWARE#OPERATION_PERFORMANCE, BATTERY#QUALITY, MOUSE#QUALITY, FANS_COOLING#DESIGN_FEATURES, MULTIMEDIA_DEVICES#QUALITY, MULTIMEDIA_DEVICES#USABILITY, SHIPPING#PRICE, DISPLAY#OPERATION_PERFORMANCE, WARRANTY#PRICE, GRAPHICS#QUALITY, HARD_DISC#DESIGN_FEATURES, DISPLAY#GENERAL, MULTIMEDIA_DEVICES#OPERATION_PERFORMANCE, MOUSE#GENERAL, SOFTWARE#PRICE, POWER_SUPPLY#OPERATION_PERFORMANCE, WARRANTY#GENERAL, HARDWARE#OPERATION_PERFORMANCE, OS#DESIGN_FEATURES, GRAPHICS#DESIGN_FEATURES, POWER_SUPPLY#GENERAL, DISPLAY#USABILITY, MULTIMEDIA_DEVICES#DESIGN_FEATURES, OPTICAL_DRIVES#QUALITY, SOFTWARE#QUALITY, SUPPORT#GENERAL, CPU#DESIGN_FEATURES, FANS_COOLING#QUALITY, CPU#QUALITY, FANS_COOLING#OPERATION_PERFORMANCE, OPTICAL_DRIVES#OPERATION_PERFORMANCE, PORTS#OPERATION_PERFORMANCE, BATTERY#GENERAL

Restaurant topics (12)

RESTAURANT#GENERAL, SERVICE#GENERAL, FOOD#QUALITY, FOOD#STYLE_OPTIONS, DRINKS#STYLE_OPTIONS, DRINKS#PRICES, RESTAURANT#PRICES, AMBIENCE#GENERAL, RESTAURANT#MISCELLANEOUS, LOCATION#GENERAL, FOOD#PRICES, DRINKS#QUALITY

---

Step [5] — raw_input/{domain}_sbta.json

Raw SBTA: topic → text segments (76 / 12 topics)

Files

semeval_stm/laptop/raw_input/laptop_sbta.json
semeval_stm/restaurant/raw_input/restaurant_sbta.json

Description
SBTA (Segment-Based Topic Assignment) format. Each topic is mapped to the list of fine-grained text segments (sub-sentences or spans) extracted by the LLM in Step [2]. Unlike DBTA, only the directly relevant portion of each review is stored, not the full document.

Data Type: dict

Keys & Values

Field	Type	Description
key	str	Aspect topic label
value	list[str]	Short text segments relevant to this topic

Stats

Domain	# Topics	Total segs	Min segs/topic	Max segs/topic
laptop	76	2,287	1	541
restaurant	12	1,837	21	497

(Same topic set as the domain's _dbta.json)

Example

{
  "LAPTOP#CONNECTIVITY": [
    "easily connects to WIFI at the library and elsewhere",
    "I have found also, it is very easy to be able to access wireless internet access; without having to have extra equipment and/or complicated routes to take to be able to do so.",
    "It does not keep internet signals no matter where you bring it."
  ]
}

---

Step [6] — dbta/{domain}_dbta_preprocessed.json

Filtered DBTA: remove long-tail topics (32 / 12 topics)

Files

semeval_stm/laptop/dbta/laptop_dbta_preprocessed.json
semeval_stm/restaurant/dbta/restaurant_dbta_preprocessed.json

Description
Generated by src/preprocess.py. Topics with ≤ 10 documents are removed from the raw DBTA (Step [4]). Only topics with statistically sufficient document counts are retained. Laptop topics are reduced from 76 → 32; restaurant topics remain at 12.

Data Type: dict

Keys & Values

Field	Type	Description
key	str	Aspect topic label
value	list[str]	Full review texts for that topic

Stats

Domain	# Topics	Total docs	Min docs/topic	Max docs/topic
laptop	32	1,825	11	395
restaurant	12	1,435	20	335

Laptop topics (32)

LAPTOP#CONNECTIVITY, DISPLAY#QUALITY, LAPTOP#USABILITY, LAPTOP#GENERAL, LAPTOP#OPERATION_PERFORMANCE, LAPTOP#PORTABILITY, SUPPORT#QUALITY, SOFTWARE#GENERAL, LAPTOP#DESIGN_FEATURES, SOFTWARE#USABILITY, DISPLAY#DESIGN_FEATURES, OS#USABILITY, COMPANY#GENERAL, KEYBOARD#DESIGN_FEATURES, BATTERY#OPERATION_PERFORMANCE, LAPTOP#PRICE, HARD_DISC#QUALITY, LAPTOP#QUALITY, KEYBOARD#QUALITY, KEYBOARD#GENERAL, OS#GENERAL, MOUSE#OPERATION_PERFORMANCE, MOUSE#USABILITY, MEMORY#DESIGN_FEATURES, GRAPHICS#GENERAL, CPU#OPERATION_PERFORMANCE, KEYBOARD#USABILITY, MOUSE#DESIGN_FEATURES, SOFTWARE#OPERATION_PERFORMANCE, MULTIMEDIA_DEVICES#QUALITY, HARD_DISC#DESIGN_FEATURES, DISPLAY#GENERAL

Example

{
  "LAPTOP#CONNECTIVITY": [
    "Purchased as a gift for a friend. The service I received from Toshiba went above and beyond the call of duty. My friend reports the notebook is astonishing in performance, picture quality, and ease of use. It is extremely portable and easily connects to WIFI at the library and elsewhere. Just what the doctor ordered.",
    "I enjoy very much my new Toshiba, purchased specifially for attending online schooling. I have found it very easy to use, very imformative, wonder alerts and tutorials making it very easy for someone like me who is not exactly technologically advanced to learn to use the various features and programs. I have found also, it is very easy to be able to access wireless internet access; without having to have extra equipment and/or complicated routes to take to be able to do so. I also love the small, convenient size of my laptop, making it a perfect tool for my academic studies. I am very pleased with my purchase!"
  ]
}

Same schema as Step [4]. Topics with ≤ 10 docs have been removed (44 topics dropped for laptop).

---

Step [7] — sbta/{domain}_sbta_preprocessed.json

Filtered SBTA: remove long-tail topics (33 / 12 topics)

Files

semeval_stm/laptop/sbta/laptop_sbta_preprocessed.json
semeval_stm/restaurant/sbta/restaurant_sbta_preprocessed.json

Description
Same threshold filtering as Step [6] (≤ 10 segments → drop), applied to the raw SBTA (Step [5]). Laptop reduces from 76 → 33 topics; restaurant stays at 12. Note: SBTA retains one extra topic (DISPLAY#OPERATION_PERFORMANCE) compared to DBTA's 32, as its segment count exceeds the threshold.

Data Type: dict

Keys & Values

Field	Type	Description
key	str	Aspect topic label
value	list[str]	Text segments for that topic

Stats

Domain	# Topics	Total segs	Min segs/topic	Max segs/topic
laptop	33	2,133	11	541
restaurant	12	1,837	21	497

Laptop topics (33) — 32 same as DBTA + DISPLAY#OPERATION_PERFORMANCE

LAPTOP#CONNECTIVITY, DISPLAY#QUALITY, LAPTOP#USABILITY, LAPTOP#GENERAL, LAPTOP#OPERATION_PERFORMANCE, LAPTOP#PORTABILITY, SUPPORT#QUALITY, SOFTWARE#GENERAL, LAPTOP#DESIGN_FEATURES, SOFTWARE#USABILITY, DISPLAY#DESIGN_FEATURES, OS#USABILITY, COMPANY#GENERAL, KEYBOARD#DESIGN_FEATURES, BATTERY#OPERATION_PERFORMANCE, LAPTOP#PRICE, HARD_DISC#QUALITY, LAPTOP#QUALITY, KEYBOARD#QUALITY, KEYBOARD#GENERAL, OS#GENERAL, MOUSE#OPERATION_PERFORMANCE, MOUSE#USABILITY, MEMORY#DESIGN_FEATURES, GRAPHICS#GENERAL, CPU#OPERATION_PERFORMANCE, KEYBOARD#USABILITY, MOUSE#DESIGN_FEATURES, SOFTWARE#OPERATION_PERFORMANCE, MULTIMEDIA_DEVICES#QUALITY, DISPLAY#OPERATION_PERFORMANCE, HARD_DISC#DESIGN_FEATURES, DISPLAY#GENERAL

Example

{
  "LAPTOP#CONNECTIVITY": [
    "easily connects to WIFI at the library and elsewhere",
    "I have found also, it is very easy to be able to access wireless internet access; without having to have extra equipment and/or complicated routes to take to be able to do so.",
    "It does not keep internet signals no matter where you bring it."
  ]
}

Same schema as Step [5]. Topics with ≤ 10 segments have been removed (43 topics dropped for laptop).

---

[6] → [8] and [7] → [9]: How the Final 23-Topic Datasets Were Built

Overview

Steps [8] and [9] converge to an identical set of 23 laptop / 11 restaurant topics, enabling a controlled comparison between DBTA and SBTA.

[6] dbta_preprocessed (32/12) ──────────────────────────────► [8] dbta_curated (23/11)
                                  substring-match via SBTA
                                  segment as lookup key

[7] sbta_preprocessed (33/12) ──── manual quality review ───► [9] sbta_curated (23/11)

---

[7] → [9]: SBTA Manual Curation

Starting from sbta_preprocessed (Step [7]), a human reviewer manually:

1. Reallocated segments incorrectly assigned to a topic by the LLM (moved to the correct topic).
2. Removed noise segments that did not belong to any topic.
3. Merged or dropped topics with insufficient clean segments after curation.

Laptop topics dropped (10): 9 shared with DBTA + DISPLAY#OPERATION_PERFORMANCE (SBTA-only)

Dropped Topic	Reason
LAPTOP#GENERAL	Too semantically broad; segments absorbed by more specific topics
SOFTWARE#GENERAL	Too semantically broad; segments absorbed by more specific topics
OS#GENERAL	Too semantically broad; merged into OS#USABILITY etc.
KEYBOARD#QUALITY	Low segment count after cleaning; merged into KEYBOARD#USABILITY
KEYBOARD#GENERAL	Low segment count after cleaning; merged into KEYBOARD#DESIGN_FEATURES
MOUSE#OPERATION_PERFORMANCE	Low segment count; merged into MOUSE#USABILITY
SOFTWARE#OPERATION_PERFORMANCE	Low segment count; merged into SOFTWARE#USABILITY
DISPLAY#GENERAL	Low segment count; merged into DISPLAY#QUALITY or DISPLAY#DESIGN_FEATURES
HARD_DISC#DESIGN_FEATURES	Low segment count; merged into HARD_DISC#QUALITY
DISPLAY#OPERATION_PERFORMANCE	SBTA-only topic; insufficient clean segments after manual review

Restaurant topics dropped (1):

Dropped Topic	Reason
RESTAURANT#MISCELLANEOUS	Too vague/catch-all; segments could not be reliably assigned

---

[6] → [8]: DBTA Alignment via Substring Matching

# Step 1. Collect all unique full review texts from dbta_preprocessed
dbta_values = list(set(sum(list(dbta_preprocessed.values()), [])))

# Step 2. For each topic and segment in sbta_curated,
#          find DBTA reviews that contain the segment as a substring
dbta_curated = defaultdict(list)
for topic, segments in sbta_curated.items():
    for segment in segments:
        for dbta_review in dbta_values:
            if segment in dbta_review:          # substring containment
                dbta_curated[topic].append(dbta_review)

# Step 3. Deduplicate reviews per topic
for topic in dbta_curated:
    dbta_curated[topic] = list(set(dbta_curated[topic]))

This ensures that:

• Topic set is inherited from sbta_curated (23/11 topics).
• Documents are linked back from SBTA segments to their original full reviews.
• Topics dropped in [7]→[9] are automatically absent in [8] as well (9 shared drops for laptop, 1 for restaurant).

Laptop topics dropped (9) — identical drop set shared with SBTA (minus DISPLAY#OPERATION_PERFORMANCE which never existed in DBTA preprocessed):

LAPTOP#GENERAL, SOFTWARE#GENERAL, OS#GENERAL, KEYBOARD#QUALITY, KEYBOARD#GENERAL, MOUSE#OPERATION_PERFORMANCE, SOFTWARE#OPERATION_PERFORMANCE, DISPLAY#GENERAL, HARD_DISC#DESIGN_FEATURES

---

Final Topic Sets

Both dbta_curated and sbta_curated share an identical topic set within each domain.

#	Laptop (23)	Restaurant (11)
1	BATTERY#OPERATION_PERFORMANCE	AMBIENCE#GENERAL
2	COMPANY#GENERAL	DRINKS#PRICES
3	CPU#OPERATION_PERFORMANCE	DRINKS#QUALITY
4	DISPLAY#DESIGN_FEATURES	DRINKS#STYLE_OPTIONS
5	DISPLAY#QUALITY	FOOD#PRICES
6	GRAPHICS#GENERAL	FOOD#QUALITY
7	HARD_DISC#QUALITY	FOOD#STYLE_OPTIONS
8	KEYBOARD#DESIGN_FEATURES	LOCATION#GENERAL
9	KEYBOARD#USABILITY	RESTAURANT#GENERAL
10	LAPTOP#CONNECTIVITY	RESTAURANT#PRICES
11	LAPTOP#DESIGN_FEATURES	SERVICE#GENERAL
12	LAPTOP#OPERATION_PERFORMANCE
13	LAPTOP#PORTABILITY
14	LAPTOP#PRICE
15	LAPTOP#QUALITY
16	LAPTOP#USABILITY
17	MEMORY#DESIGN_FEATURES
18	MOUSE#DESIGN_FEATURES
19	MOUSE#USABILITY
20	MULTIMEDIA_DEVICES#QUALITY
21	OS#USABILITY
22	SOFTWARE#USABILITY
23	SUPPORT#QUALITY

---

Step [8] — dbta/{domain}_dbta_curated.json

Primary DBTA evaluation dataset — 23 / 11 topics

Files

semeval_stm/laptop/dbta/laptop_dbta_curated.json
semeval_stm/restaurant/dbta/restaurant_dbta_curated.json

Description
Final curated DBTA dataset. Derived from dbta_preprocessed (Step [6]) via substring matching against sbta_curated segments (Step [9]). Only reviews containing at least one SBTA segment as a substring are retained, and the topic set is inherited directly from sbta_curated (23 laptop / 11 restaurant topics).

Data Type: dict

Keys & Values

Field	Type	Description
key	str	Aspect topic label (23 for laptop, 11 for restaurant)
value	list[str]	Full review texts for that topic

Stats

Domain	# Topics	Total docs	Min docs/topic	Max docs/topic
laptop	23	1,611	8	386
restaurant	11	1,344	16	332

Example

{
  "LAPTOP#CONNECTIVITY": [
    "This computer was awful! I will never buy another computer from Dell ever again do to how awful it worked and how I was treated by the company. The computer first started malfunctioning 3 months after purchasing it. It does not keep internet signals no matter where you bring it. The system it comes with does not work properly, so when trying to fix the problems with it it started not working at all. When calling Dell for help, reurn, or a new computer they were not useful and left it up to myself to figure out what to do with it.",
    "This computer periodically crashes (Office, IPhoto, Safari). It takes a long time to get on the Web and to change Web Sites. Locks up she searching for updates. Just to be clear. I have been an Apple user since they switched to Intel. I have had Macbook PROs and Macbook Airs and have never had a problem. It seems that this computer might be a lemon. I purchased it through Amazon."
  ]
}

Each review here was retrieved because it contains at least one SBTA segment for LAPTOP#CONNECTIVITY as a substring (e.g. "It does not keep internet signals no matter where you bring it.").

---

Step [8] — dbta/shuffle_test/{domain}_dbta_curated_shuffled_{0-4}.json

Shuffled copies for multi-seed evaluation — 23 / 11 topics

Files

semeval_stm/{domain}/dbta/shuffle_test/{domain}_dbta_curated_shuffled_{0,1,2,3,4}.json

Description
Five independently shuffled copies of {domain}_dbta_curated.json. Generated by globally pooling all documents across all topics, then randomly re-distributing them back to topics while preserving the original per-topic document count. As a result, the document count per topic is identical to the original, but the topic-to-document assignment is intentionally scrambled. Used as a random baseline in topic modeling evaluation.

Stats: Identical to {domain}_dbta_curated.json — laptop 23 topics, 1,611 docs; restaurant 11 topics, 1,344 docs.

Example (shuffled_0 — same per-topic document count as above, but topic-to-document assignment is globally scrambled)

{
  "LAPTOP#CONNECTIVITY": [
    "Like the price and operation. Some features arent friendly ( volume wheel, sound quality, etc. But, for the cost this is a winner.",
    "It was secure and easy to navigate. It is everything that I would want in a computer. Besides the great look, it is a great machine. I would definitely not go back to using a PC after using the mac. Once again apple stands alone at the top of the charts."
  ]
}

Per-topic document count is identical to dbta_curated, but documents are drawn from the global pool — a given document may appear under a different topic than in the original.

---

Step [9] — sbta/{domain}_sbta_curated.json

Primary SBTA evaluation dataset — 23 / 11 topics

Files

semeval_stm/laptop/sbta/laptop_sbta_curated.json
semeval_stm/restaurant/sbta/restaurant_sbta_curated.json

Description
Final curated SBTA dataset produced by manual quality review of sbta_preprocessed (Step [7]). Segments misassigned by the LLM were corrected or removed, and topics with insufficient clean segments were dropped, resulting in 23 laptop / 11 restaurant topics.

Data Type: dict

Keys & Values

Field	Type	Description
key	str	Aspect topic label
value	list[str]	Manually verified text segments for that topic

Stats

Domain	# Topics	Total segs	Min segs/topic	Max segs/topic
laptop	23	2,123	11	700
restaurant	11	1,767	16	495

Example

{
  "LAPTOP#CONNECTIVITY": [
    "easily connects to WIFI at the library and elsewhere",
    "I have found also, it is very easy to be able to access wireless internet access; without having to have extra equipment and/or complicated routes to take to be able to do so.",
    "It does not keep internet signals no matter where you bring it."
  ]
}

Segments that were misassigned by the LLM have been corrected or removed. Each string is a short, precise span directly relevant to the topic.

---

Step [9] — sbta/shuffle_test/{domain}_sbta_curated_shuffled_{0-4}.json

Shuffled copies for multi-seed evaluation — 23 / 11 topics

Files

semeval_stm/{domain}/sbta/shuffle_test/{domain}_sbta_curated_shuffled_{0,1,2,3,4}.json

Description
Generated by src/shuffle_preprocessed.py. Five independently shuffled copies of {domain}_sbta_curated.json. All segments across all topics are globally pooled, then randomly re-distributed back to topics while preserving the original per-topic segment count. As a result, segment counts per topic are identical to the original, but the topic-to-segment assignment is intentionally scrambled. Used as a random baseline in topic modeling evaluation.

Stats: Identical to {domain}_sbta_curated.json — laptop 23 topics, 2,123 segs; restaurant 11 topics, 1,767 segs.

Example (shuffled_0 — same per-topic segment count as above, but topic-to-segment assignment is globally scrambled)

{
  "LAPTOP#CONNECTIVITY": [
    "I have owned macbook pros since 2005 and this is the best yet!",
    "I am also upset with CR for giving a good rating",
    "Also the speakers are not very loud"
  ]
}

Per-topic segment count is identical to sbta_curated, but segments are drawn from the global pool — a given segment may appear under a different topic than in the original.

---

Appendix — SIT_TASK_DATA/*_sit_dataset_250517.json

Sentence Intruder Task

Files

SIT_TASK_DATA/laptop_sit_dataset_250517.json
SIT_TASK_DATA/restaurant_sit_dataset_250517.json

Description
Sentence Intruder Task (SIT) dataset derived from the curated topic data. Given 6 candidate sentences (5 from the same topic + 1 intruder from a different topic), the model must identify the intruder. Four difficulty splits vary intruder topic proximity and intruder count.

Data Type: dict

Keys & Values

Field	Type	Description
"hard_single"	list[dict]	Intruder from a closely related topic; 1 intruder
"hard_double"	list[dict]	Intruder from a closely related topic; 2 intruders
"easy_single"	list[dict]	Intruder from a distant topic; 1 intruder
"easy_double"	list[dict]	Intruder from a distant topic; 2 intruders

Each item in a split:

Field	Type	Description
topic	str	Aspect topic label
input_texts	list[str]	6 candidate sentences
intruder_idx	int	Index of the intruder sentence in input_texts

Stats

Domain	Split	# Items
laptop	hard_single	200
laptop	hard_double	200
laptop	easy_single	200
laptop	easy_double	200
restaurant	hard_single	200
restaurant	hard_double	200
restaurant	easy_single	200
restaurant	easy_double	200

Example

{
  "hard_single": [
    {
      "topic": "LAPTOP#CONNECTIVITY",
      "input_texts": [
        "The WiFi is very weak",
        "the laptop does not have an ethernet port for cable connecting",
        "Needed a 2 usb adapt. for network plug, no problem.",
        "the built-in mouse pad",
        "It has no way of connecting to the internet save wifi",
        "The internet speed is spectacular."
      ],
      "intruder_idx": 3
    }
  ]
}

Index 3 ("the built-in mouse pad") is the intruder — it belongs to LAPTOP#USABILITY, not LAPTOP#CONNECTIVITY.