| --- |
| language: |
| - tr |
| - en |
| - de |
| - ru |
| license: other |
| task_categories: |
| - text-classification |
| pretty_name: HOSIA-Intent |
| size_categories: |
| - 1K<n<10K |
| --- |
| |
| # HOSIA-Intent — synthetic hospitality intent data (draft) |
|
|
| This repository contains synthetic hospitality-intent data generated for HOSIA. It is an unreviewed working draft, not a training-ready release or a benchmark. |
|
|
| ## Status |
|
|
| - `hosia_v0.3_multilingual_ai_reviewed.jsonl`: 2,000 records covering all 20 ontology intents × 25 scenarios × TR/EN/DE/RU. Each intent-language pair has 25 records. There are 500 scenario groups, kept intact across train/validation/test (1,680/160/160). |
| - `hosia_v0.2_hard_negatives.jsonl`: separate 80-record contrast challenge set with `split: hard_negative_eval`; it covers only the original 10 intents and is excluded from the main dataset. |
| - `hosia_v0.2_multilingual_ai_reviewed.jsonl`: the original 1,000-record, 10-intent draft, kept as a versioned snapshot. |
| - `hosia_v0.2_intents.json`: the canonical intent definitions aligned to the 20 ontology intent concepts. |
| - `tfidf_baseline_report.md`: development-only word/character TF-IDF + LinearSVC report on v0.2. Main test accuracy 0.988, macro-F1 0.987; hard-negative foil accuracy 0.613. These results describe only the earlier synthetic pilot and are not generalizable benchmark claims for v0.3. |
| - EVREN Gemma/Qwen models assisted with candidate generation and translation; DeepSeek v4.1 primarily checked intents/entities and translations. No human review has been completed. |
| - In a blind second opinion, the three currently available EVREN models had 2-of-3 majority agreement with existing labels on 80/80 challenge examples; all three agreed on 74/80. This is model agreement only, not human ground truth; six examples need adjudication. |
| - Every record remains `reviewed: false` and `allowed_for_training: false`. AI checks are not human review. |
| - A basic email and phone-pattern scan found no matches. It is not a full privacy review. |
|
|
| ## Dataset fields |
|
|
| `id`, `language`, `domain`, `user_message`, `intent`, `entities`, `department`, `action`, `information_dependency`, `contains_pii`, `allowed_for_training`, `difficulty`, `scenario_group`, `source`, `generator_model`, `validator_model`, `ai_reviewed`, `ai_review_status`, `reviewed`, and `split`. |
|
|
| ## Intended use and limitations |
|
|
| The planned research concerns multilingual hospitality intent classification, entity extraction, department routing, action selection, and information dependency. This draft is for internal review and iteration only. Do not use it for training, evaluation claims, operational decisions, or downstream redistribution. |
|
|
| Synthetic LLM output and model review can contain wrong labels, awkward or culturally mismatched language, annotation omissions, semantic duplicates, and personal or sensitive information. Basic pattern screening cannot detect every privacy risk. The baseline uses a small synthetic test split and should not be read as evidence of real-world performance. All examples, entities, translations, splits, privacy, and rights need human review before any training or benchmark use. |
|
|
| The hard-negative file and its labels are public in this draft; it is not an undisclosed holdout and must not be used to claim benchmark performance. A future benchmark needs a separately reviewed, properly held-out challenge set. |
|
|
| The EVREN dashboard currently does not enable text dataset upload for this account; EVREN was used for generation and validation, not as a dataset host. |
|
|
| ## License and provenance |
|
|
| All rights reserved. No permission is granted to copy, modify, redistribute, train on, or otherwise use this dataset. Prior written permission from the HOSIA copyright holder is required. Examples were generated synthetically for HOSIA; no hotel corpus or real guest conversations were used as generation input. |
|
|
| ## Next steps |
|
|
| 1. Human-review labels, entities, language quality, challenge labels, and scenario-group splits across all 20 intents. |
| 2. Create and review hard-negative examples for the expanded intents; keep any public challenge set out of claims as a hidden benchmark. |
| 3. Complete privacy, sensitive-content, and rights review. |
| 4. Consider a training-ready or benchmark release only after review and explicit release approval. |
|
|
