NTCIR19-Lifelog Task

Lifelog Task Formal Run Topics

The Lifelog Semantic Access Task (LSAT) formal run topics are now available for download. These topics will be used for the final evaluation.

Status: Released (April 20, 2026)

Topics File: formal-run-topics.xml

The formal run includes 33 topics covering two types:

Ad-Hoc Queries (16 topics)

• AD26001–AD26016: 16 Ad-Hoc topics

Known-Item Search (17 topics)

• KIS26001–KIS26017: 17 Known-Item Search topics

Topic Format: Each topic in the XML file includes:

• ID: Unique topic identifier (AD26xxx for Ad-Hoc, KIS26xxx for Known-Item)
• Type: adhoc or knownitem
• Title: Brief topic title
• Description: Description of what to find
• Narrative: Detailed explanation of relevance criteria

CASTLE Task Formal Run Topics

The CASTLE 2024 dataset is a large-scale, multimodal collection of time-aligned video, audio, and sensor data. It consists of 15 UHD 50fps video streams (10 egocentric cameras worn by participants and 5 static cameras) recorded over four days by 10 participants, with audio and automatically generated speech transcripts. Formal run topics for CASTLE tasks are being released during the formal run phase. The CASTLE 2024 dataset is available on Hugging Face (full CASTLE2024 and audio-only repositories).

CSAT - CASTLE Semantic Access subTask

The CASTLE Semantic Access Task (CSAT) is a search-based task where participants retrieve key interactions or events from multimodal collaborative session data (video, audio, and speech transcripts) based on a semantic query.

Status: Released (April 20, 2026)

Topics File: castle-semantic-topics-formal-run.xml

The CSAT formal run includes 25 topics (CSAT26001–CSAT26025) covering various semantic queries such as finding instances of playing guitar, taking down a christmas tree, people singing, playing board games, and more.

Topic Format: Each CSAT topic includes:

• ID: Unique topic identifier (CSAT26001–CSAT26025)
• Type: semantic
• Title: Brief topic title
• Description: Description of what to find
• Narrative: Detailed explanation of relevance criteria

CAST-Seg - CASTLE Conversation Segmentation subTask

CAST-Seg asks participants to divide the CASTLE 2024 collaborative sessions into meaningful conversational units (e.g., task, discussion, decision) as contiguous, non-overlapping time spans. There is no separate topic file; the task is defined by the dataset below.

Status: Available. The task is open for the formal run.

Data: Four days of time-aligned ego- and exo-centric video and audio (CASTLE dataset), plus Whisper speech transcripts. The audio-only release is sufficient for transcript- or audio-based approaches. Sessions include workshops, meal preparation and dining, the evening "Happy Quiz", and leisure activities.

Output: For each (DAY, SOURCE) session, submit ordered segments with start/end time (HH:MM:SS) and a label. See the Submission page for the file format.

Submission: One file per run (GroupID-RunID-CASTSeg.txt) to ntcir.lifelog.dcu@gmail.com, subject line 'NTCIR-Lifelog CAST-Seg Submission'. Up to 10 runs per group; deadline August 1, 2026.

Evaluation: No ground truth; assessors review boundaries and labels for coherence. Validated submissions will be released as a community benchmark for future research. We encourage all CASTLE participants, including CSAT-only teams, to submit at least one run.

Recipe Generation

The Recipe Generation subtask challenges participants to reconstruct personal recipes (the ingredients and the step-by-step cooking procedure) for the dishes that participants prepared during the CASTLE collaborative cooking sessions. Systems combine the multimodal session data (ego- and exo-centric video, audio, and speech transcripts) with the supporting menu schedule and cookbook to produce a coherent recipe for each target dish. The task is defined entirely by the data below; there is no separate topic file to release.

Status: Available. The task is open for the formal run.

Target dishes: The dishes to generate recipes for are those in the menu schedule: Irish breakfast (fry), Shepherd's pie, Tagliatelle, Pumpkin risotto with rocket, Zürchergeschnetzeltes & Rösti, Lẩu gà, Bò kho, Italian spinach lasagne, Chicken curry, Pineapple fried rice, and Swiss Christmas biscuits (Mailänderli, Brunsli, Zimtsterne).

Provided resources: The supporting files below are available on request from the organizers by email at ntcir.lifelog.dcu@gmail.com.

• Menu schedule (cooking rota): a spreadsheet (XLSX) with one row per day, listing which participant cooked the breakfast, lunch, and dinner on each recording day and the name of each dish prepared.
• Castle Cookbook: a spreadsheet (XLSX) with per-dish ingredient lists and quantities (scaled for the group), compiled from the participants' preparation notes and shopping lists, organised by dish.
• CASTLE multimodal recordings: the video, audio, and speech transcripts covering the cooking activities (see the CASTLE dataset).

Task output: For each target dish, generate a structured recipe consisting of (i) a title, (ii) an ingredient list with quantities, and (iii) an ordered sequence of preparation and cooking steps, grounded in what is observed in the session data.

Submission: Submit one file per run containing the generated recipe for every target dish. Name the file GroupID-RunID-RecipeGen.txt and send it to ntcir.lifelog.dcu@gmail.com with the subject line 'NTCIR-Lifelog Recipe Generation Submission'. Each group may submit up to 10 runs, and submissions are due by the formal-run deadline (August 1, 2026).

Evaluation: There is no ground truth for this task. Generated recipes are reviewed manually by human assessors for whether the ingredients and steps are plausible and consistent with the cooking sessions.