RETECO · SemEval-2027 Task 1

Evaluation plan

A single official retrieval metric keeps the competition interpretable; targeted diagnostics expose temporal, conversational, and generative behavior.

Official leaderboard

Retrieval ranks by nDCG@10.

For both tracks, binary normalized Discounted Cumulative Gain at rank 10 is the official retrieval ranking metric. It rewards relevant documents appearing early while normalizing for the number of available relevant documents.

Primary metricDCG@10 = Σ(i=1..10) relᵢ / log₂(i + 1) nDCG@10 = DCG@10 / IDCG@10
No arbitrary compositeRETECO does not combine retrieval and generation into a single leaderboard number. Generation quality is reported alongside retrieval where applicable.
Sub-trackOfficial rankingAdditional reporting
1a Temporal retrievalQuery-level nDCG@10Temporal + standard diagnostics
1b Step-wise retrievalStep-level nDCG@10Temporal + standard diagnostics
2a Conversational retrievalTurn-level nDCG@10Turn depth + domain diagnostics
2b Gold-passage generationFive judge dimensions
2c Full conversational RAGTurn-level nDCG@10Five judge dimensions

Track 01 scoring

TEMPO retrieval and temporal diagnostics

Aggregation

For Sub-track 1a, nDCG@10 is computed independently for each query and macro-averaged. For Sub-track 1b, step-specific nDCG@10 values are first averaged over the supplied steps for each query, then aggregated across queries.

Temporal diagnostics

  • Temporal Precision@10: position-sensitive precision that rewards temporally relevant documents ranked earlier.
  • Temporal Relevance@10: the fraction of the top ten results judged temporally relevant.
  • Temporal Coverage@10: the fraction of required time periods covered by at least one top-ten document.
  • nDCG|FC@10: nDCG@10 over the subset of queries achieving full temporal coverage.

MAP@10, Recall@10, MRR, and per-domain results are also reported diagnostically.

Frozen temporal judgmentsTemporal relevance and period-coverage labels are created before leaderboard scoring using documented prompts calibrated with golden cases. The labels are human-checked, frozen, cached once, and reused for every run.

Track 02 scoring

RECOR retrieval diagnostics

For Sub-tracks 2a and 2c, nDCG@10 is computed per target turn and macro-averaged for official ranking. Diagnostic reporting includes MAP@10, Recall@10, MRR, per-domain results, and conversation-depth groups:

T1T2T3T4T5+

Turn-depth analysis helps distinguish single-turn retrieval quality from the ability to retain and resolve accumulated dialogue context.

Sub-tracks 2b & 2c

Five focused generation judgments

Each generated response is assessed on five independent dimensions with separate focused prompts and a fixed 1–5 Likert rubric.

DimensionQuestion answered
CorrectnessAre the response’s factual claims accurate?
CompletenessDoes it cover the information needed by the turn?
RelevanceDoes it directly address the user’s current need?
Conversational coherenceDoes it fit the preceding dialogue?
FaithfulnessIs it supported by supplied or retrieved passages?

Each integer score is normalized as (score − 1) / 4 and macro-averaged for reporting. ROUGE-L, METEOR, and BERTScore are additional diagnostics.

Track 2c evidence boundaryParticipants provide a ranking long enough to compute nDCG@10 and generate the answer using their top five retrieved passages.

Calibration & audit

Judge behavior is fixed before evaluation.

The five prompts, rubric anchors, GPT-4o version, decoding settings, and output schema are frozen before the competition. Golden-case meta-evaluation checks whether the judge distinguishes supported from unsupported, complete from incomplete, and coherent from incoherent responses.

Judge outputs are spot-validated against independent human ratings on the SemEval test set. The task report will disclose the sample, agreement results, prompts, judge version, retry procedure, and invalid or missing outputs.

Pilot validationThe RECOR pilot used three PhD annotators on 200 conversations; automatic scores tracked human ratings within +0.14 to +0.35 on the 1–5 scale.

Deterministic handling

How submitted runs are interpreted

Retrieval runs use the standard six-column TREC format:

Retrieval submissionquery_id Q0 document_id rank score run_id
  • Missing queries, steps, or target turns receive zero.
  • Duplicate document identifiers are removed after their first occurrence.
  • Unknown document identifiers are discarded.
  • Official retrieval must use the organizer-provided corpus.
  • External corpora may not replace or augment the official corpus for official ranking.

Generation output is UTF-8 JSONL containing a target turn_id and answer. Exact validation utilities and sample submissions will be released before evaluation.

Format checker · coming soonLocal scorer · coming soonParticipation rules →