GOAL Dataset: Diverse Benchmarks

• GOAL Dataset is a term that unites multiple benchmarks including multilingual script learning, sports game summarization, and knowledge-grounded soccer video captioning.
• The script learning benchmark supports 18 languages with over 444,000 articles and employs generation and retrieval evaluations using metrics like accuracy and Kendall’s tau.
• GOAL-style benchmarks in automated planning generate multiple plan variants to reduce planner bias, while careful disambiguation distinguishes them from similarly named resources.

The expression GOAL Dataset is used in the literature for several distinct resources rather than a single canonical corpus. The name appears most prominently in three benchmark families: a multilingual wikiHow-derived dataset for Goal-Oriented Script Construction, an English benchmark for sports game summarization, and a soccer benchmark for knowledge-grounded video captioning. In automated planning, closely related usage appears in GOAL-style goal-recognition benchmarks, where recent work has revised dataset construction to reduce planner-induced bias (Lyu et al., 2021, Wang et al., 2022, Qi et al., 2023, Abdelwahed et al., 16 Feb 2026).

1. Nomenclature and principal uses

The ambiguity of the term is best understood as a consequence of independent research communities reusing the word “goal” for different formal objects: procedural intentions, task-conditioned summaries, commentary generation, and goal hypotheses in planning.

Usage	Domain	Core characterization
GOAL (Lyu et al., 2021)	Script learning	Multilingual wikiHow benchmark for generating ordered steps from a goal
GOAL (Wang et al., 2022)	NLP summarization	First English sports game summarization dataset from live commentary to news
GOAL (Qi et al., 2023)	Video-language	Knowledge-grounded soccer video captioning benchmark for real-time commentary
GOAL-style benchmarks (Abdelwahed et al., 16 Feb 2026)	Automated planning	Goal-recognition datasets built from planning traces, recently extended to multi-plan versions
GOALS (2207.14447)	Medical imaging	Different acronym: Glaucoma OCT Analysis and Layer Segmentation

This plurality matters because the term “GOAL Dataset” can otherwise be misread as referring to a single established benchmark. In practice, one must identify the relevant domain and paper before interpreting scale, annotations, and evaluation protocol.

2. GOAL as a multilingual script-learning dataset

In "Goal-Oriented Script Construction" (Lyu et al., 2021), GOAL is a multilingual benchmark for constructing full procedural scripts from goals. The dataset is derived from wikiHow, and the paper describes it as the first multilingual script learning dataset supporting 18 languages. The extracted schema is explicit: each article becomes a goal-oriented script with Goal, Section, Steps, and Category. The Goal is the article title with “How to” stripped; Steps are the ordered event sequence; and Sections are retained in extraction but flattened for modeling into a single ordered script.

The scale is substantial. Appendix counts sum to approximately 444,000 articles/scripts across 18 languages, including 112,111 English articles, 64,725 Spanish, 34,194 Portuguese, 31,541 German, and smaller collections down to 1,434 Turkish. Only a subset is marked as ordered scripts; for example, English has 54,852 ordered articles, Spanish 26,620, and Turkish 520. The split protocol holds out 10% of articles in each language as test, uses 90% for training and development, and then holds out 5% of the training portion for development (Lyu et al., 2021).

The task formulation is goal-conditioned. Given a goal $g$, the system constructs a complete script as an ordered list of steps $S$, with ground-truth reference $T$, and receives the expected number of steps $l$ as input. In the retrieval setting, the model must predict a subset of steps from a candidate pool $C$ and provide their ordering. The paper defines script-construction accuracy and Kendall-style ordering metrics, including

$\textrm{acc} = (\sum_i^l [s_i \in T])/l$

and

$$\tau = \frac{NC(S\cap T,T\cap S)-ND(S\cap T,T\cap S)}{\binom{l}{2}}.$$

The benchmark evaluates both generation and retrieval. The generation baseline fine-tunes mT5-Large; the retrieval pipeline performs Step Inference and Step Ordering with mBERT or XLM-RoBERTa. Reported average retrieval performance remains modest, with mBERT at Accuracy .286, Kendall’s $\tau=.198$ and XLM-R at Accuracy .317, Kendall’s $\tau=.075$. Human evaluation shows a trade-off: retrieval outputs contain more correct steps, whereas generation outputs are better ordered but hallucinate more. The dataset also transfers zero-shot to SMILE, OMICS, DeScript, and KAIROS Schema Learning Corpus, which the paper uses to argue that goal-conditioned procedural knowledge learned from wikiHow is not merely an artifact of one source format (Lyu et al., 2021).

3. GOAL as an English sports game summarization benchmark

In "GOAL: Towards Benchmarking Few-Shot Sports Game Summarization" (Wang et al., 2022), GOAL denotes the first English sports game summarization dataset. The task is to generate a sports news article from a document of real-time soccer commentary. The corpus is crawled from Goal.com and covers four competitions: UEFA Champions League, UEFA Europa League, Premier League, and Serie A, over the period 2016 to 2020.

The paper collected 2,263 football games total. All 2,263 have commentary documents, but only 103 have paired sports news articles. Those 103 commentary-news pairs constitute the supervised dataset, while the remaining 2,160 commentary-only documents are released as unlabeled data for semi-supervised research. The labeled split is 63 training, 20 validation, and 20 test examples. The average commentary length is 2724.9 words and the average news length is 476.3 words; the corresponding 95th percentiles are 3699 and 614 (Wang et al., 2022).

The task is formalized as mapping a timestamped commentary sequence

$C=\{(t_1,c_1), (t_2,c_2),...,(t_n,c_n)\}$

to a news article

$S$0

A central difficulty is that, unlike prior Chinese sports summarization datasets, GOAL commentary omits the explicit score field $S$1 found in tuples of the form $S$2. The model must therefore infer game state implicitly from commentary text. The paper also emphasizes a strong source-target style gap: commentary is short, event-driven, and colloquial, whereas news is compact and journalistically written.

The benchmark includes extractive and abstractive baselines. Longest, TextRank, and PacSum are used as extractive systems; PGN and LED are used as abstractive systems. The best reported model is LED, with test ROUGE-1 34.7, ROUGE-2 7.8, and ROUGE-L 24.3. The paper interprets these numbers as evidence that the task remains difficult even for long-document summarization architectures, and reports frequent repetition and missed important events in generated outputs (Wang et al., 2022).

4. GOAL as a knowledge-grounded soccer commentary benchmark

In "GOAL: A Challenging Knowledge-grounded Video Captioning Benchmark for Real-time Soccer Commentary Generation" (Qi et al., 2023), GOAL stands for knowledge-Grounded videO cAptioning benchmark for reaL-time soccer commentary generation. Here the target is not summarization from text, but generation of soccer commentary from video plus external knowledge. The paper introduces the task of Knowledge-grounded Video Captioning (KGVC), formalized as

$S$3

The function $S$4 selects relevant knowledge for video $S$5, and $S$6 generates the final commentary from video and selected knowledge.

The benchmark is built from SoccerNet-v2. The authors start with 80 full-game videos narrated in English, run Azure ASR toolkit on commentary speech, filter out videos with low resolution or too sparse narration, and retain 20 games as final annotation candidates. The final resource contains over 8.9k soccer video clips, 22k sentences, and 42k knowledge triples. The average clip length is 10.31 seconds, each clip has 2.46 labeled sentences on average, and each video is relevant to 193.1 knowledge triples on average (Qi et al., 2023).

Annotation is extensive. The team consists of 10 English native speakers who are also experienced soccer fans. The 20 preserved videos are distributed among these 10 annotators; each annotator completes 4 videos, so every video is annotated by two people. Their tasks are commentary proofreading, video-text alignment, and sentence-level knowledge annotation. Each sentence is labeled as scene description, background introduction, or comment, and linked to entities such as players, teams, and soccer terms. Knowledge is collected from WhoScored and Wikidata, with entity linking performed using BLINK and 2-hop related entity pages crawled from Wikidata (Qi et al., 2023).

The paper positions this GOAL benchmark as harder than standard video captioning because commentary must identify fine-grained entities, infer event semantics, and incorporate knowledge not directly visible in the clip. Three adapted baselines are reported: HMN, SwinBERT, and ALPRO, each with and without knowledge. The strongest result is ALPRO + Knowledge, with BLEU 15.6, METEOR 6.4, ROUGE-L 11.5, and CIDEr 6.6. The gap to standard captioning benchmarks is large: the paper notes that SwinBERT averages 95.5 CIDEr on several standard datasets but only 3.7 on GOAL, which the authors interpret as evidence that KGVC remains substantially unresolved (Qi et al., 2023).

5. GOAL-style datasets in automated planning and goal recognition

A different usage appears in "Removing Planner Bias in Goal Recognition Through Multi-Plan Dataset Generation" (Abdelwahed et al., 16 Feb 2026). This work does not introduce a branded dataset called GOAL, but it explicitly addresses the design of GOAL-style goal-recognition benchmarks in automated planning. Its central claim is that existing datasets are systematically biased because they are usually generated by heuristic-based forward-search planners and therefore encode a narrow planner-specific behavioral signature.

The paper formalizes a planning task as

$S$7

and a goal-recognition task as

$S$8

Its methodological contribution is to generate multiple plans for the same goal hypothesis using top-$S$9 planning rather than a single trace. Each plan yields a different observation sequence, and each observation sequence becomes a different version of the same underlying goal-recognition instance. The generation pipeline, denoted $T$0, takes a planning task $T$1, planner, goal hypothesis $T$2, required number of variants $T$3, observability percentage $T$4, and noise percentage $T$5, then returns clean and noisy task collections. The paper states that common dataset settings use

$T$6

and

$T$7

To evaluate grouped versions, the paper introduces Version Coverage Score (VCS). For a set of goal-recognition tasks sharing the same hypothesis,

$T$8

and a recognizer $T$9,

$l$0

computes the ratio of correctly solved tasks to the number of tasks. A recognizer is said to be resilient if

$l$1

for threshold $l$2 (Abdelwahed et al., 16 Feb 2026).

Empirically, the paper regenerates the Ramírez and Geffner benchmark with $l$3 using SymK and evaluates a landmark-based recognizer. The reported result is that accuracy and positive predictive value decline as the resilience threshold increases, especially at low observability. At 10% observability, mean accuracy falls from 95.0 at threshold 0.0 to 76.0 at threshold 1.0, while PPV falls from 81.0 to 13.0. The paper’s broader implication is that GOAL-style evaluation should measure robustness across multiple valid plans for the same goal, not only success on a single planner-preferred trace (Abdelwahed et al., 16 Feb 2026).

6. Disambiguation with GOALS and other goal-oriented resources

The ambiguity surrounding GOAL is compounded by nearby acronyms and by the generic use of the adjective goal-oriented. GOALS, for example, is a different term altogether: in medical imaging it denotes the Glaucoma OCT Analysis and Layer Segmentation challenge, which released 300 circumpapillary OCT images for retinal layer segmentation and glaucoma classification (2207.14447). Despite the visual similarity of the names, GOALS is unrelated to the GOAL datasets in script learning, sports summarization, soccer commentary, or planning.

A second source of confusion is that many datasets are goal-oriented without being named GOAL. Examples include doc2dial, a goal-oriented document-grounded dialogue dataset with 4,793 conversations over 487 documents (Feng et al., 2020), and the code-mixed DSTC2 derivatives introduced in "A Dataset for Building Code-Mixed Goal Oriented Conversation Systems" (Banerjee et al., 2018). These resources are relevant to the broader study of goals in language and planning, but they are not instances of a proper noun “GOAL Dataset.”

Taken together, these usages suggest that GOAL Dataset is best treated as a disambiguation term. In script learning, it denotes a large multilingual procedural benchmark; in NLP summarization, it denotes a small but influential English soccer-news resource; in video-language research, it denotes a knowledge-grounded commentary benchmark; and in automated planning, it denotes a family of goal-recognition benchmarks whose construction has recently shifted from single-plan to multi-plan evaluation.