Dataset Card for GEM/mlb_data_to_text

Link to Main Data Card

You can find the main data card on the GEM Website.

Dataset Summary

The MLB dataset is an English sport-related data-to-text dataset in the baseball domain. The input is a large table with results of a game and the output is a description of the game.

You can load the dataset via:

import datasets
data = datasets.load_dataset('GEM/mlb_data_to_text')

The data loader can be found here.

website

Github

paper

ACL Anthology

authors

Ratish Puduppully, Li Dong, Mirella Lapata

Dataset Overview

Where to find the Data and its Documentation

Webpage

Github

Download

Github

Paper

ACL Anthology

BibTex

@inproceedings{puduppully-etal-2019-data,
    title = "Data-to-text Generation with Entity Modeling",
    author = "Puduppully, Ratish  and
      Dong, Li  and
      Lapata, Mirella",
    booktitle = "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
    month = jul,
    year = "2019",
    address = "Florence, Italy",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/P19-1195",
    doi = "10.18653/v1/P19-1195",
    pages = "2023--2035",
}

Contact Name

Ratish Puduppully

Contact Email

ratishpuduppully@gmail.com

Has a Leaderboard?

no

Languages and Intended Use

Multilingual?

no

Covered Languages

English

License

other: Other license

Intended Use

The dataset can be used to study data-to-text generation. The dataset is in sports domain. It pairs statistics of Major League Baseball (MLB) game with its summary. The summary is in the form of a document containing an average of 540 tokens. Thus it is useful to study long document generation.

Add. License Info

Restricted to non-commercial research purposes.

Primary Task

Data-to-Text

Communicative Goal

Produce a summary of MLB game from its statistics.

Credit

Curation Organization Type(s)

academic

Curation Organization(s)

University of Edinburgh

Dataset Creators

Ratish Puduppully, Li Dong, Mirella Lapata

Dataset Structure

Data Fields

        features = datasets.Features(
            {
                "home_name": datasets.Value("string"),
                "box_score": [
                    {
                        "p_l": datasets.Value("string"),
                        "last_name": datasets.Value("string"),
                        "p_h": datasets.Value("string"),
                        "sac": datasets.Value("string"),
                        "p_bb": datasets.Value("string"),
                        "pos": datasets.Value("string"),
                        "ao": datasets.Value("string"),
                        "p_bf": datasets.Value("string"),
                        "cs": datasets.Value("string"),
                        "hbp": datasets.Value("string"),
                        "ab": datasets.Value("string"),
                        "full_name": datasets.Value("string"),
                        "p_w": datasets.Value("string"),
                        "go": datasets.Value("string"),
                        "fldg": datasets.Value("string"),
                        "p_bs": datasets.Value("string"),
                        "avg": datasets.Value("string"),
                        "p_r": datasets.Value("string"),
                        "p_s": datasets.Value("string"),
                        "lob": datasets.Value("string"),
                        "first_name": datasets.Value("string"),
                        "p_sv": datasets.Value("string"),
                        "p_so": datasets.Value("string"),
                        "p_save": datasets.Value("string"),
                        "p_hr": datasets.Value("string"),
                        "po": datasets.Value("string"),
                        "p_ip1": datasets.Value("string"),
                        "p_ip2": datasets.Value("string"),
                        "bb": datasets.Value("string"),
                        "ops": datasets.Value("string"),
                        "p_hld": datasets.Value("string"),
                        "bo": datasets.Value("string"),
                        "p_loss": datasets.Value("string"),
                        "e": datasets.Value("string"),
                        "p_game_score": datasets.Value("string"),
                        "p_win": datasets.Value("string"),
                        "a": datasets.Value("string"),
                        "p_era": datasets.Value("string"),
                        "d": datasets.Value("string"),
                        "p_out": datasets.Value("string"),
                        "h": datasets.Value("string"),
                        "p_er": datasets.Value("string"),
                        "p_np": datasets.Value("string"),
                        "hr": datasets.Value("string"),
                        "r": datasets.Value("string"),
                        "so": datasets.Value("string"),
                        "t": datasets.Value("string"),
                        "rbi": datasets.Value("string"),
                        "team": datasets.Value("string"),
                        "sb": datasets.Value("string"),
                        "slg": datasets.Value("string"),
                        "sf": datasets.Value("string"),
                        "obp": datasets.Value("string"),
                    }
                ],
                "home_city": datasets.Value("string"),
                "vis_name": datasets.Value("string"),
                "play_by_play": [{
                    "top": [{
                        "runs": datasets.Value("string"),
                        "scorers": [
                            datasets.Value("string")
                        ],
                        "pitcher": datasets.Value("string"),
                        "o": datasets.Value("string"),
                        "b": datasets.Value("string"),
                        "s": datasets.Value("string"),
                        "batter": datasets.Value("string"),
                        "b1": [
                            datasets.Value("string")
                        ],
                        "b2": [
                            datasets.Value("string")
                        ],
                        "b3": [
                            datasets.Value("string")
                        ],
                        "event": datasets.Value("string"),
                        "event2": datasets.Value("string"),
                        "home_team_runs": datasets.Value("string"),
                        "away_team_runs": datasets.Value("string"),
                        "rbi": datasets.Value("string"),
                        "error_runs": datasets.Value("string"),
                        "fielder_error": datasets.Value("string")
                    }
                    ],
                    "bottom": [{
                        "runs": datasets.Value("string"),
                        "scorers": [
                            datasets.Value("string")
                        ],
                        "pitcher": datasets.Value("string"),
                        "o": datasets.Value("string"),
                        "b": datasets.Value("string"),
                        "s": datasets.Value("string"),
                        "batter": datasets.Value("string"),
                        "b1": [
                            datasets.Value("string")
                        ],
                        "b2": [
                            datasets.Value("string")
                        ],
                        "b3": [
                            datasets.Value("string")
                        ],
                        "event": datasets.Value("string"),
                        "event2": datasets.Value("string"),
                        "home_team_runs": datasets.Value("string"),
                        "away_team_runs": datasets.Value("string"),
                        "rbi": datasets.Value("string"),
                        "error_runs": datasets.Value("string"),
                        "fielder_error": datasets.Value("string")
                    }
                    ],
                    "inning": datasets.Value("string")
                }
                ],
                "vis_line": {
                    "innings": [{
                     "inn": datasets.Value("string"),
                     "runs": datasets.Value("string")
                    }
                    ],
                    "result": datasets.Value("string"),
                    "team_runs": datasets.Value("string"),
                    "team_hits": datasets.Value("string"),
                    "team_errors": datasets.Value("string"),
                    "team_name": datasets.Value("string"),
                    "team_city": datasets.Value("string")
                },
                "home_line": {
                    "innings": [{
                        "inn": datasets.Value("string"),
                        "runs": datasets.Value("string")
                    }
                    ],
                    "result": datasets.Value("string"),
                    "team_runs": datasets.Value("string"),
                    "team_hits": datasets.Value("string"),
                    "team_errors": datasets.Value("string"),
                    "team_name": datasets.Value("string"),
                    "team_city": datasets.Value("string")
                },
                "vis_city": datasets.Value("string"),
                "day": datasets.Value("string"),
                "summary": [
                    datasets.Value("string"),
                ],
                "gem_id": datasets.Value("string")
            }

Reason for Structure

The high level structure contains the following attributes: home_name, vis_name, home_city, vis_city, summary, summary_eval, day, gem_id, box_score, play_by_play, home_line, vis_line. The attributes home_name, vis_name, home_city, vis_city and day are string values. The attribute "summary" contains the summary in the form of a list of tokens. The attribute "summary_eval" contains the summary in the form of a string of tokens. The difference from "summary" field is that "summary_eval" doesn't contain "NEWPARAGRAPH" delimiters to separate the paragraphs. "summary_eval" field should be used to evaluate model outputs. "summary" field may be used during the training process. box_score contains the box score statistics of the players in the game. It is in the form of a list (of average size 90), with each element describing the statistics of a player. The box score statistics contain 53 attributes. The description of the attributes is given below. The descriptions of most of the attributes is obtained from mlb.com.

• r : Runs scored by a player in the game.
• rbi Runs Batted In (RBI): action of a batter results in a run scored by other players in the team.
• pos Position of the player.
• avg Batting Average. It is an indicator of the hits in the players' career.
• bb A walk occurs when a pitcher throws four pitches out of the strike zone, none of which are swung at by the hitter.
• hr Batter hits the ball in the air over the outfield fence.
• p_r Runs given by a pitcher in the game.
• p_bb Walks allowed by pitcher in a game.
• p_h Hits allowed by pitcher in a game.
• p_hr Home runs allowed by pitcher in a game.
• p_er Earned Run (ER): An earned run is any run that scores against a pitcher.
• p_era Earned Run Average (ERA): Earned run average represents the number of earned runs a pitcher allows per nine innings.
• p_np Number of Pitches: A pitcher's total number of pitches is determined by all the pitches he throws in game.
• p_ip1 Innings Pitched (IP1): Innings pitched measures the number of innings a pitcher remains in a game. Because there are three outs in an inning, each out recorded represents one-third of an inning pitched.
• p_ip2 Innings Pitched (IP2): Innings pitched measures the number of innings a pitcher remains in a game. Because there are three outs in an inning, each out recorded represents one-third of an inning pitched.
• p_w A pitcher receives a win when he is the pitcher of record when his team takes the lead for good.
• p_l A pitcher receives a loss when a run that is charged to him proves to be the go-ahead run in the game, giving the opposing team a lead it never gives up.
• p_so A strikeout occurs when a pitcher throws any combination of three swinging or looking strikes to a hitter.
• p_save Save: A save is awarded to the relief pitcher who finishes a game for the winning team. A pitcher cannot receive a save and a win in the same game.
• p_sv Saves: The count of saves recorded by a pitcher in his career.
• sac A sacrifice fly occurs when a batter hits a fly-ball out to the outfield or foul territory that allows a runner to score.
• p_bf Batters faced is simply a count of the number of total plate appearances against a certain pitcher or team. In a perfect game -- with 27 outs -- a pitcher will record 27 batters faced.
• cs A caught stealing occurs when a runner attempts to steal but is tagged out before reaching second base, third base or home plate.
• hbp A hit-by-pitch occurs when a batter is struck by a pitched ball without swinging at it. He is awarded first base as a result.
• ab An official at-bat comes when a batter reaches base via a fielder's choice, hit or an error (not including catcher's interference) or when a batter is put out on a non-sacrifice.
• p_bs A blown save occurs when a relief pitcher enters a game in a save situation, but allows the tying run to score.
• p_s The count of strikes thrown by a pitcher
• lob Left on base can be viewed as both an individual statistic or as a team statistic. In an individual batter's case, it refers to how many men remain on base after that batter makes an out at the plate, as the batter has failed to do his job to score those runners -- or at least put himself in a position to score. In a team's case or in an individual pitcher's case, it refers to the number of men who remain on base at the end of an inning.
• po A fielder is credited with a putout when he is the fielder who physically records the act of completing an out -- whether it be by stepping on the base for a forceout, tagging a runner, catching a batted ball, or catching a third strike
• ops OPS adds on-base percentage and slugging percentage to get one number that unites the two. It's meant to combine how well a hitter can reach base, with how well he can hit for average and for power.
• p_hld A hold occurs when a relief pitcher enters the game in a save situation and maintains his team's lead for the next relief pitcher, while recording at least one out.
• p_loss True/False- Indicates losing pitcher
• e A fielder is given an error if, in the judgment of the official scorer, he fails to convert an out on a play that an average fielder should have made.
• p_win True/False- Indicates winning pitcher
• a An assist is awarded to a fielder who touches the ball before a putout is recorded by another fielder.
• h A hit occurs when a batter strikes the baseball into fair territory and reaches base without doing so via an error or a fielder's choice.
• so A strikeout of a batter
• team Team of the player
• sb A stolen base occurs when a baserunner advances by taking a base to which he isn't entitled.
• slg Slugging percentage represents the total number of bases a player records per at-bat. Unlike on-base percentage, slugging percentage deals only with hits and does not include walks and hit-by-pitches in its equation.
• sf A sacrifice fly occurs when a batter hits a fly-ball out to the outfield or foul territory that allows a runner to score.
• obp OBP refers to how frequently a batter reaches base per plate appearance. Times on base include hits, walks and hit-by-pitches, but do not include errors, times reached on a fielder's choice or a dropped third strike.

The description of attributes in play-by-play is below:

• batter Batter in the play.
• pitcher Pitcher in play.
• b1 Player/s at first base position.
• b2 Player/s at second base position.
• b3 Player/s at third base position.
• scorers Player/s scored in the play.
• fielder_error Player committed field error.
• event Event of the play such as single, double, home run etc.
• event2 Second event of the play such as wild pitch, error etc.
• inning Inning of the play.
• top/ bottom If home team is batting it is bottom and if away team is batting it is top.
• o Count of outs
• b Count of balls
• s Count of strikes
• r Count of runs
• rbi Count of runs batted in (rbi)
• error_runs Runs due to error
• home_team_runs Score of home team
• vis_team_runs Score of visiting team

home_line and vis_line contain string value pairs for team_name, team_city, team_runs, team_hits, team_error, result, and a list of runs scored in each inning.

Data Splits

There are three splits in the dataset: train, validation and test

Splitting Criteria

The splits are random.

Dataset in GEM

Rationale for Inclusion in GEM

Why is the Dataset in GEM?

This dataset can verify if models are capable of long document generation. The challenges in long document generation conditioned on input tables include ensuring coherent output, staying faithful to the input, ensuring fluent output and avoiding repetition of text. Such aspects can be verified on models trained on this dataset

Similar Datasets

yes

Unique Language Coverage

no

Difference from other GEM datasets

Compared to the existing RotoWire (Wiseman et al. 2017) dataset, MLB summaries are longer (approximately by 50%) and the input records are richer and more structured (with the addition of play-by-play). Moreover, the MLB dataset is five times larger in terms of data size (i.e., pairs of tables and game summaries).

Ability that the Dataset measures

Long document generation, coherent ordering of information, faithfulness to the input statistics, fluency in generation and avoiding repetition of text.

GEM-Specific Curation

Modificatied for GEM?

yes

GEM Modifications

data points removed

Modification Details

Some examples have been removed from training dataset which satisfied the below criteria:

1. The examples in training dataset which overlapped with validation/test.
2. Some examples which described washed out games.

Additional Splits?

no

Getting Started with the Task

Pointers to Resources

The research paper is a good resource

Previous Results

Previous Results

Measured Model Abilities

Automatic evaluation measure can evaluate the factuality, content selection, content ordering and the fluency of the model output. The factuality, content selection and content ordering is measured using an Information Extraction based evaluation approach introduced by Wiseman et al (2017). The fluency is measured using BLEU.

Metrics

Other: Other Metrics

Other Metrics

Wiseman et al. (2017) define three metrics induced from the outputs of an Information Extraction model which is run on the model/human-written game summaries . Let ŷ be the gold summary and y the model output. • Relation Generation (RG) measures the precision and count of relations extracted from y that also appear in records r. • Content Selection (CS) measures the precision and recall of relations extracted from y that are also extracted from ŷ. • Content Ordering (CO) measures the complement of the normalized Damerau-Levenshtein distance (Brill and Moore, 2000) between the sequences of relations extracted from y and ŷ

Proposed Evaluation

We have reused the automatic metrics based on Information Extraction evaluation introduced by Wiseman et al (2017). For human evaluation, we conducted studies to evaluate the factuality, coherence, grammaticality and conciseness.

Previous results available?

yes

Relevant Previous Results

The most relevant previous results for dataset are in the TACL 2021 paper on Data-to-text Generation with Macro Planning

Dataset Curation

Original Curation

Original Curation Rationale

This dataset was curated to complement an existing data-to-text generation dataset (RotoWire by Wiseman et al. 2017) which focuses on long document generation. Compared to RotoWire , MLB summaries are longer (approximately by 50%) and the input records are richer and more structured (with the addition of play-by-play). Moreover, the MLB dataset is five times larger in terms of data size (i.e., pairs of tables and game summaries)

Communicative Goal

The goal is to study automatic generation of long documents in a data-to-text setting. The generated summaries should exhibit coherent ordering of content, be faithful to the input statistics, be fluent and avoid repetition of text.

Sourced from Different Sources

no

Language Data

How was Language Data Obtained?

Found

Where was it found?

Single website

Language Producers

The game summaries are produced by professional writers.

Topics Covered

The language focuses on the sports domain.

Data Validation

not validated

Data Preprocessing

Game summaries were tokenized using NLTK (Bird et al., 2009) and hyphenated words were separated. Sentences containing quotes were removed as they included opinions and non-factual statements unrelated to the input tables. Sometimes MLB summaries contain a "Game notes" section with incidental information which was also removed.

Was Data Filtered?

not filtered

Structured Annotations

Additional Annotations?

none

Annotation Service?

no

Consent

Any Consent Policy?

no

Justification for Using the Data

The copyright remains with the original data creators and the usage permission is restricted to non-commercial uses.

Private Identifying Information (PII)

Contains PII?

yes/very likely

Categories of PII

sensitive information, generic PII

Any PII Identification?

no identification

Maintenance

Any Maintenance Plan?

no

Broader Social Context

Previous Work on the Social Impact of the Dataset

Usage of Models based on the Data

no

Impact on Under-Served Communities

Addresses needs of underserved Communities?

no

Discussion of Biases

Any Documented Social Biases?

unsure

Considerations for Using the Data

PII Risks and Liability

Licenses

Copyright Restrictions on the Dataset

research use only

Copyright Restrictions on the Language Data

research use only

Known Technical Limitations