Accepted at INLG 2025

Input Matters: Evaluating Input Structure’s Impact on LLM Summaries of Sports Play-by-Play

Barkavi Sundararajan, Somayajulu (Yaji) Sripada, Ehud Reiter, University of Aberdeen

Input structure strongly influences factual accuracy in LLM-generated NBA game summaries. Here, factual accuracy refers to faithfulness to the given input data. This page provides a brief overview of the input-format comparison, the error-type reference table used in evaluation, and annotated examples.

Error Categories (Annotation Protocol)

Shorthand used in annotations: U=Number, N=Name, WO=Word Objective, WS=Word Subjective, C=Context, X=Not Checkable, O=Other.

Category Description Example
NUMBERU Incorrect numeric statistics (scores, rebounds, points). Errors highlighted: “The score was 12U8U.”
Correction: The correct score was 25–22.
NAMEN Wrong names of players, teams, or locations. Errors highlighted: “Assist by LeBron JamesN.”
Correction: The assist was by Jaylen Brown.
WORD_OBJECTIVEWO Factually wrong but checkable descriptions (shot type, foul type, etc.). Errors highlighted: Kevin Durant's “free throwsWO brought the Rockets within range.”
Correction: The player actually made two layups.
WORD_SUBJECTIVEWS Unverifiable opinions or subjective phrasing. Errors highlighted: “The bench provided a sparkWS.”
Explanation: They scored 6 points in the quarter (whether that is a “spark” is subjective).
CONTEXTC Player/team context wrong (e.g., implied wrong team/season fit for the event). Errors highlighted: “Christian WoodC hit a buzzer-beater for the Rockets.”
Correction: Christian Wood was not on the Rockets in this context; the correct player was Cam Whitmore.
NOT_CHECKABLEX Requires outside knowledge beyond the play-by-play (career highs, league rankings). Errors highlighted: “Brown had his career-highX tonight.”
Explanation: Cannot be verified from play-by-play logs; needs external career data.
OTHERO Any other mistake not covered above. Miscellaneous mistake not fitting other categories.

Input Structures (Partial game excerpts from the NBA play-by-play logs)

Time LA Lakers LA Pts Score ATL Pts Atlanta
2:56.0A. Davis makes free throw 1 of 21.0123-123
2:56.0A. Davis makes free throw 2 of 21.0124-123
2:45.0124-123D. Hunter misses 2-pt layup from 10 ft (block by L. James)
2:44.0124-123Offensive rebound by Team
2:33.0124-1252.0T. Young makes 2-pt layup from 2 ft
2:13.0L. James makes 3-pt jump shot from 25 ft3.0127-125
2:02.0127-125T. Young misses 3-pt jump shot from 25 ft
1:58.0127-125Offensive rebound by C. Capela
1:58.0127-1272.0T. Young makes 2-pt jump shot from 2 ft (assist by C. Capela)
1:44.0G. Vincent makes 3-pt jump shot from 26 ft3.0130-127
1:24.0130-1292.0D. Daniels makes 2-pt jump shot from 7 ft (assist by T. Young)
1:07.0G. Vincent misses 2-pt jump shot from 8 ft130-129
1:03.0130-129Defensive rebound by D. Daniels

Raw unstructured log view with per-event points by side (LA Pts / ATL Pts) and running score.

Time Period Team Primary Player Secondary Player Play Description Event Type Action Type Outcome Distance (ft) Current Points Scored LAL Cum. Score ATL Cum. Score
2:56.0 OT1 Los Angeles Lakers Anthony Davis N/A Anthony Davis makes free throw 1 of 2 Shot Free Throw Made N/A LAL: 1 LAL: 123 ATL: 123
2:56.0 OT1 Los Angeles Lakers Anthony Davis N/A Anthony Davis makes free throw 2 of 2 Shot Free Throw Made N/A LAL: 1 LAL: 124 ATL: 123
2:45.0 OT1 Atlanta Hawks De'Andre Hunter LeBron James De'Andre Hunter misses 2-pt layup from 10 ft (block by LeBron James) Shot Layup Missed 10 No points scored LAL: 124 ATL: 123
2:44.0 OT1 Atlanta Hawks N/A N/A Offensive rebound by Team Rebound Offensive N/A N/A No points scored LAL: 124 ATL: 123
2:33.0 OT1 Atlanta Hawks Trae Young N/A Trae Young makes 2-pt layup from 2 ft Shot Layup Made 2 ATL: 2 LAL: 124 ATL: 125
2:13.0 OT1 Los Angeles Lakers LeBron James N/A LeBron James makes 3-pt jump shot from 25 ft Shot Jump Shot Made 25 LAL: 3 LAL: 127 ATL: 125

Single 13-column table. Not-applicable attributes appear as “N/A”; in the released dataset these are stored as None.

{
      "time": {
        "period": "OT1",
        "clock": "2:13.0"
      },
      "team": "Los Angeles Lakers",
      "play_details": {
        "description": "LeBron James makes 3-pt jump shot from 25 ft",
        "event": {
          "type": "Shot",
          "action": "Jump Shot",
          "outcome": "Made",
          "distance": "25 ft",
          "points": 3
        },
        "players": {
          "primary_player": "LeBron James"
        }
      },
      "score": {
        "points_scored": "LAL: 3",
        "cumulative": {
          "LAL": 127,
          "ATL": 125
        }
      }
    }

Hierarchical JSON groups time (period/clock), play details (event and players), and scoring (points in play and cumulative totals).

Summary Examples with Annotations

Unstructured input: Llama3.1-70B (partial summary)

The Lakers fought back, however, with James and Davis leading the charge. James hit a pair of three-pointersWO, and Davis added a dunk and a layupWO to tie the game at 130U-129U. Throughout the game, James and Davis were the Lakers’ top performers, combining for 63U points. Young and BogdanovicN led the Hawks, combining for 61U points.

Row input: Qwen2.5-72B (partial summary)

The game saw several lead changes, with Anthony Davis making a pair of free throwsWO to tie the game at 123. LeBron James then hit a three-pointer to give the Lakers a 127–125 lead, but Trae Young answered with a layupWO to tie it again. The Lakers regained the lead with a three-pointer from Gabe Vincent, but the Hawks stayed close with a jump shot from Dyson Daniels. LeBron James finished with a doubleWO-double, contributing significantly on both ends of the court.

JSON input: Qwen2.5-72B (partial summary)

The Hawks took an early lead with a layup from Onyeka Okongwu, but the Lakers quickly tied itWO with a three-pointer from Max Christie. Throughout the game, LeBron James and Anthony Davis were the standout players for the Lakers, contributing significantly with scoring, rebounds, and assists.

Dataset

Please cite the paper if these resources are used.

BibTeX