My Master's Thesis in Mathematical Engineering
- Find dataset to RAG
- Use ElasticSearch to RAG new events that are not in the dataset
- Find golden-standard dataset with accurate and rich comment
For now I used FantasyPremierLeague, I only considered last 3 seasons and generated textual description for each seasons and for each games week for each player.
- https://www.football-data.co.uk/ • CSV con risultati, date, squadre, goal, xG, etc. (Serie A inclusa). • Perfetto per basi statistiche.
- https://github.com/statsbomb/open-data • Estremamente dettagliato (eventi in campo, passaggi, tiri). • Include alcuni campionati (non Serie A), ma puoi usarlo per prototipi.
- https://www.transfermarkt.com/ • Profili giocatori, statistiche stagionali, cronologia gol, assist, etc. • Puoi costruirti un dataset RAG-friendly.
- https://www.kaggle.com/datasets/hugomathien/soccer • Database SQLite con partite, squadre, player data (2008–2016).
- https://www.whoscored.com/ • Molto dettagliato ma da usare con cautela per TOS.
Limitations: Just a brief description of the event, not a real comment
Pro: Very rich comment, with insights, statistics ecc Cons: Difficult allineate events and comment, but I could use gpt4 to do that for me
Contacted them, waiting for a response:
- https://www.api-football.com/
- https://sportradar.com/
- https://www.statsperform.com/
- https://www.football-data.org/
- Youtube comment scraping(rich but dirty) and dataset from papers(poor but clean)
- retrieval ready dataset preprocessed and prepared from FantasyPremierLeague
- Indexing(faiss)+Embedding(SentenceTrasformer)
- created file .faiss and metadata.json
- Implement query retrieval functionality with FAISS and Sentence Transformers
Below are simple, factual templates (no emotion) for common football events. Fill in the placeholders ({player}, {team}, {minute}, etc.) as needed.
- “{minute}’ Goal by {player} for {team}.”
- “Goal at {minute}’: {player} scores for {team}.”
“{minute}’ Goal by {player} for {team}, assisted by {assist_player}.” “Goal at {minute}’: {player} scores for {team}, assisted by {assist_player}.”
- “{minute}’ Own goal by {player} (for {opponent}).”
- “Own goal at {minute}’ credited to {player}.”
- “{minute}’ Penalty awarded to {team}.”
- “Penalty at {minute}’ for {team}.”
- “Penalty converted by {player} at {minute}’.”
- “{minute}’ Penalty scored by {player}.”
- “{minute}’ Penalty by {player} missed.”
- “{minute}’ Penalty saved by {keeper}.”
- “{minute}’ Shot on target by {player}.”
- “{player} tests the keeper at {minute}’.”
- “{minute}’ Shot wide by {player}.”
- “{player} misses target at {minute}’.”
- “{minute}’ Shot by {player} blocked.”
- “Blocked attempt at {minute}’ from {player}.”
- “{minute}’ Corner to {team}.”
- “Corner awarded at {minute}’ for {team}.”
- “{minute}’ Free kick to {team}.”
- “Free kick conceded by {player} at {minute}’.”
- “{minute}’ Yellow card shown to {player}.”
- “Booking for {player} at {minute}’.”
- “{minute}’ Red card shown to {player}.”
- “{player} sent off at {minute}’.”
- “{minute}’ Substitution for {team}: {player_off} off, {player_on} on.”
- “{team} change at {minute}’: {player_on} replaces {player_off}.”
- “{minute}’ Offside called against {player}.”
- “{player} flagged offside at {minute}’.”
- “{minute}’ Foul by {player}.”
- “Foul at {minute}’ on {player}.”
- “{minute}’ Foul won by {player}.”
- “{player} wins a free kick at {minute}’.”
- “{minute}’ Play stopped for treatment to {player}.”
- “{player} down injured at {minute}’.”
- “{minute}’ VAR check underway.”
- “VAR overturns on-field decision at {minute}’.”
- “Kick‑off.”
- “Half‑time.”
- “Full‑time.”
+--------------------------------------------+ | Input: sentences with pre-set templates | +----------------------+---------------------+ | v +----------------------+---------------------+ | Extraction of statistics and data | +----------------------+---------------------+ | v +--------------+----------------+ | | v v +-------+---------+ +---------+---------+ | Extracted | | Template sentence | | statistics | | | +-------+---------+ +---------+---------+ \ / \ / \ / v v +------------------------------------+ | Generation via LLM | | (input: extracted stats + template)| +------------+-----------------------+ | +-------------+--------------+ | | v v +-------+---------+ +-------+----------+ | Fine-tuning LLM | | Prompt Engineering| | with real sports| +-------------------+ | commentary data | +-----------------+
Examples:
[1) Event Input] “Goal by de Bruyne with a left-foot strike at 12’” │ ▼ [2) Predefined queries based on the template sentence] • extract: event_type, player, team, minute • call → get_event_stats({event_type, player, team, season}) │ ▼ [3) Execute Query] get_event_stats(…) → from CSV/DB: • goals_this_season = 15 • left_foot_goals = 7 │ ▼ [4) LLM Generates Commentary] Input to LLM: • Event text • Stats dict System Prompt: “You are a sports commentator. Generate a lively, natural-sounding sentence using only the provided data.” Output: “De Bruyne’s left-foot strike from outside the box is his 15th goal of the season, 7 of which have come from his left foot—what a campaign for him!”
https://github.com/felipeall/transfermarkt-api
From video to commentary directly Other sports in the same way Voice of past telechronist Eventi precedenti nel prompt+tiri in porta sostituzioni, infortuni ecc Enfsasi sulla base del minuto, risultato, posizione in classifica, eventi precedenti