A Novel Approach to Automated Live-Ticker Generation in Football:

Using Large Language Models and Audio Data

James Anurathan, Manfred R

ossle

and Marco Klaiber

Aalen University of Applied Sciences, Beethovenstr. 1, Aalen 73430, Germany

Keywords:

Large Language Model, Audio Data, Football, Named Entity Recognition, Natural Language Processing,

Live-Ticker.

Abstract:

Football (soccer) is one of the most popular sports in the world, with fans enjoying real-time coverage of

their favorite team’s from anywhere. Explicitly, the progress in the ﬁeld of Artiﬁcial Intelligence (AI) holds

great potential to further improve this experience and optimize the delivery of content. In this context, our

work investigates the integration of Large Language Models (LLMs) – in our case GPT-4 – with Advanced

Speech Recognition (ASR) systems to automate the creation of live football ticker commentary. For this

purpose, we present an approach for transcribing live audio commentary from real football matches, utilizing

a whisper model to prepare the transcribed text for correct input to the LLM. This approach is leveraged by

Named Entity Recognition (NER) and BERT-based models to provide clear, user-friendly, and multilingual

texts for live tickers. In addition, we evaluate our approach with an objective and metric-based method to

transparently assess the effectiveness of our approach. The study shows the potential of LLMs in automating

sports commentary, but also emphasizes the importance of reﬁning entity recognition and addressing content

accuracy issues. Future work should focus on improving transcription accuracy, reﬁning NER models, and

mitigating LLM hallucinations to develop more reliable and scalable automated live ticker systems.

1 INTRODUCTION

Football (soccer) is considered one of the most pop-

ular sports in the world, captivating a large number

of people and experiencing strong continuous growth

in recent years (Cotta, 2016, Anzer and Bauer, 2022).

This popularity developed football into a very lucra-

tive business, generating billions of dollars from var-

ious sources, with fans supporting their teams from

all over the world (Goes et al., 2019,

Cwiklinski

et al., 2021). This global need for the availability

of information from football matches has signiﬁcantly

changed the rapid development of digital journalism

in recent years with respect to the landscape of sports

reporting (Cheng et al., 2024). As one of the most el-

ementary components of sports reporting, live ticker

systems have established themselves as an indispens-

able element (Ojomo and Olomojobi, 2021).

These systems have become an integral part of

sports coverage, providing fans with real-time up-

dates and an immersive experience that bridges the

https://orcid.org/0000-0002-9038-9317

https://orcid.org/0009-0007-7070-3413

gap for those unable to access live broadcasts on tele-

vision or radio (Ojomo and Olomojobi, 2021). For

most football matches, the textual live-tickers focus

on highlighting the most important events such as

goals, shots, yellow or red cards, and substitutions

ochtefeld et al., 2015). In addition, impressions

of the game and other important information are con-

veyed as vividly as possible, to simulate the feeling

of being almost live in the stadium. However, man-

ual creation of live-tickers remains a time-consuming

and demanding task, requiring constant monitoring

and quick text generation by reporters (Huang et al.,

2020). In addition, the vivid live-tickers are usually

only offered for more popular games, with smaller

leagues sometimes not having a live-ticker or only be-

ing able to report simple events based on structured

data, such as event data.

This challenge emphasises the need for automa-

tion, which relies heavily on data being available

to accurately capture and report key events (Kunert,

2020). Although structured data captures the basic

details of a match, unstructured data, such as au-

dio commentary, provides comprehensive informa-

tion, including context and detailed insights (Behera

134

Anurathan, J., Rössle, M. and Klaiber, M.

A Novel Approach to Automated Live-Ticker Generation in Football: Using Large Language Models and Audio Data.

DOI: 10.5220/0013665700003988

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 13th International Conference on Sport Sciences Research and Technology Support (icSPORTS 2025), pages 134-141

ISBN: 978-989-758-771-9; ISSN: 2184-3201

and Saradhi, 2024, Tuyls et al., 2021). In combina-

tion with advances in Artiﬁcial Intelligence (AI), in

particular Large Language Models (LLMs), a promis-

ing solution is emerging to address the automation

of live ticker creation (Bonner et al., 2023). This is

partly because state-of-the-art LLMs have the ability

to integrate multimodal data, including audio, to ef-

fectively deliver results (Strand et al., 2024). This is

particularly apparent for Automatic Speech Recogni-

tion (ASR) systems (Lakomkin et al., 2024). As a re-

sult, the combination of the ASR system and LLM of-

fers the possibility to use the mentioned properties of

audio commentaries, for high-quality automated live-

tickers (Fathullah et al., 2024,Min and Wang, 2024).

Consequently, a research gap exists, as there are

currently no studies that effectively integrate LLMs

with audio commentary for the automatic generation

of live tickers. Our work aims to address this research

gap by developing a novel approach that demonstrates

how LLMs can be effectively used to generate and

optimise automated live tickers based on audio data

from football radio broadcasts. To do this, we propose

an approach that involves the process of transcribing,

processing and structuring audio data to make the data

understandable and usable for LLMs, with a focus on

accurately capturing and reporting key events. In ad-

dition, we are also exploring the potential of trans-

lating the transcribed texts into multiple languages,

which should improve the accessibility and global

reach of the live ticker system. Therefore, this work

offers the following contributions:

1. Development of an approach to automatically

recognise relevant football events from audio

broadcasts.

2. Implementation of a Large Language Model for

the conversion of identiﬁed events into clear, user-

friendly and multilingual texts for live tickers.

3. Objective metrics-based evaluation of the results

to transparently assess the effectiveness of our ap-

proach.

4. Based on our challenges, recommendations for

future research in the ﬁeld of football analysis and

speech processing are given.

2 RELATED WORK

The use of LLMs to generate live tickers from audio

sources is a novel area of research in football. Nev-

ertheless, some studies have pursued similar research

approaches, which we present below and distinguish

from our own approach.

Cook and Karakus (Cook and Karakus¸, 2024)

introduced the “LLM Commentator”, a system de-

signed to automate real-time football commentary us-

ing LLMs. Their approach leverages advanced speech

processing techniques and raw football data to gen-

erate accurate descriptions of match events. A sig-

niﬁcant contribution of their work is the exploration

of ﬁne-tuning strategies for LLMs, speciﬁcally tai-

lored to improve the models’ performance in captur-

ing and articulating live football commentary. Fur-

thermore, Sarkhoosh et al.(Sarkhoosh et al., 2024)

developed the “SoccerSum” framework, which em-

ploys LLMs such as GPT-4 to automate the sum-

marisation of football events. Their approach inte-

grates multimodal data, including audio, to generate

narrative content. However, in contrast to the focus

of our study, “SoccerSum” extends beyond the cre-

ation of live tickers by incorporating video analytics

and social media content creation. Another approach

was presented by Strand et al. (Strand et al., 2024)

in the area of football analytics. “SoccerRAG” is a

framework that combines Retrieval Augmented Gen-

eration (RAG) with LLMs to effectively respond to

natural language queries and ﬁnd relevant informa-

tion. This framework integrates multiple data modal-

ities, including video, audio, and recorded commen-

tary, to handle complex queries and improve user in-

teraction with sports archives.

In contrast to presented articles, our work focuses

on transcribing and analyzing live audio data from

football radio broadcasts to capture key events and

produce multilingual live-tickers. In addition, we use

a NER approach specially ﬁne-tuned for football, as

well as BERT to prepare the events in the best possi-

ble way, focusing exclusively on the audio data. We

also differentiate ourselves by integrating the possi-

bility of real-time translation into other languages to

increase global accessibility.

3 METHODOLOGY

For this work, a german four-minute audio ﬁle

was obtained from Sportschau.de

, a popular sports

show in Germany that focuses on commentary on

sporting events, especially football. Therefore, we

used a match from the German Bundesliga be-

tween VfL Wolfsburg and VfB Stuttgart (Matchday 24,

2024/03/02). The MP3 ﬁle contains the highlights of

the game and is used as a dataset for our approach.

The overall methodology is shown in Fig. 1.

Sportschau.de - Audio ﬁle

A Novel Approach to Automated Live-Ticker Generation in Football: Using Large Language Models and Audio Data

135

Figure 1: Overview of the individual steps of the proposed approach.

3.1 Transcription of Audio Data

The ﬁrst step was to provide the audio data in text

form; therefore, the data had to be transcribed. The

Whisper model (OpenAI, 2025) was used, which is

an advanced speech recognition system (ASR) from

OpenAI (Radford et al., 2023). Whisper is known

for its high speech recognition accuracy and has the

ability to process multiple languages, including Ger-

man, which is why Whisper is considered effective

as an ASR system for our approach (Radford et al.,

2023,Wills et al., 2023). Of the various versions avail-

able, the medium-sized Whisper model was selected

based on the trade-off between performance and re-

source efﬁciency. In addition, for optimised process-

ing efﬁciency, the audio ﬁle was divided into smaller

segments and each segment was then transcribed us-

ing Whisper.

3.2 Entity Extraction with BERT-Based

NER and Fuzzy Matching

To correctly recognize key events, players, and team

names from transcribed audio commentaries, Named

Entity Recognition (NER) was used. NER is a core

element of Natural Language Processing (NLP), used

to classify objects in text (Jehangir et al., 2023). This

involves assigning speciﬁc tags to each word or to-

ken, indicating its role. The integration of NER

into the pre-training phase is expected to improve

the overall performance of LLMs by providing them

with more accurate and structured data inputs (Devlin

et al., 2019). An effective model for NER tasks is the

Bidirectional Encoder Representations of Transform-

ers (BERT) (Devlin et al., 2019), which was therefore

selected for our work. The BERT-based multilingual

model is a pre-trained transformer model that captures

the context of words in a sentence taking into account

the preceding as well as following words, and is also

able to accurately recognise and process named ob-

jects in multiple languages (Chizhikova et al., 2023).

The BERT-based NER model was speciﬁcally ﬁne-

tuned for the football context in our study.

Since the pronunciation of the player names in the

audio ﬁle can differ, which can happen especially with

non-native names, our dataset was adapted to contain

several variants of player names to account for pos-

sible transcription differences. Each element in the

transcribed text was labeled with speciﬁc tags, such

as ‘B-PER’ for the beginning of a person’s name,

‘I-PER’ for the continuation of that name, and sim-

ilar tags for events. The tagged data was then split

into training and validation sets to ensure that the

NER model could accurately recognise and categorise

football-speciﬁc entities, which is essential for creat-

ing accurate and contextually relevant live ticker up-

dates. In addition to the NER model, fuzzy match-

ing techniques were employed to handle variations

and misspellings in player names and other entities.

Fuzzy matching enhances the model’s ability to cor-

rectly identify entities even when they are not per-

fectly transcribed, ensuring greater accuracy in the

recognition process (Bhasuran et al., 2016).

After training, the NER model was used to pro-

cess the transcriptions, automatically identifying and

labeling the key entities. These labeled entities were

then mapped to the correct teams and events, provid-

ing the LLM with the structured information needed

to generate accurate live-ticker updates. The trained

NER model was further applied to extract and classify

relevant entities from live audio comments.

3.3 Live-Ticker Generation and

Translation

The ﬁnal step was to create the live ticker commentary

using the GPT-4 model. This model was selected be-

cause of its advanced capabilities in natural language

processing, particularly its ability to understand con-

text, generate coherent text, and integrate transcrip-

tion data. In addition, a simple prompt was deﬁned

that places the model in the role of a sports commen-

tator to capture the context:

icSPORTS 2025 - 13th International Conference on Sport Sciences Research and Technology Support

136

“You are a sports commentator providing

live text commentary. Make sure you correctly

report key events, player names, scores, and goal

scorers, and assign them to the correct clubs.”

To ensure that the live commentary was accessible

to a global audience, the commentary was translated

from German to English using the Google Translator

API

. Google Translator was selected as it provides a

fast and reliable balance between translation quality

and speed (Baek et al., 2024). In addition, we wanted

to outsource the translation task from the LLM to save

resources.

3.4 Objective Evaluation Metrics

To ensure an objective evaluation, we integrate three

metrics: (1) BERTScore, (2) ROUGE-L, and (3)

BLEU. For each metric, we used LLM-generated

texts compared to the transcribed original from the

audio data, which represents the ground truth. The

LLM-generated live-ticker is not intended to repro-

duce the audio description identically, but to repro-

duce the most important information on the basis of

the audio description, whereby there should be a sim-

ilarity, which is why the comparison is reasonable.

The BERTScore (Zhang et al., 2020) measures the

quality of texts based on semantic similarity to the

original text. Furthermore, we integrate ROUGE-L,

based on the Longest Common Subsequence (LCS)

(Lin and Och, 2004), which considers structural simi-

larity at sentence level and automatically identiﬁes the

longest n-grams occurring in the sequence (Briman

and Yildiz, 2024). In addition, we use BLEU, which

analyses the accuracy of n-grams with a focus on the

precision of the model output (Tran et al., 2019).

4 RESULTS AND DISCUSSION

Considering the transcription of the audio commen-

tary using the Whisper model, the relevant informa-

tion was captured and accurately reproduced. This

includes, for example, information on the ﬂow and

dynamics of the game, as well as speciﬁc events such

as goals, penalties, and notable moves. Explicitly,

the quieter passages were captured almost perfectly.

In contrast, the more exciting and louder passages,

mostly related to goals, posed difﬁculties. In addi-

tion, the names of the players represent a major chal-

lenge for our model, with M

ale being transcribed in-

GoogleTranslator API

stead of Maehle or Girassi instead of Guirassy, for

example. The model struggled to accurately recog-

nize and transcribe player names from the audio com-

mentary, which is a critical aspect of generating pre-

cise live-tickers. It can also be assumed that this dif-

ﬁculty may arise with more complex club names, but

this was not the case in our work. This highlights

the need for more domain-speciﬁc language models

trained speciﬁcally on sports commentary data. An-

other option would be to provide the model with the

entire squad data of the respective teams. For this, a

comprehensive dataset of the teams must be collected

accordingly, which could be implemented using free-

accessible data from Transfermarkt (TM) (Transfer-

markt, 2025) or Fbref (FBref, 2025), for example. Al-

ternatively, using a larger and more advanced version

of Whisper, such as the “large” model, could poten-

tially improve transcription accuracy. Nevertheless, it

is conceivable that this challenge is not only due to

the model, but it could also be due to the pronuncia-

tion of the players’ names by the commentators. For

instance, players with names of non-German origin,

such as “Guirassy”, which has a French pronuncia-

tion, might be more challenging to recognize accu-

rately compared to names like “M

uller”, which have

a more straightforward German pronunciation. In ad-

dition, different commentators also have different lan-

guage styles and pronunciations, making exact recog-

nition even more difﬁcult.

In Table 1 excerpts from the LLM-generated live

ticker are shown. Our approach was able to report the

most relevant match events, for example, goals and

scores. Moreover, the disallowed goal was also recog-

nised and recorded in the live ticker. Despite the in-

tegration of the NER and fuzzy matching techniques,

there are still inconsistencies in the naming of player

names. For example, our model incorrectly named the

scorer of the 0:1, where J

uhrich (which corresponds

to the player F

uhrich) was named, although the real

scorer was Guirassy, which was correctly recorded

in the transcription. It should also be noted that our

approach generated artiﬁcial game events by insert-

ing game minutes that were not present in the original

transcribed text, which is referred to as hallucination

(Li et al., 2024). Accordingly, the LLM occasionally

generated non-existent details for the live commen-

tary, which poses a signiﬁcant challenge to ensuring

accurate and reliable reporting. To mitigate this, the

LLM could be ﬁne-tuned speciﬁcally for summariz-

ing football commentary, with a strong emphasis on

grounding its outputs in actual input data. Integrat-

ing a fact-checking mechanism or cross-referencing

(Jiang et al., 2024) system within the LLM could

also help reduce instances of hallucination. Another

A Novel Approach to Automated Live-Ticker Generation in Football: Using Large Language Models and Audio Data

137

Table 1: Excerpts of generated live ticker texts and the translation.

Min. German Live Ticker English Live Ticker

90’ +

Stuttgart sichert sich 3 wichtige Punkte. . .

Tore durch J

urich, Girassi und Wagnumann.

Stuttgart secures 3 important points. . . Goals

by J

urich, Girassi, and Wagnumann.

85’

TOR! Matcher erzielt den Anschlusstreffer

ur Wolfsburg!

GOAL! Matcher scores the equalizer for

Wolfsburg!

81’

TOR ABERKANNT! Wagnumanns Tref-

fer wird aufgrund einer Abseitsposition nicht

gegeben.

GOAL DISALLOWED! Wagnumann’s

goal is disallowed due to an offside position.

79’ TOR! Ein Flitzer von Wagnumann! GOAL! A streaker from Wagnumann!

72’ Girassi l

auft an. . . Tor! 2:1 VfB Stuttgart. Girassi runs up. . . Goal! 2-1 VfB Stuttgart.

70’

Elfmeter f

ur Stuttgart! M

ale bringt Mio zu

Fall und sieht Gelb. Girassi legt sich den Ball

zurecht.

Penalty for Stuttgart! M

ale brings Mio down

and sees yellow. Girassi sets the ball up.

65’

TOR! Wolfsburg! M

ale trifft mit einem wun-

derbaren Schuss aus 18 Metern in das linke

obere Eck. 1:1

GOAL! Wolfsburg! M

ale scores with a won-

derful shot from 18 meters into the top left

corner. 1:1

45’

Halbzeit! Stuttgart f

uhrt 1:0 in Wolfsburg.

Das Tor von J

urich, ein Distanzschuss, der

wirklich atemberaubend war.

Halftime! Stuttgart leads 1:0 in Wolfsburg.

urich’s goal, a long-range shot that was truly

breathtaking.

15’

TOR! Stuttgart! J

urich zieht von links nach

rechts und versenkt den Ball aus der zweiten

Reihe seelenruhig ins Netz. Super Start f

den VfB. Wolfsburg 0, Stuttgart 1.

GOAL! Stuttgart! J

urich moves from left to

right and calmly sinks the ball into the net

from the second row. Great start for VfB.

Wolfsburg 0, Stuttgart 1.

1’

Anpﬁff! Das Spiel ist er

offnet, Wolfsburg hat

Anstoß.

Kick-off! The game is underway, Wolfsburg

has the kick-off.

way to prevent hallucination would be to adjust the

prompt, which in our case was kept relatively straight-

forward, but the addition that the model “should not

hallucinate” could improve the results (Tonmoy et al.,

2024). However, our approach correctly updated the

scores and assigned the goals to the correct teams.

Accordingly, the matching at team level is effective,

as the model correctly assigns the players to the re-

spective teams. In addition, the translation from Ger-

man to English has been successfully implemented so

that live ticker entries have been translated correctly

in terms of content. The incorrect player names have

been adopted identically in the translation.

For an objective evaluation of our approach, we

used the three presented metrics to compare 10 seg-

ments with the reference and LLM-generated text,

which is shown in Table 2. It should be noted that

the LLM changed the order between the segments

6 – 9, which explains why the values in this area

became signiﬁcantly worse, especially for ROUGE-

L and BLEU. To counteract this in the future, it is

conceivable to divide the commentary into chunks,

for example, whenever a highlight was discovered,

whereby all chunks could then be processed in se-

quential order. The BERTscore, on the other hand, is

relatively constant across all segments and has an av-

erage value of ≈ 0.70, which shows that our approach

is semantically close to the references. On the other

hand, an average ROUGE-L of ≈ 0.20 and BLEU of

≈ 5.3 show that there is relatively little exact word

choice or n-gram overlap. This pattern is typical when

the outputs are heavily paraphrased, content has been

rearranged, or differs in length from the ground truth.

Overall, the evaluation show that the system is suc-

cessful in capturing general context and key events,

but that there are signiﬁcant inaccuracies and incon-

sistencies, particularly in exact word matching and in-

sertion of missing parts. Since the live ticker is not in-

tended to directly reproduce the audio ﬁle, the results

can be considered a benchmark, which applies in par-

ticular to the BERTscore. Improvements in the BERT-

based NER model, LLM tuning, and post-processing

steps could enhance the quality of live ticker genera-

tion, ensuring that outputs are more reliable and closer

to actual commentary.

However, we were able to successfully present

an initial approach for a multilingual automated live

ticker system with Whisper AI, NER, BERT, and an

icSPORTS 2025 - 13th International Conference on Sport Sciences Research and Technology Support

138

Table 2: Evaluation of the generated segments.

Segment BERTScore ROUGE-L BLEU

1 0.8150 0.2667 8.91

2 0.6638 0.2014 15.32

3 0.7414 0.2833 22.45

4 0.6814 0.2410 18.75

5 0.6675 0.2117 16.89

6 0.7325 0.1714 4.33

7 0.6541 0.1111 0.9954

8 0.6695 0.0833 1.1174

9 0.7171 0.2143 1.7118

10 0.6898 0.2090 6.8816

Average 0.70321 0.20159 5.2703

LLM. With our research, we are laying the founda-

tions for the automatic generation of live tickers using

advanced AI technologies. The live ticker texts gen-

erated emphasize the potential of our approach, as the

audio ﬁles were captured correctly and a meaningful

live ticker could be presented.

5 LIMITATIONS AND FUTURE

WORK

One challenge was training the NER model to identify

key players and game events from the transcribed text.

However, the training might not have been sufﬁcient,

leading to inconsistent recognition of entities and po-

tential inaccuracies in the live-ticker. To address this,

the NER model could beneﬁt from more extensive

training on a larger and more diverse dataset. In

addition, experimenting with different NER models,

e.g. those based on transformer architectures such as

RoBERTa (Mehta and Varma, 2023), could lead to

better results. Furthermore, it is conceivable to eval-

uate the effectiveness of individual components, such

as the NER model, which can be carried out in the

future with ablation studies.

For our approach, we used a game with the high-

lights audio ﬁle, which was sufﬁcient for our study,

but it is conceivable that more data could improve the

results. The accuracy of both the transcription and

entity recognition processes could be limited by the

small size of the training data, limiting the model’s

ability to generalize effectively. Collecting a larger

dataset, including a wide range of football matches

with varying commentary styles and audio quality,

could signiﬁcantly improve model performance. The

use of data augmentation techniques to synthetically

increase the size of the dataset could also be beneﬁcial

(Moreno-Barea et al., 2020). It is also conceivable

that not only highlights are used, but entire broad-

casts of matches, which would signiﬁcantly increase

the amount of data but would prioritize the recogni-

tion of highlights for the live ticker.

We used three different metrics for the evalua-

tion, namely BERTScore, ROUGE-L, and BLEU, al-

though other metrics or evaluation approaches could

be considered in the future. ROUGE-L and BLEU

in particular are susceptible to deterioration in results

when formulations differ despite semantic correct-

ness, which could potentially obscure the validity of

the live ticker passages. Consequently, COMET (Tas-

nim et al., 2019) or BARTScore (Yuan et al., 2021),

which have been used in other domains, could be con-

sidered in the future. It should also be considered

for future evaluation that, especially for a 90-minute

game, a relevant challenge is to ﬁnd the appropriate

highlights that are worth mentioning in a live ticker.

Furthermore, our approach focused on using a

previous game to demonstrate feasibility. How-

ever, live commentaries are generated on match days,

which then have to be converted into a live ticker

in real-time. The performance of our approach was

sufﬁcient for our use case, but it remains to be seen

how the system will handle real-time processing, es-

pecially when processing large audio ﬁles over 90

minutes. In the future, further optimization of the

system for real-time processing could be promoted,

which includes streamlining audio segmentation and

transcription pipelines, possibly using lighter mod-

els that offer a better compromise between accuracy

A Novel Approach to Automated Live-Ticker Generation in Football: Using Large Language Models and Audio Data

139

and speed. Exploring parallel processing techniques

(Brakel et al., 2024) and GPU acceleration (Huang

et al., 2024) could also improve performance.

6 CONCLUSION

This study demonstrates the potential of integrating

LLMs with advanced speech recognition systems to

automate live-ticker generation in football. The pro-

posed approach, which includes the implementation

of the Whisper model for audio transcription, fol-

lowed by NER and fuzzy matching techniques, ef-

fectively processes live commentary to generate accu-

rate, real-time textual updates. Our approach show-

cases both the possibilities and challenges involved

in creating a robust system for real-time sports com-

mentary automation. The results showed that while

the system successfully captured the overall context

and key events of a football match, there were notable

challenges with player name recognition and the gen-

eration of non-existent match details. These inaccura-

cies underscore the need for further reﬁnement in both

the transcription process and the entity recognition

models. Speciﬁcally, enhancing the training datasets

with more diverse and extensive football commen-

tary could improve the system’s ability to general-

ize across different pronunciation variations and com-

mentary styles.

The limitations identiﬁed in this study, such as

the LLM’s tendency to hallucinate and the challenges

in real-time processing, point to several avenues for

future research. Fine-tuning LLMs to better han-

dle football-speciﬁc commentary, improving NER

model accuracy, and optimizing real-time process-

ing capabilities are essential steps forward. Further-

more, exploring alternative transformer-based models

like RoBERTa (Liu et al., 2019) or developing more

domain-speciﬁc LLMs (Jeong, 2024) could further

enhance system performance. Future research should

also consider the scalability of such systems, particu-

larly in multilingual environments and across various

sports. Implementing cloud-based distributed com-

puting solutions could address these scalability con-

cerns, allowing simultaneous processing of multiple

matches in different languages.

REFERENCES

Anzer, G. and Bauer, P. (2022). Expected passes: Determin-

ing the difﬁculty of a pass in football (soccer) using

spatio-temporal data. Data Mining and Knowledge

Discovery, 36(1):295–317.

Baek, S., Lee, S., and Seok, J. (2024). Strategic Insights in

Korean-English Translation: Cost, Latency, and Qual-

ity Assessed through Large Language Model. 2024

Fifteenth International Conference on Ubiquitous and

Future Networks (ICUFN), pages 551–553.

Behera, S. R. and Saradhi, V. V. (2024). Visualization of

Unstructured Sports Data – An Example of Cricket

Short Text Commentary.

Bhasuran, B., Murugesan, G., Abdulkadhar, S., and Natara-

jan, J. (2016). Stacked ensemble combined with fuzzy

matching for biomedical named entity recognition of

diseases. Journal of Biomedical Informatics, 64:1–9.

Bonner, E., Lege, R., and Frazier, E. (2023). Large

Language Model-Based Artiﬁcial Intelligence in the

Language Classroom: Practical Ideas for Teaching.

Teaching English with Technology, 23(1):23–41.

Brakel, F., Odyurt, U., and Varbanescu, A.-L. (2024).

Model Parallelism on Distributed Infrastructure: A

Literature Review from Theory to LLM Case-Studies.

Briman, M. K. H. and Yildiz, B. (2024). Beyond ROUGE:

A Comprehensive Evaluation Metric for Abstrac-

tive Summarization Leveraging Similarity, Entail-

ment, and Acceptability. International Journal on Ar-

tiﬁcial Intelligence Tools, 33(05):1–23.

Cheng, L., Deng, D., Xie, X., Qiu, R., Xu, M., and Wu, Y.

(2024). SNIL: Generating Sports News From Insights

With Large Language Models. IEEE Transactions on

Visualization and Computer Graphics, pages 1–14.

Chizhikova, A., Konovalov, V., Burtsev, M., Kryzhanovsky,

B., Dunin-Barkowski, W., Redko, V., and Tiumentsev,

Y. (2023). Multilingual Case-Insensitive Named En-

tity Recognition. In Advances in Neural Computation,

Machine Learning, and Cognitive Research VI, pages

448–454.

Cook, A. and Karakus¸, O. (2024). LLM-Commentator:

Novel ﬁne-tuning strategies of large language models

for automatic commentary generation using football

event data. Knowledge-Based Systems, 300:1–22.

Cotta, L. (2016). Using FIFA Soccer video game data for

soccer analytics. Workshop on large scale Sports An-

alytics, pages 1–4.

Cwiklinski, B., Giełczyk, A., and Chora

s, M. (2021). Who

Will Score? A Machine Learning Approach to Sup-

porting Football Team Building and Transfers. En-

tropy, 23(1):1–12.

Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K.

(2019). BERT: Pre-training of Deep Bidirectional

Transformers for Language Understanding.

Fathullah, Y., Wu, C., Lakomkin, E., Jia, J., Shangguan, Y.,

Li, K., Guo, J., Xiong, W., Mahadeokar, J., Kalinli,

O., Fuegen, C., and Seltzer, M. (2024). Prompt-

ing Large Language Models with Speech Recognition

Abilities. In ICASSP 2024 - 2024 IEEE International

Conference on Acoustics, Speech and Signal Process-

ing (ICASSP), pages 13351–13355.

FBref (2025). Fbref.com. https://fbref.com/en/. Accessed:

08.01.2025.

Goes, F. R., Kempe, M., Meerhoff, L. A., and Lemmink,

K. A. (2019). Not Every Pass Can Be an Assist: A

icSPORTS 2025 - 13th International Conference on Sport Sciences Research and Technology Support

140

Data-Driven Model to Measure Pass Effectiveness in

Professional Soccer Matches. Big Data, 7(1):57–70.

Huang, K. H., Chen, L., and Chang, K. W. (2020). Generat-

ing Sports News from Live Commentary: A Chinese

Dataset for Sports Game Summarization. In Proc. 1st

Conf. of the Asia-Paciﬁc Chapter of the Ass. for Com-

putational Linguistics and the 10th Int. Joint Conf. on

Natural Language Processing, pages 609–615.

Huang, Y., Wan, L. J., Ye, H., Jha, M., Wang, J., Li, Y.,

Zhang, X., and Chen, D. (2024). Invited: New Solu-

tions on LLM Acceleration, Optimization, and Appli-

cation. In Proceedings of the 61st ACM/IEEE Design

Automation Conference, pages 1–4.

Jehangir, B., Radhakrishnan, S., and Agarwal, R. (2023).

A survey on Named Entity Recognition — datasets,

tools, and methodologies. Natural Language Process-

ing Journal, 3:1–12.

Jeong, C. (2024). Domain-specialized LLM: Finan-

cial ﬁne-tuning and utilization method using Mistral

7B. Journal of Intelligence and Information Systems,

30(1):93–120.

Jiang, L., Jiang, K., Chu, X., Gulati, S., and Garg, P.

(2024). Hallucination Detection in LLM-enriched

Product Listings. In Malmasi, S., Fetahu, B., Uefﬁng,

N., Rokhlenko, O., Agichtein, E., and Guy, I., editors,

Proc. Seventh Workshop on e-Commerce and NLP @

LREC-COLING, pages 29–39.

Kunert, J. (2020). Automation in Sports Reporting: Strate-

gies of Data Providers, Software Providers, and Media

Outlets. Media and Communication, 8(3):1–11.

Lakomkin, E., Wu, C., Fathullah, Y., Kalinli, O., Seltzer,

M. L., and Fuegen, C. (2024). End-to-End Speech

Recognition Contextualization with Large Language

Models. In ICASSP 2024 - 2024 IEEE International

Conference on Acoustics, Speech and Signal Process-

ing (ICASSP), pages 12406–12410.

Li, H., Chi, H., Liu, M., and Yang, W. (2024). Look Within,

Why LLMs Hallucinate: A Causal Perspective.

Lin, C.-Y. and Och, F. J. (2004). Automatic evaluation

of machine translation quality using longest common

subsequence and skip-bigram statistics. In Proceed-

ings of the 42nd Annual Meeting on Association for

Computational Linguistics - ACL ’04, pages 1–8.

Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D.,

Levy, O., Lewis, M., Zettlemoyer, L., and Stoyanov,

V. (2019). RoBERTa: A Robustly Optimized BERT

Pretraining Approach.

ochtefeld, M., J

ackel, C., and Kr

uger, A. (2015). TwitSoc-

cer: Knowledge-Based Crowd-Sourcing of Live Soc-

cer Events. In Proceedings of the 14th International

Conference on MUM, pages 1–4.

Mehta, R. and Varma, V. (2023). LLM-RM at SemEval-

2023 Task 2: Multilingual Complex NER using XLM-

RoBERTa.

Min, Z. and Wang, J. (2024). ”Exploring the Integration

of Large Language Models into Automatic Speech

Recognition Systems: An Empirical Study”. In Luo,

B., Cheng, L., Wu, Z.-G., Li, H., and Li, C., editors,

Neural Information Processing, pages 69–84.

Moreno-Barea, F. J., Jerez, J. M., and Franco, L. (2020).

Improving classiﬁcation accuracy using data augmen-

tation on small data sets. Expert Systems with Appli-

cations, 161:1–14.

Ojomo, O. W. and Olomojobi, O. T. (2021). Viewing

the Game Textually: Online Consumption of Live

Text Commentary as Alternate Spectatorship Among

Nigerian Football Fans. Communication & Sport,

9(3):496–521.

OpenAI (2025). Introducing Whisper.

https://openai.com/index/whisper/. Accessed:

15.01.2025.

Radford, A., Kim, J. W., Xu, T., Brockman, G., Mcleavey,

C., and Sutskever, I. (2023). Robust speech recog-

nition via large-scale weak supervision. In Proceed-

ings of the 40th International Conference on Machine

Learning, volume 202 of Proceedings of Machine

Learning Research, pages 28492–28518.

Sarkhoosh, M. H., Gautam, S., Midoglu, C., Sabet, S. S.,

Torjusen, T., and Halvorsen, P. (2024). The Soccer-

Sum Dataset for Automated Detection, Segmentation,

and Tracking of Objects on the Soccer Pitch. In Pro-

ceedings of the 15th ACM Multimedia Systems Con-

ference, MMSys ’24, page 353–359.

Strand, A. T., Gautam, S., Midoglu, C., and Halvorsen, P.

(2024). SoccerRAG: Multimodal Soccer Information

Retrieval via Natural Queries. arXiv.

Tasnim, M., Collarana, D., Graux, D., Galkin, M., and

Vidal, M.-E. (2019). COMET: A Contextualized

Molecule-Based Matching Technique. In Lecture

Notes in Computer Science, pages 175–185.

Tonmoy, S. M. T. I., Zaman, S. M. M., Jain, V., Rani, A.,

Rawte, V., Chadha, A., and Das, A. (2024). A Com-

prehensive Survey of Hallucination Mitigation Tech-

niques in Large Language Models.

Tran, N., Tran, H., Nguyen, S., Nguyen, H., and Nguyen, T.

(2019). Does BLEU Score Work for Code Migration?

In 2019 IEEE/ACM 27th International Conference on

Program Comprehension (ICPC), pages 165–176.

Transfermarkt (2025). transfermarkt.com.

https://www.transfermarkt.de/. Accessed:

12.04.2025.

Tuyls, K., Omidshaﬁei, S., Muller, P., Wang, Z., Connor, J.,

Hennes, D., Graham, I., and et al. (2021). Game Plan:

What AI can do for Football, and What Football can

do for AI. JAIR, 71:41–88.

Wills, S., Bai, Y., Tejedor-Garc

ıa, C., Cucchiarini, C., and

Strik, H. (2023). Automatic Speech Recognition of

Non-Native Child Speech for Language Learning Ap-

plications (Short Paper). In arXiv:2306.16710, pages

1–8.

Yuan, W., Neubig, G., and Liu, P. (2021). BARTScore:

Evaluating generated text as text generation. In Ran-

zato, M., Beygelzimer, A., Dauphin, Y., Liang, P., and

Vaughan, J. W., editors, Advances in Neural Informa-

tion Processing Systems, volume 34, pages 27263–

27277.

Zhang, T., Kishore, V., Wu, F., Weinberger, K. Q., and

Artzi, Y. (2020). BERTScore: Evaluating Text Gener-

ation with BERT.

A Novel Approach to Automated Live-Ticker Generation in Football: Using Large Language Models and Audio Data

141