Indonesian Sports Text Classification Modeling Based on Few-Shot NER

2.1 Sports technology innovation

Technology integration in sports has significantly transformed, with digital platforms and virtual environments becoming central to various sporting activities. For instance, platforms like Zwift offer mixed-reality environments where users’ real-world performances are translated into virtual settings, enhancing motivation and physical performance through social interactions and immersive experiences [28]. Adopting technology-driven innovations like flywheel resistance training has gained traction in elite sports, providing neuromuscular and strength benefits through controlled, variable loading exercises [29]. These technologies reshape how sports are practiced and understood, pushing the boundaries of traditional sports concepts into the digital realm [30].

Despite these advancements, there are ongoing debates about how technology redefines the nature of sport. Scholars have highlighted that the terms used to describe these new forms, such as eSports, cybersport, and virtual sport, lack consistency, which creates challenges for comprehensive research [30]. However, the convergence of digital and traditional sports provides unique opportunities, especially for sports technology entrepreneurs, who utilize digital platforms to overcome gender biases and foster innovation [31]. As society continues to embrace digital transformation, sport will likely evolve, balancing tradition with cutting-edge technologies [28, 30].

The application of NLP in sports is still in its early stages, though it has shown significant potential across various domains. One example is using machine learning techniques for text classification in sports news [11]. The progress, however, is often constrained by factors such as limited data resources [12] and the complexity of language in sports commentary [10]. Despite the rise of LLMs, which have demonstrated success in sports commentary, the challenge of creating coherent and accurate automated commentary remains, as it requires handling domain-specific knowledge and diverse events.

Additionally, NLP techniques have been explored in the mental training of athletes, demonstrating their utility in improving psychosomatic skills like managing stress and enhancing self-confidence [9]. A study highlighted the benefits of NER-based interventions for managing anxiety and improving performance. This demonstrates that while NLP's application in sports is still developing, its impact is already evident in commentary generation and athlete mental training [9, 10]. These early advancements suggest that NLP could become a more integral part of sports.

2.2 Advances and constraints in domain-specific Named Entity Recognition

NER is an NLP technique that identifies and categorizes specific entities in a text, such as people, organizations, locations, and medical terms. NER plays a significant role in various domains, where it extracts critical information from electronic records and scientific documents [32, 33]. The application of NER has been widely adopted in machine learning and AI models, helping to structure unstructured data [13, 34]. Additionally, NER has been used to create knowledge graphs and supporting intelligent systems [35, 36]. One of NER's primary advantages is its ability to enhance the accuracy and efficiency of information extraction, which leads to improved decision-making [37]. Furthermore, domain adaptation in NER models allows them to handle specific linguistic patterns or other specialized domains, reducing errors in entity recognition and improving the performance of downstream tasks [33, 34].

NER systems can be implemented using different approaches, such as rule-based, machine learning-based, or hybrid techniques, each offering varying degrees of flexibility, accuracy, and resource dependence [23, 38]. Machine learning-based methods are popular due to their adaptability across domains and languages, though they require extensive annotated data. Combining rule-based and machine-learning techniques, hybrid models can enhance system performance, especially in complex domains [39]. Recent developments have introduced deep learning architectures, such as Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), and Convolutional Neural Networks (CNN), which improve NER performance by better capturing semantic and syntactic features [40]. These advanced methods significantly enhance NER tasks, particularly in under-resourced languages where traditional methods fall short [41].

NER has become an essential task in various specialized fields, but domain-specific challenges, data limitations, and the complexity of linguistic structures often hinder its effectiveness. For instance, the lack of publicly available annotated datasets hampers advancements in identifying key entities related to crop diseases and pests in agricultural domains [42]. Similarly, materials science faces challenges with overlapping entities in scientific literature, prompting Machine Reading Comprehension (MRC) techniques to enhance accuracy by converting entity extraction into a question-answering task [43]. In historical and low-resource domains, NER's complexity increases due to the lack of training data and the degradation of historical documents. In ancient Chinese texts, the concise nature of the language complicates the recognition process, while integrating external knowledge, such as dictionaries, can enhance the representation of entities [44]. In contrast, LLMs offer a potential solution for historical document NER but face challenges when handling noisy inputs, as their performance still lags behind dedicated NER systems [45].

Challenges in NER often arise from the limitations of data and the complexity of nested structures. One challenge is the recognition of nested entities, where entities overlap, making traditional models inadequate. Sequence labeling methods face performance issues due to the complexity of hierarchical structures, which results in poor accuracy and error propagation when external entities are misclassified after internal ones are recognized [44, 45]. While effective in theory, Span-based approaches require the enumeration of all possible spans, leading to high computational complexity and frequent misidentification of non-entities [46-48]. These issues highlight the difficulty in dealing with entities that have complex positional dependencies, further complicating the task of NER.

Another significant area for improvement stems from language-specific challenges. For example, Chinese NER is complicated by the need for word delimiters, making it difficult to identify entity boundaries, particularly in nested entities. Multi-feature fusion, which incorporates glyphs and pinyin, is essential but increases the complexity of the models [49]. Similarly, agglutinative languages such as Turkish present challenges in informal text settings, where abbreviations, irregular grammar, and orthographic errors are standard, leading to a substantial decrease in the performance of NER models [50]. These limitations underscore the need for more adaptable models capable of handling diverse linguistic features and overcoming the challenges posed by complex and noisy data.

2.3 Few-Shot learning

Few-Shot learning has become a crucial research subject aimed at tackling the issue of training models with few labelled data across several fields. Initial research primarily focused on meta-learning, which instructs a model on the learning process, using episodic training to improve generalization across tasks with few instances [49, 50]. Nonetheless, obstacles such as overfitting to the support set and meta-dataset persist, as seen in studies on ensemble meta-learning and adaptive transformers [51-53].

A notable strategy in Few-Shot learning is model-centric techniques, whereby researchers enhance designs or use ensemble learning to increase learning efficacy from limited datasets [52, 53]. Examples include the integration of multi-input and multi-output setups inside meta-learning to improve generalization performance by circumventing local minima during training [52]. The fine-tuning of pre-trained networks has proven effective, especially in image-based damage detection, where models such as ProtoNet exhibit encouraging outcomes after fine-tuning with a limited support set [52, 54].

Conversely, data-centric solutions emphasize enhancing data quality and addressing domain transitions, often hindering model performance in Few-Shot contexts [49, 51]. Cross-domain adaptation techniques seek to alleviate the effects of distribution discrepancies between the source and target domains. Moreover, self-supervised learning is often used to exploit unlabeled data successfully, as shown in recent initiatives for cross-domain threat intelligence extraction and segmentation tasks [23, 51].

Unsupervised and lifetime learning are becoming significant as academics investigate techniques to enhance model flexibility without extensive annotated datasets [55-57]. Unsupervised meta-learning networks use task-specific reconstruction techniques for domain adaptability, enhancing feature extraction across tasks with few labelled data. Lifelong ensemble learning in robotics utilizes many representations to adapt to new object categories dynamically, optimizing the balance between identification accuracy and processing efficiency.

LLMs and prompt-based fine-tuning have gained significance in text-based Few-Shot applications, including code review automation and named entity identification in cybersecurity. Few-Shot strategies, including personal-specific instructions and embedding-based retrieval methods, provide effective solutions for LLMs, especially when extensive fine-tuning is unfeasible [58]. These developments underscore the progression of Few-Shot learning towards more flexible, efficient, and scalable solutions relevant across many domains and workloads.

4.1 Data collecting

The news dataset was compiled through automated retrieval using Google News RSS feeds, which aggregate articles from a wide range of publishers. This approach enables efficient collection of diverse headlines while minimizing manual intervention. A sample of the extracted data showing the structured format containing the news title, publication date, and associated URL is presented in Figure 2. The final dataset comprises 1543 Indonesian sports-related news headlines, which were subsequently categorized according to the sport taxonomy listed in Table 1.

(a) Indonesia language

(b) English language

Figure 2. Data collecting result

Table 1. Indonesian sport entities

4.2 Pre-processing pipeline for Few-Shot NER

To prepare the text for Few-Shot NER, a comprehensive pre-processing pipeline was developed for the Indonesian language. This process includes lowercasing, punctuation and numeric removal, stop word elimination, tokenization, and stemming. The outcome of the Indonesian pipeline is shown in Figure 3(a), demonstrating how raw input is transformed into a cleaned and standardized token sequence. For comparison, Figure 3(b) presents the same pipeline applied to English text, highlighting cross-language consistency and generalizability of the pre-processing steps. These transformations ensure uniform input structure, thereby enhancing downstream model robustness and reducing noise-related variability.

(a) Indonesia language

(b) English language

Figure 3. Data collecting result

4.3 Data annotation

Annotation follows the widely adopted BIO tagging scheme, enabling explicit identification of entity boundaries within each sentence. An example of the annotated output is illustrated in Figure 4. In this example, tokens such as golf are labeled as B-Sport_Individual-aiming, marking the beginning of an entity span, while other non-entity tokens receive the O label. This annotation framework provides structural clarity and facilitates accurate supervised training of NER models, particularly in specialized domains such as sports news.

Figure 4. Data annotation result

4.4 Dataset diagnostic analysis

A deeper evaluation of the dataset’s structure was conducted to assess its suitability for Few-Shot NER and to identify potential sources of bias. Sentence length distributions for both training and test sets are presented in Figure 5. The boxplot reveals that the corpus is dominated by short headlines, with average lengths of 10.43 tokens (train) and 10.61 tokens (test). This characteristic reflects the typical structure of Indonesian news headlines, which tend to be concise and highly compressed.

Figure 5. Sentence length distribution in train and test sets

Figure 6. ECDF of sentence lengths

Further analysis using the Empirical Cumulative Distribution Function (ECDF) shown in Figure 6 demonstrates near-identical cumulative patterns between the two splits. This consistency indicates an effective and well-balanced data partition without structural drift.

Token-level label comparison, visualized in Figure 7, highlights a pronounced class imbalance. The O label accounts for approximately 89% of all tokens across both splits, while each sport-related entity label constitutes less than 1.3%. Although the distribution between splits is stable (points cluster closely along the diagonal), the extreme imbalance underscores the importance of compensatory strategies such as class weighting or data augmentation during model training.

A span-level analysis of entity categories is provided in Figure 8. All eight sport categories exhibit comparable proportions, each contributing between 9% and 14% of total spans, indicating balanced representation at the entity level. However, lexical diversity within categories remains limited some categories contain only one to three unique surface forms. This low diversity raises concerns regarding potential lexical bias, where the model may memorize specific entity words rather than learn generalized patterns.

Figure 7. Token-level label distribution comparison

(a) Entity-type distribution train

(b) Entity-type distribution train

Figure 8. Entity-type distribution across train and test sets

Finally, Figure 9 compares entity-type proportions across the two splits. Points aligning closely with the diagonal confirm that the train-test partitioning maintained consistent entity distributions, ensuring fair evaluation conditions without category-specific skew.

Collectively, these analyses provide an in-depth understanding of dataset characteristics and highlight both strength such as well-balanced entity-type distribution and limitations, particularly in token-level imbalance and low surface-form diversity. These insights directly inform the design and expected performance of Few-Shot NER models applied to Indonesian sports news.

Figure 9. Span-level entity-type proportion comparison

4.5 Few-Shot NER fine tuning model

The construction of a few-Shot model comprises two essential phases: model initialization and fine-tuning. Model initialization involves loading a pre-trained pipeline that provides a foundation for customization, guaranteeing the model has essential linguistic characteristics such as tokenization, part-of-speech tagging, and named entity identification. The initialization stage is formally described in Algorithm 1, which summarizes the procedural pipeline for constructing the base Indonesian SpaCy model prior to downstream Few-Shot adaptation. Subsequently, fine-tuning entails adjusting or training the model using a limited, task-specific dataset by updating pertinent components, such as including new labels for text categorization or custom entities. When used with appropriate batching, the nlp.update () method in SpaCy modifies the model parameters according to the supplied Few-Shot data. Subsequently, following training, the optimized model is assessed for efficacy and preserved for deployment using nlp.to_disk () to guarantee its suitability for the designated application.

The function train_ner_model initializes a blank SpaCy model for Indonesian, adds an NER pipeline, and loads training data from a DocBin file. It registers entity labels from the training data and begins model training with an optimizer. The training loop runs for a specified number of iterations, where the data is divided into batches using minibatch and compounding. Each batch is used to update the model with Example objects that map entity spans, and the losses are tracked and printed for each iteration. Finally, the trained model is returned for further use.

The pseudocode Algorithm 2 describes the process of fine-tuning an existing NER model using Few-Shot data. It begins by loading the Few-Shot data from a DocBin file and extracting the documents with their annotated entities. The NER component of the NLP model is accessed, and any new entity labels found in the data are added to the model. The training resumes using an optimizer, running for a specified number of iterations (n_iter). In each iteration, the data is divided into batches of increasing size, and the model is updated with examples created from the documents and their annotations, while tracking the training losses. After all iterations have been completed, the refined model is returned for further application.

Figure 10 shows how loss decreased a NER model across five iterations. The first graph reveals a high loss of roughly 450, which decreases markedly by the second repetition, signifying substantial improvement. Nonetheless, after the fourth cycle, the loss exhibits a tiny rise, indicating a little variability. The second graph, by comparison, shows a smaller range of loss values, commencing at around 20 and progressively declining. Although the pattern seems more consistent in the first rounds, a significant decrease in loss transpires in the final iteration. These graphs illustrate the enhancement of the model's performance, with considerable diversity in optimization across the training cycles, particularly after fine-tuning.

Loss iteration in NER refers to the process of calculating and minimizing the error (loss) during the training of a machine learning model. The model predicts which words in a sentence correspond to specific named entities in each iteration. The model's predictions are compared to the actual labeled entities, and a loss function measures how far off the predictions are. This loss is then used to update the model's parameters via an optimization algorithm (like gradient descent), improving its accuracy over time. Each iteration refines the model, gradually reducing the loss until it reaches an optimal performance.

(a) SpaCy

(b) Few-Shot fine tuning

Figure 10. Comparison of loss iteration between SpaCy and Few-Shot fine tuning

4.6 Evaluation model

The assessment of model performance must be conducted using the commonly used metrics based on accuracy. This study includes evaluation metrics such as precision (P), recall (R), and F1-score. The formulas for calculating precision (P), recall (R), and F1-score are as follows:

Precision (P) $=\frac{\text {True Positive}}{(\text {True Positive}+ \text {False Positive})}$     (1)

Recall (R) $=\frac{\text {True Positive}}{\text {True Positive}+ \text {False Negatives}}$     (2)

$F 1=\frac{2 \times \text {Precision} \times  \text {Recall}}{\text {Precision}+ \text {Recall}}$     (3)

Table 2. Performance metrics of the Indonesian Few-Shot NER model across evaluation scenarios

The results presented in Table 2 demonstrate a clear performance advantage of the proposed Few-Shot Fine-Tuning model over the baseline methods. While the SpaCy model achieves a strong F1-score of 0.983 and perfect recall, its slightly lower precision indicates a modest tendency to over-predict entities. The Bi-LSTM-CRF model performs considerably weaker, particularly in recall (0.795), suggesting that it fails to identify many true entities likely due to its limited ability to capture contextual nuances in Indonesian sports headlines. The Transformer-IndoBERT model provides a substantial improvement, reaching an F1-score of 0.961 with a well-balanced precision and recall profile, reflecting the benefits of contextualized language modeling. However, the proposed Few-Shot Fine-Tuning approach surpasses all baselines with an F1-score of 0.995, supported by near-perfect precision (0.991) and perfect recall. This exceptional performance indicates that adapting a pre-trained model with a domain-specific Few-Shot dataset enables highly effective entity representation, allowing the model to generalize well within the sports domain while minimizing false predictions.

4.7 Error analysis and improvement methods

The entity-level confusion matrices across the four evaluated NER architectures SpaCy baseline, BiLSTM-CRF, Transformer-IndoBERT, and the proposed Few-Shot fine-tuned model provide a detailed view of misclassification behaviour among the eight Indonesian sport categories (Figure 11). While the SpaCy baseline, BiLSTM-CRF, and IndoBERT models display strong diagonal dominance with only negligible off-diagonal mass, indicating stable cross-category discrimination, the Few-Shot fine-tuned model exhibits marked deviations from this ideal pattern. In particular, a substantial proportion of Sport_Combat instances are confused with Sport_Target and, to a lesser extent, with Sport_Individual-Aesthetic, despite the generally high span-level F1 achieved in aggregate evaluation. This asymmetry is consistent with recent findings showing that NER models can appear robust at the corpus level while still exhibiting systematic, label-specific errors that only become visible through fine-grained confusion analysis and per-label diagnostics [59-62]. Given that Indonesian sports headlines are short and lexically compressed, with limited surface-form diversity within each category, the Few-Shot model is particularly prone to over-generalising from a small number of support examples and collapsing semantically adjacent classes into a single dominant prototype.