Cryptocurrency continues to dominate online discourse, with platforms such as Twitter, Reddit, and YouTube serving as major hubs for public opinion, market speculation, and user-driven Q&A. The CryptoQA track at FIRE 2025 aims to develop systems that can automatically analyze and categorize cryptocurrency-related social media posts. This year, the track consisted of two tasks: (a) classifying posts into a three-level hierarchical label space covering categories such as Noise, Objective, Subjective (and its subtypes), and (b) detecting whether an answer is relevant to a given question in question-answer pairs. In total, five teams participated in the track, submitting a wide range of approaches which were evaluated primarily using the macro F1-score across all datasets and task levels. The participating systems demonstrated strong performance in handling challenges such as sentiment ambiguity, noisy and unstructured text, and multi-platform variability, ofering valuable insights into modeling cryptocurrency discourse on social media.
Over the past decade, the emergence of new cryptocurrencies has had a profound impact on digital
ifnancial ecosystems. This shift has sparked ongoing conversations across various social media platforms.
These platforms, such as Twitter, Reddit, and YouTube, are key places where users share opinions,
ask questions, and discuss updates related to cryptocurrency technologies and market changes. The
amount, variety, and speed of user-generated content create a significant and challenging issue for
researchers who want to analyze large data streams consisting of text, images, and videos. Previous
studies show that social media discussions about cryptocurrencies often reflect a mix of feelings, ranging
from positive and negative responses to neutral comments, factual statements, ads, and user questions
[
Classifying these sentiments can ofer useful insights into public opinion, new market trends, and
user behavior. Correctly identifying subjective and objective content has been found to aid
decisionmaking, improve market analysis, and allow better tracking of cryptocurrency-related conversations
[
In addition to opinionated content, cryptocurrency discussions also include question-and-answer
exchanges, where users seek guidance or clarification about investments, market conditions, or technical
details. Often, community responses can be incomplete, irrelevant, or misleading. Thus, determining
whether a response efectively addresses the posed question is crucial for ensuring reliable access to
information in cryptocurrency communities [
To tackle these issues, the aim of the proposed research is focused on creating systems that can perform two main tasks: (i) detailed classification of cryptocurrency-related posts into a hierarchical set of labels that diferentiate noise, objective information, and subjective opinions, and (ii) assessing the relevance of responses in question-and-answer pairs. These tasks aim to support automated and ongoing monitoring of social media discussions, fostering a more organized understanding of cryptocurrency narratives across various platforms.
There are three sources of datasets, namely Reddit, Twitter, and YouTube, which contain social media posts related to cryptocurrency. Now this dataset is again divided for classification and Q&A tasks, respectively. The classification dataset has three-level annotations: 1. Level 1: In level 1, there are three classes: Noise, Objective, and Subjective, and these three classes are marked with 0, 1, and 2, respectively. 2. Level 2: In this level, the Subjective class is further divided into three categories: Neutral,
Negative, and Positive. These are marked with 0, 1, and 2, respectively, in the dataset. 3. Level 3: In the last level, there are four classes: Neutral Sentiments, Questions, Advertisement, and Miscellaneous, and these are marked with 0, 1, 2, and 3, respectively. This set of classes is branched from the Neutral category in level 2.
The hierarchical data distribution in the Twitter, Reddit, and YouTube datasets is shown in Figure 1. The Q&A task has a total of 31,692 samples across all the data sources (Twitter, Reddit, and YouTube combined). This dataset is further classified as Relevant or non-relevant, with 25,369 and 6,323 samples from the training and test sets, respectively. The dataset contains 3,704 datapoints for the Relevant class and 21,531 datapoints for the Non-relevant class.
The training data provided for this evaluation contain posts from Reddit, Twitter, and YouTube, annotated across three hierarchical levels and mapped to eight final categories. The distribution for each platform is reported in Figures 2, 3, and 4.
For Reddit, the Level 1 labels contain 645 Noise, 503 Objective, and 3,852 Subjective posts. In Level 2, the Subjective class is divided into 259 Positive, 410 Negative, and 3,183 Neutral posts. In Level 3, the Neutral posts are further split into 476 Neutral Sentiments, 2390 Questions, 105 Advertisements, and 212 Miscellaneous samples.
For Twitter, the Level 1 distribution includes 1,338 Noise, 1,702 Objective, and 1,248 Subjective posts. These Subjective posts are divided in Level 2 into 270 Positive, 78 Negative, and 902 Neutral posts. The Level 3 breakdown of Neutral posts consists of 178 Neutral Sentiments, 136 Questions, 544 Advertisements, and 46 Miscellaneous entries.
For YouTube, Level 1 consists of 786 Noise, 32 Objective, and 4,182 Subjective posts. In Level 2, the Subjective posts include 207 Positive, 1,574 Negative, and 2,401 Neutral samples. Level 3 further categorizes the Neutral posts into 1,391 Neutral Sentiments, 1,000 Questions, 1 Advertisement, and 9 Miscellaneous posts.
The test split contains 500 posts each from Reddit, Twitter, and YouTube, totaling 1,500 posts across platforms. Each platform follows the same three-level hierarchical label design as the training data, ensuring consistency for evaluation.
For Q&A, the dataset includes 6,323 paired entries consisting of a question, its corresponding comment, and a binary relevance label indicating whether the comment addresses the question.
Task 1 is to develop a classification model to classify cryptocurrency-related social media posts into eight classes, namely, Noise, Objective, Positive, Negative, Neutral, Question, Advertisement, and Miscellaneous.
Task 2 required participants to identify all answers relevant to a given question on cryptocurrency.
There are final submissions from five teams representing various academic institutions in the CryptOQA
shared task at FIRE 2025, focusing on classifying social media posts related to cryptocurrency. The
varied strategies and advanced mathematical models employed by the respective teams to deal with the
given task are mentioned below:
1. Team KLU (Koneru Lakshmaiah University) submitted a multi-stage transformer architecture
built on DeBERTa-V3-Small [
The submitted solutions across teams participating in the CryptOQA shared task employed a range of techniques for classifying cryptocurrency-related social media posts and determining the relevance of question–answer pairs. These methodologies can be broadly categorized into four techniques, namely, transformer-based models, hierarchical classification, recurrent neural architectures (LSTM/GRU), and parameter-eficient LLM fine-tuning. To tackle the challenges posed by noisy, multi-platform data and the multi-level label structure, each team adopted one or more of these methodological strategies.
Transformer-based Models Transformer architectures remained the dominant choice due to their ability to capture long-range dependencies and contextual relationships through self-attention. These models are particularly well-suited for domain-specific and informal language prevalent in cryptocurrency discourse. Several teams fine-tuned pre-trained transformer encoders, such as DeBERTa and BERT, for both subtasks.
• DeBERTa-V3-Small was used by Team KLU, integrating multiple task-specific classification heads and specialized loss functions to improve discriminability across the hierarchical label space. • BERT and CryptoBERT were employed by Team NLPFusion in a multi-task learning framework, allowing the model to jointly learn representations for all hierarchical levels and Q&A relevance prediction. • BERT-base (uncased) was utilized by Team MSEC as the primary encoder for a single unified hierarchical classifier fine-tuned on cleaned and normalized text from all platforms.
Hierarchical Classification Several teams incorporated the hierarchical structure of the CryptOQA labels directly into their modeling pipelines. Hierarchical classification enables predictions to be made at increasing levels of granularity while reducing confusion between semantically distant classes. • Hierarchical multi-head design was implemented by Team KLU, in which Level 1 outputs are conditionally routed to deeper classification heads at Levels 2 and 3, enabling fine-grained discrimination among sentiment and question categories. • Multi-task hierarchical modeling was adopted by Team NLPFusion, where shared transformer encoders simultaneously optimize for Level 1, Level 2, Level 3, and Q&A loss functions, leveraging inter-level dependencies without requiring separate models for each stage. • Recursive hierarchical prediction was incorporated in Team MSEC’s BERT-based system, where all hierarchical outputs are produced within a single unified classifier, simplifying the inference pipeline.
Recurrent Neural Models (LSTM/GRU) Recurrent neural networks were also explored due to their efectiveness in modeling sequential dependencies and sentiment-bearing patterns in short social media posts. These models remain competitive for noisy and domain-specific text when combined with appropriate embeddings.
• GRU-based hierarchical classifiers were adopted by Team SVNIT, enhanced with attention mechanisms to focus on discriminative tokens across levels. • Siamese GRU networks were applied by the same team for the Q&A task, embedding questions and answers separately and computing their semantic alignment for relevance prediction. • LSTM-based architectures were explored in initial experimentation across some teams, although
GRU variants generally provided more stable performance in the final submitted runs.
Parameter-eficient LLM Fine-tuning With the availability of compact open-weight LLMs, teams also explored parameter-eficient tuning approaches to reduce computational cost while retaining strong contextual understanding.
• Gemma-1B was adapted by Team IISER by replacing the model’s output projection with an eight-class head and updating only the final transformer block along with normalization layers, significantly reducing training overhead. • Selective layer unfreezing and lightweight optimization strategies were employed to stabilize fine-tuning on the small task-specific datasets while mitigating overfitting.
These approaches were accompanied by extensive data preprocessing steps, including text cleaning, noise filtering, token normalization, and class-weighted loss functions, to compensate for the limited parameter updates.
KLU NLP Fusion
IISER SVNIT MSEC
Reddit 1.0000 0.5694 0.6102 0.4744 0.2881
Avg. 1.0000 0.6254 0.6029 0.5222 0.2500
Task 2
The results from the CryptOQA 2025 shared task highlight that transformer-based models demonstrate clear superiority in categorizing cryptocurrency-related social media posts. Team KLU achieved the best performance in Task 1, attaining a perfect macro F1 score of 1.00 across Reddit, Twitter, and YouTube using a multi-stage DeBERTa-V3–based hierarchical architecture enriched with multiple loss functions and ensembling. Their system substantially outperformed the remaining submissions. Team NLPFusion followed with a macro F1 of 0.6253, supported by a multi-task transformer framework using BERT and CryptoBERT encoders, while Team IISER’s parameter-eficient adaptation of Gemma-1B secured a close macro F1 of 0.6028. Recurrent neural models performed moderately well, with Team SVNIT reporting a macro F1 score of 0.5222, indicating that GRU-based architectures can still be competitive on noisier platforms, despite lacking the representational power of transformers. Team MSEC’s unified BERTbased classifier achieved a macro F1 of 0.2500, indicating dificulty in handling the deeper sentiment granularity required at Level 2 and Level 3. Overall, transformer-driven approaches remained the most successful for hierarchical classification.
For the Question and Answering scenario in Task 2, the findings exhibit clearer variation across architectures. Team KLU again attained the highest score with a macro F1 of 1.00, demonstrating the strength of supervised contrastive and multi-loss training strategies for semantic relevance prediction. Team NLPFusion achieved the next best performance with 0.7940, followed by Team SVNIT’s Siamese GRU network, which obtained 0.7575, reflecting that recurrent similarity models remain efective for paired-sentence inference. Team IISER’s lightweight Gemma-1B fine-tuned system reported 0.7432, showing that selective parameter tuning can still yield strong generalization. Team MSEC did not submit a ranked run for Task 2. In general, transformer-based systems dominated Task 1 with significantly higher scores, while both transformers and recurrent Siamese models showed competitive performance in the Q&A task. All contributions and their final results are listed in Table 1.
The CryptOQA 2025 shared task aimed to evaluate machine learning and NLP approaches for hierarchical post classification and Q&A relevance detection in cryptocurrency-related social media content. Transformer-based models have proven to be most efective, with architectures built on DeBERTa-V3, BERT, and CryptoBERT delivering the strongest results, exemplified by Team KLU’s multi-stage transformer system achieving perfect macro F1 scores in both tasks. While recurrent architectures, such as GRU-based classifiers, performed reliably, they remained less efective than transformers for making deeper sentiment distinctions. Parameter-eficient LLM tuning, as seen with Gemma-1B, also showed competitive potential, particularly under computational constraints. Hierarchical modeling benefited from multi-task transformer strategies, though label imbalance and fine-grained categories continued to pose challenges. For Q&A relevance, both transformer models and Siamese GRU networks produced strong outcomes, with transformers ultimately leading.
During the preparation of this work, the author(s) used Grammarly in order to: Grammar and spelling check, and reword. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content. [19] Y. Wu, M. Schuster, Z. Chen, Q. V. Le, M. Norouzi, W. Macherey, M. Krikun, Y. Cao, Q. Gao, K. Macherey, et al., Google’s Neural Machine Translation System: Bridging the Gap Between Human and Machine Translation, arXiv:1609.08144 (2016). [20] I. Loshchilov, F. Hutter, Decoupled Weight Decay Regularization, arXiv:1711.05101 (2017).