In the dynamic landscape of cryptocurrency, understanding public sentiment is essential for tracking market trends and investor behavior. Social media platforms like Twitter, Reddit, YouTube and Facebook have become key spaces where users express diverse and often unfiltered opinions about cryptocurrencies. However, the informal and unstructured nature of social media content presents significant challenges for accurate sentiment analysis. Posts vary widely in tone, language, and format, making it dificult to categorize sentiments with precision. To tackle these complexities, the shared task "CryptOQA: Opinion Extraction and Question Answering from Crypto Currency-Related Tweets and Reddit posts" organized at FIRE 2025, calls on the research community to develop innovative solutions. The goal is to enhance the identification and classification of cryptocurrencyrelated opinions and questions expressed in English across social platforms. To address these challenges, this paper introduces Multi-task Learning (MTL) models that leverage pretrained transformers to efectively capture the hierarchical structure of data for Task 1. For Task 2, we implement question-answering frameworks built upon pretrained transformer models to understand and classify cryptocurrency-related questions from social media posts. Our best-performing model for the opinion classification task achieved macro F1 scores of 0.5694, 0.5570, and 0.7496 on the Reddit, Twitter, and YouTube datasets, respectively, as provided by the task organizers. Additionally, our top model for Task 2 attained a macro F1 score of 0.7940, demonstrating the efectiveness of our approach in handling diverse and unstructured social media content.
Cryptocurrency is a paradigm change in the world of finance, defined by digital or virtual money that
uses cryptographic methods to secure transactions [
Social media websites have emerged as the primary source of information and public sentiment,
giving immediate insights into attitudes and behavior [
To tackle the challenges of identifying cryptocurrency-related opinion posts on social media, the "CryptOQA: Opinion Extraction and Question Answering from CryptoCurrency-Related Tweets and Reddit posts" shared task was organized at FIRE 2025. This initiative invited the research community to develop models capable of detecting diferent categories of cryptocurrency posts. The shared task included two components: Task 1, focused on opinion classification from cryptocurrency-related social media posts, and Task 2, centered on Question Answering (QA) from such posts. The dataset provided by the organizers comprised 5,000 posts per platform and was annotated across three hierarchical levels, as illustrated in Figure 1. This structured dataset ofered a foundation for developing models to accurately capture the nuances of cryptocurrency-related opinions expressed on social media. To address the challenges provided by the shared task, MTL models that utilize pretrained transformers to capture the hierarchical structure of the given data are implemented. Two separate runs are submitted to the organizers, each based on diferent transformer models: BERT and CryptoBERT. Among these, the model ifne-tuned using CryptoBERT achieved the best performance across the provided datasets, demonstrating its efectiveness in handling cryptocurrency-related content. For Task 2, we developed two questionanswering frameworks built on BERT and DistilBERT to identify and classify cryptocurrency-related questions from social media posts. These models performed relevance classification to determine the relationship between questions and candidate answers. Overall, the proposed approaches show promising results in both opinion categorization and question-answering tasks.
The rest of the paper is organized as follows: Section 2 sheds light on related works, and methodology is discussed in Section 3. Experimental results are presented in Section 4 and the paper concludes in Section 5 along with spreading light on future works.
The market for cryptocurrencies has grown very fast, presenting challenges and benefits for machine learning-style data analysis. To have any success in this extremely volatile market, it is crucial to understand the sentiment of the public and predict market directions. Numerous studies have shown that machine learning, particularly deep learning, can be used to address this issue, with a focus on sentiment analysis of social media posts. A number of research works have investigated the role of social media sentiment in cryptocurrency price forecasting with encouraging outcomes. Among such works, one presented an MAPE of 1.3251 as evidence of its suitability for predicting the price of Bitcoin (BTC). Sahal [11] introduced a method that combines Bidirectional LSTM (BiLSTM) networks and Embeddings from Language Models (ELMo) in examining Twitter data to determine profitable cryptocurrencies using contextual sentiment analysis. Their model was able to produce a high accuracy of 86.30%, demonstrating its capability in forecasting price variations afected by social media. In another highly cited paper, Huang et al. [12] looked into sentiment analysis on Sina-Weibo, a well-known Chinese social networking site, to forecast cryptocurrency price volatility. They constructed a crypto-oriented sentiment dictionary and used an LSTM-based Recurrent Neural Network (RNN) and autoregressive models, incorporating sentiment with past price information. The LSTM model performed better than conventional autoregressive models, recording a precision of 0.87 and a recall of 0.94, testifying to its performance in reflecting market movements.
Das et al. [13] proposed an explanatory multitasking model for analysis in financial markets by combining sentiment, emotion, and cause extraction from social media text. To this end, Hegde and Shashirekha [14] constructed the FinEMA dataset by merging existing financial news/tweet datasets with 6,500 newly scraped and manually labeled tweets containing sentiment, emotion, and cause annotations. Their methodology utilizes FinBERT embeddings and a hierarchical Emotion-Sentiment Attention Network (ESAN), supplemented with a CentralNet-based multitasking framework, for modeling sentiment classification, emotion detection, and cause identification jointly. The experimental results indicate that ESAN attained 89% accuracy for sentiment classification and 79% for emotion detection, surpassing baselines like BERT, CNN-BiLSTM, and ensemble LSTM-GRU models.
The associated research emphasizes that scholars have used various methods, such as LSTM-GRU ensembles, self-attention mechanisms, and hierarchical graph neural networks, to examine cryptocurrency sentiments. These models make use of varied features and information sources, such as tweets and financial metrics, to analyze intricate market relationships. Although, they have exhibited robust performance in sentiment classification and trend identification, the highly volatile nature of cryptocurrency markets introduces new challenges. In addition, volatile and loud social media content makes the identification of sentiments even more challenging. Therefore, considerable potential exists to develop methodologies and find novel techniques to comprehend cryptocurrency market dynamics.
The proposed approach introduces two distinct methodologies for Task 1 and Task 2. For Task 1, a hierarchical classification model is built in order to structure labels in a multi-level format, in which there can be more specific and context-sensitive classification by taking into account relationships at various label levels. For Task 2, a relevance classification model is used to train the correspondence between a question and its respective answers, categorizing whether an answer is relevant or not with respect to the question. These approaches collectively try to enhance both answer relevance detection and label organization in the given tasks. It may be noted that common pre-processing for both Task 1 and Task 2 is employed in this paper. The description of each of these is provided in the below subsections (3.2 and 3.3).
Pre-processing operation is applied to remove noise from and standardize social media text for opinion extraction and QA [15]. A pre-processing pipeline is implemented that begins with the elimination of user mentions (e.g., *@username*) that appear frequently in social media updates but contribute negligibly to semantic meaning. The text is then converted to lowercase to ensure consistency and eliminate redundancy in word representation. Character-wise normalization is also performed to achieve standardized formatting across the dataset.
MTL is a machine learning paradigm in which a model learns to undertake a number of associated tasks simultaneously by sharing some layers and parameters. This approach is particularly beneficial when there is not enough data for individual but similar tasks. MTL allows the model to learn common lowerlevel features while still acquiring task-specific higher-level features. With an integrated framework, it enhances generalization and increases prediction accuracy [16].
The model for Task 1 proposed in this study is a typical MTL design and uses pre-trained transformer models to process input sequences and derive text representations that have meaningful contextual and semantic information [17, 18]. These representations of context and semantics are then pooled across tasks and used for task-specific classification, as shown in Figure 2 (a). By transforming text to feature vectors, the model is able to discern patterns and relationships more efectively, resulting in better accuracy in classification. Two transformer-based models, namely BERT and CryptoBERT, are employed in this work, with Hugging Face’s ‘BertTokenizer‘ and ‘BertModel‘ for tokenization and loading pre-trained models. Description of each of these transformer-based models is given below: • CryptoBERT - is a transformer-based model that has been borrowed from BERT and fine-tuned for cryptocurrency text data. The model is specifically tailored to understand the distinctive vocabulary, jargon, and discourse tendencies present in cryptocurrency conversations on social media platforms such as Twitter and Reddit. This targeted training enables CryptoBERT to grasp sentiment and opinion more efectively in the crypto space, enhancing its performance in tasks including cryptocurrency sentiment prediction and opinion mining. • BERT - is a Google-developed pre-trained language model that has revolutionized natural language processing (NLP). It works to comprehend the context of words by taking into account their left and right contexts via a bidirectional attention mechanism. BERT can be used for various applications, including text classification, QA, and sentiment analysis, as it ofers rich contextual embeddings.
As illustrated in Figure 2 (a), the MTL architecture under consideration utilizes contextualized representations produced by transformer-based models to process the input sequence. These are then fed to two distinct Recurrent Neural Network (RNN) modules that specialize in processing a given subtask. The embeddings coming out of the RNNs are utilized to produce probability distributions for target labels of a given task. For improved learning, the outputs are fed through a Feed Forward Network (FFN), which adds non-linearity and better picks up the more intricate patterns within the data. The produced logits are further processed using a softmax layer for translation into probability values for the final predictions. The total loss is the weighted combination of individual losses for each subtask, = ∑︀
=1 , where is the number of labels and is the associated loss weight. (a) Task 1 (b) Task 2 This architecture allows the model to share feature vectors among tasks to enhance generalization and task-specific performance.
In QA, relevance classification is responsible for classifying a candidate text (e.g., answer or passage) as relevant or irrelevant to a given question, often applying binary or graded labels [19]. Relevance classification is important in filtering and ranking out content that actually answers the question, thus enhancing answer quality and user experience. For example, in Community QA settings, numerous responses can be of low quality or unrelated, so relevance measures become crucial for useful answers’ identification. Work on community QA demonstrates how relevance (as well as measurements such as similarity and completeness) is central to evaluating candidate answers’ quality [20].
In this study, we tackle the task of relevance classification between a question and its respective comment or answer in social media text on cryptocurrency and the proposed framework is shown in Figure 2 (b). The dataset has three major columns: MAIN (the question), comment_body (the comment/answer), and relevance (binary labels on whether the question-answer pair is relevant). To preprocess the data, text from both the MAIN and comment_body columns was cleaned (as mentioned in Subsection 3.1). Further, all values are transformed into strings and missing entries are replaced with empty strings. The labels were also standardized into integers for consistency in training. For training and evaluation purposes, the dataset was transformed into Hugging Face Dataset objects to ensure compatibility with the transformers library and enable eficient tokenization and batching. This configuration enabled us to frame the task as a binary classification problem in which the model predicts whether a comment pertains to a particular question.
In order to encode the text pairs, we employed pre-trained transformer-based models (DistilBERT and BERT models), which yields contextual embeddings for sequence classification tasks. Description of DistilBERT is given below (description of BERT model is provided in Subsection 3.2): • DistilBERT - is a lower, quicker, and lighter alternative to BERT, developed through knowledge distillation to capture nearly all of BERT’s performance but with much lower computational costs. It has approximately 40% fewer parameters than BERT and processes around 60% more quickly, for which it is particularly suited for real-time or for resource-limited applications. In spite of its small size, DistilBERT retains about 97% of BERT’s language comprehension ability, with excellent performance across a broad variety of natural language processing tasks including text classification, sentiment analysis, and question answering. Its trade-of of eficiency and accuracy has earned it a coveted spot among both academic research and industry applications.
We passed MAIN (question) and comment_body (answer) concurrently to the tokenizer, with MAIN being the main sequence and comment_body as the paired sequence. Tokenization, truncation, and padding were accomplished by the Hugging Face tokenizer, converting input text pairs into input_ids and attention_masks that are ingestible by the model. Transformer models took these inputs and processed them to create hidden representations, which were fed into a classification head with a dense layer that produced binary labels.
A set of experiments were conducted across various transformer models to solve Task 1 using the MTL framework and Task 2 using relevance classification models. Hyperparameters and their respective values applied to both models are shown in Tables 1 and 2. Models with excellent performance on the validation sets were also tested on the test sets. Since the datasets are not balanced, the Macro F1 score was adopted as the metric of evaluation, and results on the performance of the proposed models are presented in Table 3.
According to the shared task guidelines, only predictions of two models from the Test sets could be submitted by participants. We thus trained our models on the given Train sets and obtained predictions on the Test sets provided by the organizers. The organizers used the macro F1 score as the measure of performance and evaluated our predictions. Our suggested MTL model, which uses CryptoBERT for text representation, posted competitive performance with macro F1 values of 0.5694, 0.5570, and 0.749, with 5th, 5th, and 3rd ranks on the Reddit, Twitter, and YouTube datasets, respectively, for Task 1. Globally, this model posted the 4th rank in Task 1. For Task 2, the suggested relevance classification model using BERT for text representation posted good macro F1 score of 0.7940 with 3rd rank. A comparison of macro F1 scores achieved by all the participating teams for Task 1 and Task 2 is depicted in Figures 3 and 4, respectively.
This paper describes the methods of our run submissions for the CryptOQA shared task at FIRE 2025 underscores the eficacy of transformer-based models in dealing with the issues of opinion extraction and QA from social media posts about cryptocurrencies. Through the application of MTL strategies and ifne-tuning pre-trained models like BERT and CryptoBERT, we were able to capture the hierarchical nature of opinions expressed on Twitter, Reddit, and YouTube. Our findings illustrate that domainspecific models such as CryptoBERT yield significant improvements over universal models, especially for dealing with subtle and unstructured conversations prevalent in the cryptocurrency community. On Task 1, our models reported competitive macro F1 scores across various platforms, indicating stability in separating distinct categories of opinion. Analogously, in Task 2, the proposed BERT and DistilBERT-based models achieved good performance in relevance classification, thus further confirming the viability of pretrained models for question-answering tasks in user-generated noisy text. Among the proposed models, our best-performing model for the opinion classification task achieved macro F1 scores of 0.5694, 0.5570, and 0.7496 on the Reddit, Twitter, and YouTube datasets, respectively, as provided by the task organizers. Further, our top model for Task 2 attained a macro F1 score of 0.7940. Although our achievements are promising, the current results still leave much room for improvement in dealing with sarcasm, code-switching, and implicit sentiment, which pervade social media language. Another constraint is the dependency on annotated datasets since annotated data at scale is hard and expensive to attain in the rapidly changing cryptocurrency space. Subsequent work must investigate semi-supervised and unsupervised approaches like self-training, co-training, or weak supervision to take advantage of the vast amounts of unlabeled social media data.
During the preparation of this work, the author(s) used ChatGPT-4 in order to:Grammar and spelling check. Using this tool, the author(s) reviewed and edited the content as needed and take full responsibility for the publication’s control [7] L. S. Kumar, A. Hegde, B. R. Chakravarthi, H. Shashirekha, R. Natarajan, S. Thavareesan, R. Sakuntharaj, T. Durairaj, P. K. Kumaresan, C. Rajkumar, Overview of second shared task on sentiment analysis in code-mixed tamil and tulu, in: Proceedings of the Fourth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages, 2024, pp. 62–70. [8] A. Hegde, G. Kavya, S. Coelho, P. Lamani, H. L. Shashirekha, Munlp@ dravidianlangtech2023: Learning approaches for sentiment analysis in code-mixed tamil and tulu text, in: Proceedings of the Third Workshop on Speech and Language Technologies for Dravidian Languages, 2023, pp. 275–281. [9] S. Sarkar, A. Badwal, A. Roy, K. Rudra, K. Ghosh, Cryptopiqa: A new opinion and question answering dataset on cryptocurrency, in: Proceedings of the 31st International Conference on Computational Linguistics, 2025, pp. 11107–11120. [10] A. Deroy, S. Maity, Cryptollm: Unleashing the power of prompted llms for smartqna and classification of crypto posts, arXiv preprint arXiv:2411.07917 (2024). [11] R. Sahal, Predicting Optimal Cryptocurrency using Social Media Sentimental Analysis, Ph.D. thesis,
Dublin, National College of Ireland, 2023. [12] X. Huang, W. Zhang, X. Tang, M. Zhang, J. Surbiryala, V. Iosifidis, Z. Liu, J. Zhang, Lstm based sentiment analysis for cryptocurrency prediction, in: International conference on database systems for advanced applications, Springer, 2021, pp. 617–621. [13] S. Das, U. Chowdhury, N. Lijin, A. Deep, S. Saha, A. Maurya, Investigate how market behaves: Toward an explanatory multitasking based analytical model for financial investments, IEEE Access 12 (2024) 30928–30940. [14] A. Hegde, H. Shashirekha, Transformer-driven multi-task learning for fake and hateful content detection, in: Proceedings of the 21st International Conference on Natural Language Processing (ICON): Shared Task on Decoding Fake Narratives in Spreading Hateful Stories (Faux-Hate), 2024, pp. 29–35. [15] S. Coelho, A. Hegde, P. Lamani, H. L. Shashirekha, et al., Mucsd@ dravidianlangtech2023: Predicting sentiment in social media text using machine learning techniques, in: Proceedings of the Third Workshop on Speech and Language Technologies for Dravidian Languages, 2023, pp. 282–287. [16] Y. Zhang, Q. Yang, A Survey on Multi-task Learning (2021) 5586–5609. [17] B. R. Chakravarthi, P. Kumaresan, R. Priyadharshini, P. Buitelaar, A. Hegde, H. Shashirekha, S. Rajiakodi, M. Á. García, S. M. Jiménez-Zafra, J. García-Díaz, et al., Overview of Third Shared Task on Homophobia and Transphobia Detection in Social Media Comments, in: Proceedings of the Fourth Workshop on Language Technology for Equality, Diversity, Inclusion, 2024, pp. 124–132. [18] A. Hegde, G. Kavya, S. Coelho, H. L. Shashirekha, MUCS@ LT-EDI2023: Homophobic/Transphobic Content Detection in Social Media Text using mBERT, in: Proceedings of the Third Workshop on Language Technology for Equality, Diversity and Inclusion, 2023, pp. 287–294. [19] X. Hu, Enhancing answer selection in community question answering with pre-trained and large language models. corr abs/2311.17502 (2023), 2023. [20] G. A. Boroujeni, H. Faili, Y. Yaghoobzadeh, Answer selection in community question answering exploiting knowledge graph and context information, Semantic Web 13 (2022) 339–356.