The objective of this study is to improve speech recognition performance for low-resource Hakka, a language spoken by a specific ethnic group. Our team conducted experiments by fine-tuning different base versions of Whisper (e.g., the original model and the Mandarin-focused Belle model). We found that fine-tuning on different bases yielded distinct advantages and varying results in Hakka character and phonetic recognition tasks. To further enhance model accuracy, we experimented with replacing the q, k, and v linear layers in the attention blocks of the Whisper encoder with a mixture-of-experts model combined with RoLA. In addition, we augmented the training data with synthesized speech generated with diverse voice styles and varying speaking rates. The results showed a 0.73% reduction in character error rate for Task 1 and a 0.2% reduction in word error rate for Task 2. These findings confirm that both architectural adjustments to the model and the strategic use of limited synthetic speech data in low-resource dialect corpora can effectively improve recognition performance.

A cancer registry is a critical and massive database for which various types of domain knowledge are needed and whose maintenance requires labor-intensive data curation. In order to facilitate the curation process for building a high-quality and integrated cancer registry database, we compiled a cross-hospital corpus and applied neural network methods to develop a natural language processing system for extracting cancer registry variables buried in unstructured pathology reports. The performance of the developed networks was compared with various baselines using standard micro-precision, recall and F-measure. Furthermore, we conducted experiments to study the feasibility of applying transfer learning to rapidly develop a well-performing system for processing reports from different sources that might be presented in different writing styles and formats. The results demonstrate that the transfer learning method enables us to develop a satisfactory system for a new hospital with only a few annotations and suggest more opportunities to reduce the burden of cancer registry curation.

In this paper, we described our systems for the first and second subtasks of Social Media Mining for Health Applications (SMM4H) shared task in 2020. The two subtasks are automatic classi-fication of medication mentions and adverse effect in tweets. Our systems for both subtasks are based on Robustly optimized BERT approach (RoBERTa) and our previous work at SMM4H’19. The best F1-scores achieved by our systems for subtask 1 and 2 were 0.7974 and 0.64 respec-tively, which outperformed the average F1-scores among all teams’ best runs by at least 0.13.

In this study, we describe our methods to automatically classify Twitter posts conveying events of adverse drug reaction (ADR). Based on our previous experience in tackling the ADR classification task, we empirically applied the vote-based under-sampling ensemble approach along with linear support vector machine (SVM) to develop our classifiers as part of our participation in ACL 2019 Social Media Mining for Health Applications (SMM4H) shared task 1. The best-performed model on the test sets were trained on a merged corpus consisting of the datasets released by SMM4H 2017 and 2019. By using VUE, the corpus was randomly under-sampled with 2:1 ratio between the negative and positive classes to create an ensemble using the linear kernel trained with features including bag-of-word, domain knowledge, negation and word embedding. The best performing model achieved an F-measure of 0.551 which is about 5% higher than the average F-scores of 16 teams.

The increasing popularity of social media lead users to share enormous information on the internet. This information has various application like, it can be used to develop models to understand or predict user behavior on social media platforms. For example, few online retailers have studied the shopping patterns to predict shopper’s pregnancy stage. Another interesting application is to use the social media platforms to analyze users’ health-related information. In this study, we developed a tree kernel-based model to classify tweets conveying pregnancy related information using this corpus. The developed pregnancy classification model achieved an accuracy of 0.847 and an F-score of 0.565. A new corpus from popular social media platform Twitter was developed for the purpose of this study. In future, we would like to improve this corpus by reducing noise such as retweets.

Traditional disease surveillance systems depend on outpatient reporting and virological test results released by hospitals. These data have valid and accurate information about emerging outbreaks but it’s often not timely. In recent years the exponential growth of users getting connected to social media provides immense knowledge about epidemics by sharing related information. Social media can now flag more immediate concerns related to out-breaks in real time. In this paper we apply the long short-term memory recurrent neural net-work (RNN) architecture to classify tweets conveyed influenza-related information and compare its performance with baseline algorithms including support vector machine (SVM), decision tree, naive Bayes, simple logistics, and naive Bayes multinomial. The developed RNN model achieved an F-score of 0.845 on the MedWeb task test set, which outperforms the F-score of SVM without applying the synthetic minority oversampling technique by 0.08. The F-score of the RNN model is within 1% of the highest score achieved by SVM with oversampling technique.