We describe a novel unsupervised approach to subtitle segmentation, based on pretrained masked language models, where line endings and subtitle breaks are predicted according to the likelihood of punctuation to occur at candidate segmentation points. Our approach obtained competitive results in terms of segmentation accuracy across metrics, while also fully preserving the original text and complying with length constraints. Although supervised models trained on in-domain data and with access to source audio information can provide better segmentation accuracy, our approach is highly portable across languages and domains and may constitute a robust off-the-shelf solution for subtitle segmentation.

Adaptive Machine Translation purports to dynamically include user feedback to improve translation quality. In a post-editing scenario, user corrections of machine translation output are thus continuously incorporated into translation models, reducing or eliminating repetitive error editing and increasing the usefulness of automated translation. In neural machine translation, this goal may be achieved via online learning approaches, where network parameters are updated based on each new sample. This type of adaptation typically requires higher learning rates, which can affect the quality of the models over time. Alternatively, less aggressive online learning setups may preserve model stability, at the cost of reduced adaptation to user-generated corrections. In this work, we evaluate different online learning configurations over time, measuring their impact on user-generated samples, as well as separate in-domain and out-of-domain datasets. Results in two different domains indicate that mixed approaches combining online learning with periodic batch fine-tuning might be needed to balance the benefits of online learning with model stability.

We aimed to predict an individual suicide risk level from longitudinal posts on Reddit discussion forums. Through participating in a shared task competition hosted by CLPsych2019, we received two annotated datasets: a training dataset with 496 users (31,553 posts) and a test dataset with 125 users (9610 posts). We submitted results from our three best-performing machine-learning models: SVM, Naïve Bayes, and an ensemble model. Each model provided a user’s suicide risk level in four categories, i.e., no risk, low risk, moderate risk, and severe risk. Among the three models, the ensemble model had the best macro-averaged F1 score 0.379 when tested on the holdout test dataset. The NB model had the best performance in two additional binary-classification tasks, i.e., no risk vs. flagged risk (any risk level other than no risk) with F1 score 0.836 and no or low risk vs. urgent risk (moderate or severe risk) with F1 score 0.736. We conclude that the NB model may serve as a tool for identifying users with flagged or urgent suicide risk based on longitudinal posts on Reddit discussion forums.