Building for Tomorrow: Assessing the Temporal Persistence of Text Classifiers

Performance of text classification models can drop over time when new data to be classified is more distant in time from the data used for training, due to naturally occurring changes in the data, such as vocabulary change. A solution to this is to continually label new data to retrain the model, which is, however, often unaffordable to be performed regularly due to its associated cost. This raises important research questions on the design of text classification models that are intended to persist over time: do all embedding models and classification algorithms exhibit similar performance drops over time and is the performance drop more prominent in some tasks or datasets than others? With the aim of answering these research questions, we perform longitudinal classification experiments on three datasets spanning between 6 and 19 years. Findings from these experiments inform the design of text classification models with the aim of preserving performance over time, discussing the extent to which one can rely on classification models trained from temporally distant training data, as well as how the characteristics of the dataset impact this.