Filtering data in transformers

For decades, natural language processing (NLP) has provided methods for computers to understand language in a way that mimics humans. Since they are built on transformers, complex neural network layers, these large language models' decision making processes are usually incomprehensible to humans and require large amounts of data to be trained properly. In the past, researchers have tried to remedy this by having models explain their decisions by providing rationales, short excerpts of data that contributed most to the label. To create these models, large amounts of rational data are needed. In order to decrease the high costs of gathering data, active learning is often implemented to increase the efficiency of training data for NLP models.

While researching this topic, recent UT Austin graduate Didi Zhou found that although transformers are widely used for NLP tasks, active learning has not been put into practice with transformer models and rationales before. Overseen by Professor Matt Lease, Postdoctoral Researcher Venelin Kovachev, and PhD student Anubrata Das, Zhou was able to explore techniques that would make it possible to incorporate active learning into transformer-based NLP models with rationales. Eventually, Zhou’s work could allow for future developments providing transparency in decision making processes and further enable interactive machine learning where humans can correct machine rationales in decision making. 

In order to grasp the importance of Zhou's research, there first needs to be an understanding of how active learning works. Active learning is commonly used to decrease the amount of total data needed for a model while maintaining performance. It does this through identifying examples that provide the most amount of useful information to the model. The model is then trained with the pre-selected data, which significantly decreases the cost and time needed to label large amounts of data. 

Didi simulated active learning on a model that would both make a decision on its given task and give a rationale for that decision. Active learning was simulated by taking an existing labeled rationale dataset on fact-checking and withholding the labels from the model, only providing the label when the data is selected as important. To do this, first, a portion of the full data set is labeled and used to train the model. The remaining, unlabeled, portion is what the model is evaluated on. After this process occurs, the data is ranked based on different selection criteria. Several selection criteria were evaluated, some of which followed traditional active learning methods and were based on the model’s performance on the end task, predicting whether a claim is supported or refuted by a document, while others explored using the model’s predicted rationales. Top data points are removed from the unlabeled data group then placed into the labeled group. The model is then re-trained with this new labeled data and the process repeats itself until the set amount of data is used. In doing this, Zhou’s research emulated active learning, proving its effectiveness on transformer-based models with rationales.

In the future, this work could potentially help create models where humans could correct a model’s rationales to enable more interactive machine learning. Zhou’s research helps to bridge the gap between machine and human knowledge and has laid an important foundation for future work utilizing active learning on transformer models with rationales. Didi reflected on her experience doing research in the Turing Scholars Honor Program, and expressed appreciation for her mentors’ guidance, support from her cohort, and the opportunity to participate in such a valuable research experience. The Turing Scholars Honor Program is an honors program for undergraduate students working toward a degree in computer science. It enables students to enhance their research experience through close interaction with faculty and an opportunity to write an honors thesis. Zhou emphasized that while the thesis initially felt like a daunting task, it instilled confidence in her ability to produce meaningful research. Didi Zhou is now an associate product manager at Google and continues to explore her research interests, hoping to one day pursue the research field full-time.