Jiayuan He

Aparna Elangovan, Jiayuan He, Yuan Li et al.

The NLP community typically relies on performance of a model on a held-out test set to assess generalization. Performance drops observed in datasets outside of official test sets are generally attributed to "out-of-distribution" effects. Here, we explore the foundations of generalizability and study the factors that affect it, articulating lessons from clinical studies. In clinical research, generalizability is an act of reasoning that depends on (a) internal validity of experiments to ensure controlled measurement of cause and effect, and (b) external validity or transportability of the results to the wider population. We demonstrate how learning spurious correlations, such as the distance between entities in relation extraction tasks, can affect a model's internal validity and in turn adversely impact generalization. We, therefore, present the need to ensure internal validity when building machine learning models in NLP. Our recommendations also apply to generative large language models, as they are known to be sensitive to even minor semantic preserving alterations. We also propose adapting the idea of matching in randomized controlled trials and observational studies to NLP evaluation to measure causation.

Aparna Elangovan, Jiayuan He, Yuan Li et al.

BERT-based models have had strong performance on leaderboards, yet have been demonstrably worse in real-world settings requiring generalization. Limited quantities of training data is considered a key impediment to achieving generalizability in machine learning. In this paper, we examine the impact of training data quality, not quantity, on a model's generalizability. We consider two characteristics of training data: the portion of human-adversarial (h-adversarial), i.e., sample pairs with seemingly minor differences but different ground-truth labels, and human-affable (h-affable) training samples, i.e., sample pairs with minor differences but the same ground-truth label. We find that for a fixed size of training samples, as a rule of thumb, having 10-30% h-adversarial instances improves the precision, and therefore F1, by up to 20 points in the tasks of text classification and relation extraction. Increasing h-adversarials beyond this range can result in performance plateaus or even degradation. In contrast, h-affables may not contribute to a model's generalizability and may even degrade generalization performance.

Jiayuan He, Jianzhong Qi, Kotagiri Ramamohanarao

Trip recommendation is an important location-based service that helps relieve users from the time and efforts for trip planning. It aims to recommend a sequence of places of interest (POIs) for a user to visit that maximizes the user's satisfaction. When adding a POI to a recommended trip, it is essential to understand the context of the recommendation, including the POI popularity, other POIs co-occurring in the trip, and the preferences of the user. These contextual factors are learned separately in existing studies, while in reality, they impact jointly on a user's choice of a POI to visit. In this study, we propose a POI embedding model to jointly learn the impact of these contextual factors. We call the learned POI embedding a context-aware POI embedding. To showcase the effectiveness of this embedding, we apply it to generate trip recommendations given a user and a time budget. We propose two trip recommendation algorithms based on our context-aware POI embedding. The first algorithm finds the exact optimal trip by transforming and solving the trip recommendation problem as an integer linear programming problem. To achieve a high computation efficiency, the second algorithm finds a heuristically optimal trip based on adaptive large neighborhood search. We perform extensive experiments on real datasets. The results show that our proposed algorithms consistently outperform state-of-the-art algorithms in trip recommendation quality, with an advantage of up to 43% in F1-score.