Question generation (QG) approaches based on large neural models require (i) large-scale and (ii) high-quality training data. These two requirements pose difficulties for specific application domains where training data is expensive and difficult to obtain. The trained QG models’ effectiveness can degrade significantly when they are applied on a different domain due to domain shift. In this paper, we explore an unsupervised domain adaptation approach to combat the lack of training data and domain shift issue with domain data selection and self-training. We first present a novel answer-aware strategy for domain data selection to select data with the most similarity to a new domain. The selected data are then used as pseudo-in-domain data to retrain the QG model. We then present generation confidence guided self-training with two generation confidence modeling methods (i) generated questions’ perplexity and (ii) the fluency score. We test our approaches on three large public datasets with different domain similarities, using a transformer-based pre-trained QG model. The results show that our proposed approaches outperform the baselines, and show the viability of unsupervised domain adaptation with answer-aware data selection and self-training on the QG task.
Quality control is essential for creating extractive question answering (EQA) datasets via crowdsourcing. Aggregation across answers, i.e. word spans within passages annotated, by different crowd workers is one major focus for ensuring its quality. However, crowd workers cannot reach a consensus on a considerable portion of questions. We introduce a simple yet effective answer aggregation method that takes into account the relations among the answer, question, and context passage. We evaluate answer quality from both the view of question answering model to determine how confident the QA model is about each answer and the view of the answer verification model to determine whether the answer is correct. Then we compute aggregation scores with each answer’s quality and its contextual embedding produced by pre-trained language models. The experiments on a large real crowdsourced EQA dataset show that our framework outperforms baselines by around 16% on precision and effectively conduct answer aggregation for extractive QA task.
According to the Probability Ranking Principle (PRP), ranking documents in decreasing order of their probability of relevance leads to an optimal document ranking for ad-hoc retrieval. The PRP holds when two conditions are met: [C1] the models are well calibrated, and, [C2] the probabilities of relevance are reported with certainty. We know however that deep neural networks (DNNs) are often not well calibrated and have several sources of uncertainty, and thus [C1] and [C2] might not be satisfied by neural rankers. Given the success of neural Learning to Rank (LTR) approaches—and here, especially BERT-based approaches—we first analyze under which circumstances deterministic neural rankers are calibrated for conversational search problems. Then, motivated by our findings we use two techniques to model the uncertainty of neural rankers leading to the proposed stochastic rankers, which output a predictive distribution of relevance as opposed to point estimates. Our experimental results on the ad-hoc retrieval task of conversation response ranking reveal that (i) BERT-based rankers are not robustly calibrated and that stochastic BERT-based rankers yield better calibration; and (ii) uncertainty estimation is beneficial for both risk-aware neural ranking, i.e. taking into account the uncertainty when ranking documents, and for predicting unanswerable conversational contexts.
Knowledge tracing serves as a keystone in delivering personalized education. However, few works attempted to model students’ knowledge state in the setting of Second Language Acquisition. The Duolingo Shared Task on Second Language Acquisition Modeling provides students’ trace data that we extensively analyze and engineer features from for the task of predicting whether a student will correctly solve a vocabulary exercise. Our analyses of students’ learning traces reveal that factors like exercise format and engagement impact their exercise performance to a large extent. Overall, we extracted 23 different features as input to a Gradient Tree Boosting framework, which resulted in an AUC score of between 0.80 and 0.82 on the official test set.