Weakly supervised learning aims to construct effective predictive models from imperfectly labeled data. The recent trend of weakly supervised learning has focused on how to learn an accurate classifier from completely unlabeled data, given little supervised information such as class priors. In this paper, we consider a newly proposed weakly supervised learning problem called multi-class classification from multiple unlabeled datasets, where only multiple sets of unlabeled data and their class priors (i.e., the proportions of each class) are provided for training the classifier. To solve this problem, we first propose a classifier-consistent method (CCM) based on a probability transition matrix. However, CCM cannot guarantee risk consistency and lacks of purified supervision information during training. Therefore, we further propose a risk-consistent method (RCM) that progressively purifies supervision information during training by importance weighting. We provide comprehensive theoretical analyses for our methods to demonstrate the statistical consistency. Experimental results on multiple benchmark datasets and various prior matrices demonstrate the superiority of our proposed methods.

The learning to defer (L2D) framework allows models to defer their decisions to human experts. For L2D, the Bayes optimality is the basic requirement of theoretical guarantees for the design of consistent surrogate loss functions, which requires the minimizer (i.e., learned classifier) by the surrogate loss to be the Bayes optimality. However, we find that the original form of Bayes optimality fails to consider the dependence between the model and the expert, and such a dependence could be further exploited to design a better consistent loss for L2D. In this paper, we provide a new formulation for the Bayes optimality called dependent Bayes optimality, which reveals the dependence pattern in determining whether to defer. Based on the dependent Bayes optimality, we further present a deferral principle for L2D. Following the guidance of the deferral principle, we propose a novel consistent surrogate loss. Comprehensive experimental results on both synthetic and real-world datasets demonstrate the superiority of our proposed method.

In conventional supervised classification, true labels are required for individual instances. However, it could be prohibitive to collect the true labels for individual instances, due to privacy concerns or unaffordable annotation costs. This motivates the study on classification from aggregate observations (CFAO), where the supervision is provided to groups of instances, instead of individual instances. CFAO is a generalized learning framework that contains various learning problems, such as multiple-instance learning and learning from label proportions. The goal of this paper is to present a novel universal method of CFAO, which holds an unbiased estimator of the classification risk for arbitrary losses—previous research failed to achieve this goal. Practically, our method works by weighing the importance of each instance and each label in the group, which provides purified supervision for the classifier to learn. Theoretically, our proposed method not only guarantees the risk consistency due to the unbiased risk estimator but also can be compatible with arbitrary losses. Extensive experiments on various problems of CFAO demonstrate the superiority of our proposed method.

In the construction of multiple-choice Machine Reading Comprehension(MRC), in addition to questions and answers, it is necessary to generate distractors corresponding to them. Recent models based on Seq2Seq have shown good results in text generation, while previous work has only succeeded in producing a few interfering words or phrases per question. Our goal is to generate more meaningful distractors based on reading comprehension of articles that are closer to the semantics of the question to help better diagnose gaps in text understanding. A major drawback of recent studies is that they do not take the relationship between the distractors and background text into account when generating distractors. This often results in the distractors being either too general or too close to the correct answer. We propose a Topic Enhanced Multi-head Co-Attention model (TMCA) based on hierarchical networks to better capture the interactions between sentences. By adding query-relevance loss, our model enables the distractors to be as semantically relevant to the question as possible based on the reading comprehension of the article, while trying to ensure that they are false answers. The results show that the proposed approach achieves superior performance over the baselines in terms of automatic metrics on two multiple-choice datasets (RACE and DREAM). For further evaluation, we use another advanced reading comprehension model and human evaluation to demonstrate that our model outperforms several strong baselines in generating high-quality and educationally meaningful distracters.

Aspect-based sentiment analysis (ABSA) aims to detect the sentiment polarities of a sentence with a given aspect. Aspect-term sentiment analysis (ATSA) is a subtask of ABSA, in ATSA, the aspect is given by aspect term: a word or a phrase in sentence. In previous work, most models apply attention mechanism or gating mechanism to capture the key part of the sentence and detect the sentiment polarity by classifying the weighted sum vector, which are regardless of the aspect information during the classification. However, same contexts may show different sentiment polarity with differnet aspects. For example, in sentence “The scene hunky waiters dub diner darling and it sounds like they mean it.” the context “dub diner dearling” shows negative polarity towards the aspect term “waiters”. But in sentence “The diner's husband dub diner darling.”, the same context would show neural polarity towards “husband”. The absent of aspect information would make model get a wrong result in some sentence. To solve this problem, we propose a model called Trasfomer-based context-aspect Relation Detect Model (TRDM), and add two special tokens “[CLSA]” and “[SEPA]” to specify the aspect term in the sentence. TRDM uses the embedding of two tyes of tokens (i.e. “[CLS]” and “[CLSA]”) for classification where “[CLS]” is used to represent the entire sentence in BERT. This scheme enable TRDM to combine the sentence information and aspect information in the step of sentiment polarity classification. We evaluate the performance of our model on two datasets: Restaurant dataset from SemEval2014 and MAMS from NLPCC 2020. Experiment results show that our model obtain noticeable improvement compared with state-of-art transfromer-based models.