Most few-shot learning approaches can’t get satisfactory results in fine-grained multi-category sentiment classification tasks. To solve this problem, a joint learning approach is proposed to few-shot learning for multi-category sentiment classification. Specifically, we utilize the pre-trained token-replaced detection model as few-shot learners and concurrently reformulate fine-grained sentiment classification tasks as both classification and regression problems by appending classification and regression templates and label description words to the input at the same time. For joint learning, several fusion methods are proposed to fuse the classification prediction and regression prediction. Experimental results show that, compared to mainstream few-shot methods, the proposed approach apparently achieves better performances in F1-Score and accuracy rate.

This paper proposes a new translation quality estimation model that does not rely on the reference translation to score the translation of each sentence in the source language. The authors model the sentence-level semantic difference and word-level referential difference between the source and translation and design additional loss function to make the model constrain the differences as much as possible when predicting scores. The experimental results show that proposed method can effectively improve the performance of quality estimation model. Compared with the baseline system, the proposed method improves the Pearson correlation coefficient by up to 6.68 percentage points.

Different from the previous studies with only text, this paper focuses on multimodal data (text and audio) to perform emotion recognition. To simultaneously address the characteristics of multimodal data, we propose a novel joint learning framework, which allows auxiliary task (multimodal sentiment classification) to help the main task (multimodal emotion classification). Specifically, private neural layers are designed for text and audio modalities from the main task to learn the uni-modal independent dynamics. Secondly, with the shared neural layers from auxiliary task, we obtain the uni-modal representations of the auxiliary task and the auxiliary representations of the main task. The uni-modal independent dynamics is combined with the auxiliary representations for each modality to acquire the uni-modal representations of the main task. Finally, in order to capture multimodal interactive dynamics, we fuse the text and audio modalities’ representations for the main and auxiliary tasks separately to obtain the final multimodal emotion and sentiment representations with the self attention mechanism. Empirical results demonstrate the effectiveness of our approach to multimodal emotion classification task as well as the sentiment classification task.