The emergence of Large Language Models (LLMs) provides a new solution to text generation tasks that involve high complexity, such as text style transfer (TST) tasks. However, previous studies have not fully explored the TST capabilities of different LLMs, and have faced issues with a lack of uniform standards in the human evaluation stage. This makes the results of human evaluation difficult to reproduce and less credible. To address this, this paper designs a prompt template to guide the cutting-edge LLMs to perform effective text style transfer and carries out an in-depth comparative analysis of various small-scale language models. In the stage of human evaluation, this paper eschews the conventional rating system, opting instead for a comparative human assessment methodology, which we refer to as duel-ranking. This method determines the relative ranking of models through mutual comparison, serving as an alternative to direct scoring. Detailed evaluation instructions are provided herein, to enhance the reproducibility of this method and ensure consistency throughout the evaluation process. This manual evaluation process reveals that GPT-3.5 and GPT-4 exhibit excellent performance in the TST tasks.

Abstract Multimodal sentiment analysis is a downstream branch task of sentiment analysis with high attention at present. Previous work in multimodal sentiment analysis have focused on the representation and fusion of modalities, capturing the underlying semantic relationships between modalities by considering contextual information. While this approach is feasible for simple contextual comments, more complex comments require the integration of external knowledge to obtain more accurate sentiment information. However, incorporating external knowledge into sentiment analysis to enhance information complementarity has not been thoroughly investigated. To address this, we propose a multichannel cross-modal feedback interaction model that incorporates the knowledge graph into multimodal sentiment analysis. Our proposed model consists of two main components: the cross-modal feedback recurrent interaction module and the external knowledge module for capturing latent information. The cross-modal interaction employs a self-feedback mechanism during network training, extracting feature representations of each modality and using these representations to mask sensory inputs, allowing the model to perform feedback-based feature masking. The external knowledge graph captures potential semantic information representations in the textual data through knowledge graph embedding. Finally, a global feature fusion module is employed for multichannel multimodal information integration. On two publicly available datasets, our method demonstrates good performance in terms of accuracy and F1 scores, compared to state-of-the-art models and several baselines.

Xianbing Zhou, Yang Yong, Xiaochao Fan, Ge Ren, Yunfeng Song, Yufeng Diao, Liang Yang, Hongfei Lin. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Pre-trained distributed word representations have been proven useful in various natural language processing (NLP) tasks. However, the effect of words’ geometric structure on word representations has not been carefully studied yet. The existing word representations methods underestimate the words whose distances are close in the Euclidean space, while overestimating words with a much greater distance. In this paper, we propose a word vector refinement model to correct the pre-trained word embedding, which brings the similarity of words in Euclidean space closer to word semantics by using manifold learning. This approach is theoretically founded in the metric recovery paradigm. Our word representations have been evaluated on a variety of lexical-level intrinsic tasks (semantic relatedness, semantic similarity) and the experimental results show that the proposed model outperforms several popular word representations approaches.

Heterographic pun plays a critical role in human writing and literature, which usually has a similar sounding or spelling structure. It is important and difficult research to recognize the heterographic pun because of the ambiguity. However, most existing methods for this task only focus on designing features with rule-based or machine learning methods. In this paper, we propose an end-to-end computational approach - Pronunciation Spelling Understanding Gated Attention (PSUGA) network. For pronunciation, we exploit the hierarchical attention model with phoneme embedding. While for spelling, we consider the character-level, word-level, tag-level, position-level and contextual-level embedding with attention model. To deal with the two parts, we present a gated attention mechanism to control the information integration. We have conducted extensive experiments on SemEval2017 task7 and Pun of the Day datasets. Experimental results show that our approach significantly outperforms state-of-the-art methods.