Emotion analysis is currently a popular research direction in the field of natural language processing. However, existing research focuses primarily on tasks such as emotion classification, emotion extraction, and emotion cause analysis, while there are few investigations into the relationship between emotions and their impacts. To address these limitations, this paper introduces the emotion-driven behavior extraction (EDBE) task, which addresses these limitations by separately extracting emotions and behaviors to filter emotion-driven behaviors described in text. EDBE comprises three sub-tasks: emotion extraction, behavior extraction, and emotion–behavior pair filtering. To facilitate research in this domain, we have created a new dataset, which is accessible to the research community. To address the EDBE task, we propose a pipeline approach that incorporates the causal relationship between emotions and driven behaviors. Additionally, we adopt the prompt paradigm to improve the model’s representation of cause-and-effect relationships. In comparison to state-of-the-art methods, our approach demonstrates notable improvements, achieving a 1.32% improvement at the clause level and a 1.55% improvement at the span level on our newly curated dataset in terms of the $F_1$ score, which is a commonly used metric to measure the performance of models. These results underscore the effectiveness and superiority of our approach in relation to existing methods.

Background: Prompt recognition of patients predisposed to acute kidney injury (AKI) within 72 hours of intensive care unit (ICU) admission holds significant clinical importance as it can considerably lower mortality rates. However, existing AKI prediction models often require complex data collection yet yield only moderate performance. This study aims to develop a straightforward and efficient AKI prediction model, providing ICU physicians with a powerful tool to expedite the detection of AKI patients. Methods: This study proposed a novel generative adversarial imputation networks-least absolute shrinkage and selection operator-extreme gradient boosting (Gain-Lasso-XGBoost) framework and developed an AKI prediction model on the basis of the medical information mart for intensive care (MIMIC-III) database. All the steps, including data preprocessing, feature selection, development, and optimization of prediction models, are organically integrated into the framework which has strong scalability. To compare the performance of our model with current models, we conducted a systematic review to collect all studies on the basis of the MIMIC-III database with similar objectives. Results: From 15 demographic and clinical variables, 8 features and 5 features were identified as the optimal group of features and processed into the model development. The model optimization further improved the performance of our proposed framework, and the area under curve (AUC) results with 8 and 5 feature vectors achieved 0.849 and 0.830, respectively. Compared with other studies, our method extracted only 8 or 5 feature vectors and obtained superior performance, with an average AUC 1.9% higher than the state-of-the-art approaches in the same type. Conclusions: Our study suggested that the onset of AKI be effectively and quickly predicted using simplified features, and not just for more specific patient groups. It may help clinicians accurately identify patients at risk of AKI after ICU admission and provide timely monitoring and treatment.

In this paper, we present a novel self-supervised framework for Sentiment Cue Extraction (SCE) aimed at enhancing the interpretability of text sentiment analysis models. Our approach leverages self-supervised learning to identify and highlight key textual elements that significantly influence sentiment classification decisions. Central to our framework is the development of an innovative Mask Sequence Interpretation Score (MSIS), a bespoke metric designed to assess the relevance and coherence of identified sentiment cues within binary text classification tasks. By employing Monte Carlo Sampling techniques optimized for computational efficiency, our framework demonstrates exceptional effectiveness in processing large-scale text data across diverse datasets, including English and Chinese, thus proving its versatility and scalability. The effectiveness of our approach is validated through extensive experiments on several benchmark datasets, including SST-2, IMDb, Yelp, and ChnSentiCorp. The results indicate a substantial improvement in the interpretability of the sentiment analysis models without compromising their predictive accuracy. Furthermore, our method stands out for its global interpretability, offering an efficient solution for analyzing new data compared to traditional techniques focused on local explanations.

With the rapid development of Internet and mobile devices, a vast number of short texts are produced by users, which also post great challenges to topic modeling because of the severe sparsity in context. The traditional topic model cannot do well in short text because of lacking word co-occurrence patterns. An effective approach bi-term topic model(BTM) has been proposed which models the word co-occurrence at the whole corpus directly and performs better than conventional topic models. However, BTM only consider the frequency of bi-term simply and ignore the latent semantic information between bi-terms which cause the words with similar semantic having a great risk of being grouped under different topic. In this paper, we propose a latent semantic augmented bi-term topic model(LS-BTM) which incorporates semantic information as prior knowledge to infer the topic more reasonable. The experimental result shows that our model gets better result than other short text topic models over real-world dataset.

According to the characteristics of sparseness, poor focus and lack of semantic information in short text, the existing studies mainly improved the topic representation from two aspects: Adjusting the structure of the topic model to increase word co-occurrence and incorporating word embedding to enrich the semantic information. In this paper, we review the existing topic representation methods for short text from these two aspects, select DMM, BTM and LF-DMM as targets and compare the quality of topic representations from them. Considering the traditional topic model does not incorporate word embedding, and single short text often lacks context information due to length limitation, we try to combine it with word embedding, we use two real-world datasets to evaluate the quality of topic representation in the classification task: Although the LF-DMM model incorporates word embedding, it performs poorly on short text, and the performance of DMM and BTM integrated with word embedding improve greatly.

Deep learning has been demonstrated to be very effective in many difficult tasks of computer vision(CV) and natural language processing(NLP). But this usually depends on a large number of available training samples to learn, and powerful computing resources (like GPU). In the tasks of computer vision, multiple deep layers of the model are initialized with weights pre-trained on the large-scale datasets like ImageNet, which improves the training speed of the network but does not reduce the performance of the network. Inspired by this, in this paper, in order to demonstrate that transfer is also effective in natural language processing(NLP), we transfer the encoder parts of two models trained under different tasks of NLP to sentiment classification. In the result, we can find that even if we use a simple classifier, we still get good accuracy. And we propose a persuasive explanation for the phenomenon that why the similarity of the representation vectors obtained by different encoders under the same sentence is not high.