Financial sentiment analysis (FSA) has garnered considerable attention for its potential to detect bullish and bearish sentiments that drive stock market fluctuations. Nonetheless, extracting salient sentiments from analyst reports encounters two main challenges. First, the highly specialized terms and expressions prevalent in these reports make it difficult for general Large Language Models (LLMs) to interpret financial expertise. Second, sentiment cues are implicit and dispersed across long-range dependencies, whereas existing LLM-based FSA methods relying on a single fine-tuning strategy lack fine-grained control during adaptation, thus leading to key information loss. To tackle these issues, we propose SEHLP, the first LLM that integrates summary information with a hybrid adaptation strategy that combines Low-rank Adaptation (LoRA) and dynamic Prefix Tuning to enhance FSA. Specifically, we employ prompt engineering on Qwen-2.5-14B to generate concise summaries that distill salient insights of each report, and construct FinLLaMA as SEHLP’s backbone through Supervised Fine-tuning (SFT) on extensive domain-specific instructions, enhancing financial knowledge comprehension. To inject summary information and enable fine-grained control during fine-tuning, we propose a hybrid adaptation strategy that concatenates LoRA-updated attention projections with dynamic summary-enhanced key-value prefixes, thereby fully utilizing sentiment cues in analyst reports and their summaries. Moreover, we construct a large-scale LCFR-Instruct corpus with 16,912 samples to address the lack of high-quality FSA instruction data. Comprehensive experiments on the LCFR-Instruct and FinTHUC-Instruct benchmark datasets indicate that SEHLP, with only 1.3B parameters, consistently surpasses competing LLMs, exhibiting ACC gains of 1.89% and 1.59% over the larger FinGPT-7B model on both datasets while maintaining superior efficiency. Our code is publicly accessible at https://github.com/lhz9999/SEHLP.

Robustness against word substitution attacks is crucial for text classifiers and fundamental to broader NLP robustness. Such attacks use semantically similar word substitutions. Existing certified defenses often compute loose outer bounds for the convex hull of word embeddings, including irrelevant words, degrading performance on clean and adversarial tests. Additionally, convex hull-based methods also struggle to emulate worst-case scenarios accurately. This letter proposes the Masked Adversarial Convex Combination (MACC) method, which models the solution space as a convex hull of word vectors and uses Variational Information Bottleneck theory to mask unnecessary words. We also introduce an empirical masking method based on the volume of the convex hull to enhance the performance. By reducing the number of words explored within the convex hull, MACC enables more precise mimicry of worst-case attacks. Experiments across models and datasets show MACC outperforms existing methods in clean and adversarial accuracy against word-level attacks.

Due to the rapid development of the Internet, the abundance of views and emotions surrounding news articles has become significant. Accurate prediction of these emotions plays a crucial role in social media applications, such as sentiment retrieval. In this paper, building upon previous research, we address the challenge of limited emotion tendencies and emotional words in social news by leveraging prompt knowledge and emotional comments posted under news, and develop an end-to-end architecture to predict news readers' emotion. Specifically, our approach introduces a prompt knowledge enhanced large language model as news encoder, which use prompt knowledge to emphasize parts of news which are likely to evoke reader's emotions. For encoding emotional and topic comments, we first propose Neural Emotion Topic Model(NETM), which can generate emotional topics of comments for emotion prediction. Then we use them to develop emotion-topic attention mechanism to capture emotional and topical information while encoding comments. By sharing emotion topic embedding, NETM can be co-trained with attention mechanism and large language models, further enhancing the performance of our architecture. Experiments on four social news datasets show that our proposed end-to-end framework outperforms the state-of-art methods.

Despite the impressive performance of neural language models in natural language processing, they remain vulnerable to adversarial attacks, particularly those based on synonym substitutions at the word level. Recent defense methods, however, often generate worst-case adversarial examples or upper bounds on propagation by independently perturbing each data sample. This approach limits the model’s ability to generalize to unseen data. To address this limitation, we propose a distributional robustness framework to defend against synonym-based word-level adversarial attacks. Our framework identifies the worst-case distribution within a known uncertainty set to craft adversarial examples and models the solution space as the convex hull of word vectors. This convex hull is sufficiently inclusive to cover all potential substitutions while excluding unnecessary ones. Furthermore, we provide an efficiently computable, data-dependent upper bound on the worst-case loss, ensuring that the worst-case performance of the output from our principled adversarial training procedure does not exceed this bound. Additionally, we certify the generalizability of the robustness level to the training set. Experiments on benchmark datasets demonstrate that our framework achieves comparable adversarial robustness to other defense methods under synonym-based word-level attacks.

Social emotion prediction aims to predict readers’ emotion, for example, emotion distributions evoked by documents (e.g., news articles). It makes a significant contribution to social media applications, such as opinion summary, election prediction, and emotions investigation of society. While recent studies have focused on encoding consecutive word sequences in documents using neural network models and leveraging topical information, it is essential to acknowledge the influence of documents sharing similar topics or being related to similar events on evoking readers’ emotions. The interactions among documents can significantly impact social emotion prediction. In this paper, we propose a novel approach to model the interactions among documents by constructing a heterogeneous graph. This graph captures the interaction among documents based on global word co-occurrence patterns in a corpus and the emotional scores of words obtained from emotion lexicons. Additionally, we develop heterogeneous graph convolution attention network (HGCA) to embed the heterogeneous graph. This network effectively captures the importance of different neighboring nodes and different node types, enabling comprehensive emotion prediction. Furthermore, we develop Taylor series expansion-based Transformer (Tayformer) to derive initialized node representations that can be co-trained with our graph network while having low memory complexity. Experimental results on four benchmark datasets show the effectiveness of our method.

General graph neural networks (GNNs) implement convolution operations on graphs based on polynomial spectral filters. Existing filters with high-order polynomial approximations can detect more structural information when reaching high-order neighborhoods but produce indistinguishable representations of nodes, which indicates their inefficiency of processing information in high-order neighborhoods, resulting in performance degradation. In this article, we theoretically identify the feasibility of avoiding this problem and attribute it to overfitting polynomial coefficients. To cope with it, the coefficients are restricted in two steps, dimensionality reduction of the coefficients' domain and sequential assignment of the forgetting factor. We transform the optimization of coefficients to the tuning of a hyperparameter and propose a flexible spectral-domain graph filter, which significantly reduces the memory demand and the adverse impacts on message transmission under large receptive fields. Utilizing our filter, the performance of GNNs is improved significantly in large receptive fields and the receptive fields of GNNs are multiplied as well. Meanwhile, the superiority of applying a high-order approximation is verified across various datasets, notably in strongly hyperbolic datasets. Codes are publicly available at: https://github.com/cengzeyuan/TNNLS-FFKSF.

Financial reports serve as crucial resources for investors and researchers, providing analysts' assessments of stocks that play a vital role in stock market applications. However, detecting analysts' opinions and sentiments in financial reports is challenging. First, the formal and professional language used in these reports makes it difficult for previous methods to comprehend domain-specific knowledge. Second, financial reports often adopt lengthy and elaborate expressions to convey rich semantics, which exposes the existing methods to contextual information loss, especially on long-term dependencies. To address these problems, we propose a summary-enhanced hierarchical framework (SEHF), which leverages summary information to enhance financial report sentiment analysis. Our framework incorporates financial bidirectional and auto-regressive transformer (FinBART), equipped with extended position encoding to summarize lengthy report articles and capture long-range interactions. To mitigate information loss, we initially divide each report into segments and then propose the hierarchical analyst sentiment representation network (ASRN), which utilizes financial bidirectional encoder representation from transformer (FinBERT), bidirectional long short-term memory (BiLSTM)-Attention, and dendrite (DD) network to fuse information in the generated summary and report segments. Notably, FinBART and FinBERT are pretrained on large-scale financial corpora to effectively understand professional expressions. Furthermore, we construct a new dataset large-scale Chinese financial report (LCFR) for the lack of supervised datasets. Experimental results on LCFR and a benchmark dataset show that SEHF significantly outperforms state-of-the-art (SOTA) baselines, and the ablation study highlights the effectiveness of aggregating sentiment information in the summary and report segments.

With the rapid development of the Internet, readers tend to share their views and emotions about news events. Predicting these emotions provides a vital role in social media applications (e.g., sentiment retrieval, opinion summary, and election prediction). However, news articles usually consist of objective texts that lack emotion words, making emotion prediction challenging. From prior studies, we know that comments that come directly from readers are full of emotions. Therefore, in this article, we propose a deep learning framework that first merges article and comment information to predict readers' emotions. At the same time, in the prediction process, we design a pseudo comment representation for unpublished news articles by the comments of published news. In addition, a better model is required to encode articles that contain implicit emotions. To solve this problem, we propose a block emotion attention network (BEAN) to encode news articles better. It includes an emotion attention mechanism and a hierarchical structure to capture emotion words and generate structural information during encoding. Experiments performed on three public datasets show that BEAN achieves the state-of-the-art average Pearson (AP) and accuracy (Acc@1). Moreover, results on four self-collected datasets show that both the introduction of emotional comments and BEAN in our framework improve the ability to predict readers' emotions.

Nowadays, multivariate time series data is increasingly collected in many large-scale application systems, which often has periodic, repetitive patterns that can be affected by advertisements, workdays, holidays, and some user behavior activities. However, existing density and distance-based anomaly detection approaches suffer from detecting anomalies related to periodicity and seasonality. To address this problem, we propose a generic and scalable adaptation framework (GSPAD) for unsupervised anomaly detection in time series with periodic patterns. Our framework mainly consists of a time series predictor and an anomaly detector. Therefore, we present a Convolutional Attention-skip Network (CASNet) as a predictor responsible for predicting both short- and long-term patterns. These two types of patterns are modeled by the CASNet combining the Convolutional Neural Network (CNN) and the Dual Branch Attention-skip Network. Moreover, the proposed anomaly detector can deduce the anomaly according to the severity of the deviations between the actual and predicted values. Compared with other related researches on public datasets, GSPAD shows better performance with an average F-score over 0.76.

Financial news summarization (FNS) has been an attractive research problem in recent years, which aims to generate a shorter highlight of the news article while preserving key factual aspects, emotions, and opinions, providing significant assistance in stock trading and investment decision-making. However, FNS faces two challenges compared to the common domain. Firstly, financial news involves professional qualitative and quantitative information and salient content always scatters across long-range interactions. Secondly, financial news contains latent causal relationships, where historical information in the early generated sequence can significantly affect the subsequent decoding process. To address these difficulties, we propose an enhanced Seq2Seq model named TLGA, where the hierarchical Transformer-BiLSTM encoder can capture long-range interactions and sequential semantics while the Graph Attention-based decoder can fully utilize the historical information of decoded tokens and capture key causal relations. Moreover, we propose history-enhanced attention to concentrate on salient input content based on history semantics, guiding our decoder to generate the summary around the corresponding contents. It is also the first attempt to reuse history information of previously generated summary sequences in FNS using the idea of the Graph Attention Mechanism. Additionally, we construct the LCFNS dataset with 430,820 news-summary pairs for the lack of large-scale high-quality datasets in FNS. Experimental results on two financial datasets and two benchmark datasets indicate that our model outperforms other baselines.

BACKGROUND:The zone adjacent to a transcription start site (TSS), namely, the promoter, is primarily involved in the process of DNA transcription initiation and regulation. As a result, proper promoter identification is critical for further understanding the mechanism of the networks controlling genomic regulation. A number of methodologies for the identification of promoters have been proposed. Nonetheless, due to the great heterogeneity existing in promoters, the results of these procedures are still unsatisfactory. In order to establish additional discriminative characteristics and properly recognize promoters, we developed the hybrid model for promoter identification (HMPI), a hybrid deep learning model that can characterize both the native sequences of promoters and the morphological outline of promoters at the same time. We developed the HMPI to combine a method called the PSFN (promoter sequence features network), which characterizes native promoter sequences and deduces sequence features, with a technique referred to as the DSPN (deep structural profiles network), which is specially structured to model the promoters in terms of their structural profile and to deduce their structural attributes.RESULTS:The HMPI was applied to human, plant and Escherichia coli K-12 strain datasets, and the findings showed that the HMPI was successful at extracting the features of the promoter while greatly enhancing the promoter identification performance. In addition, after the improvements of synthetic sampling, transfer learning and label smoothing regularization, the improved HMPI models achieved good results in identifying subtypes of promoters on prokaryotic promoter datasets.CONCLUSIONS:The results showed that the HMPI was successful at extracting the features of promoters while greatly enhancing the performance of identifying promoters on both eukaryotic and prokaryotic datasets, and the improved HMPI models are good at identifying subtypes of promoters on prokaryotic promoter datasets. The HMPI is additionally adaptable to different biological functional sequences, allowing for the addition of new features or models.

Native speaker detection has always focused on speech data. However, the writing style of native speakers is also different from that of non-native speakers. Therefore, for the first time, we performed native speaker detection on text data so that non-native speakers can better learn the writing style of native speakers. Native speaker writing detection is relatively difficult due to the long sequences and complex semantics in the writings. Therefore, we use BERT-based methods. However, BERT suffers from the exponentially increasing computational complexity because of the self-attention mechanism, which limits the length of text input. Consequently, in this paper, we present a hierarchical BERT model to solve this problem. Our model first cuts the long text into segments and obtains segment representation vectors from BERT. Then, we extract the temporal and interactional information between segments to form a text-level representation vector for writing detection. We conducted experiments on a self-made native speaker writing detection dataset. The results demonstrate that our model can accurately recognize native speakers' writing. In addition, we have successfully used it in various long text classification tasks and achieved improvement over the baseline models. We also show the importance of both temporal and interaction information for text-level representation.

In the era of data technology, data growth is occurring at an unprecedented scale. Business data and information are among the most valuable assets. Massive data analysis now drives nearly every aspect of society and can facilitate informed decision-making by businesses. Fully automated data flow detection of anomalies plays a crucial role in maintaining data service stability and preventing malicious attacks. This paper presents an extensible and generic real-time monitoring system framework (EGRTMS) for large-scale time-series data. EGRTMS employs a prediction module and an anomaly detection module within an anomaly filtering layer for the accurate identification of anomalies. Moreover, the alarm module and anomaly handling module within an anomaly trace processing layer enables the system to respond swiftly to the detected threats. Our solution does not rely on the labelling of anomalies; instead, a predictor module with a deep learning attention-based mechanism learns the normal behaviour of the data, and an anomaly handling module determines the dynamic alarm-threshold by utilizing a sliding window. The results of this study demonstrate that our framework significantly outperforms other anomaly detection systems on most real and synthetic datasets.

With the rapid development of social media, sentiment classification of financial microblogs has been a research hotpot. Online financial microblogs contain a wealth of information on market and usually express emotions or opinions. However, microblogs contain so much noise that it is difficult to make sentiment classification correctly. In this paper, we propose a two-stage model to improve the performance of sentiment classification through automatic text summarization techniques. The automatic text summarization model built on encoder-decoder model as well as attention mechanism can compress the original text and obtain the informative words, which can filter the noise in microblogs and capture significant semantics. Then we build the sentiment classifier based on hierarchical attention network. To make full use of sentiment semantics, we fuse every original blog and its generated summary to train our classifier. We establish two datasets based on Sina Weibo, one is for training the summarization model, another is for sentiment classification. Experimental results indicate that our model outperforms the other baselines on real datasets.

The promoter is a region located near the transcription start site (TSS) and responsible for the initiation and regulation of DNA transcription. Hence, accurate identification of promoters is essential for further building and understanding the mechanism of genetic regulatory networks. Numerous approaches for eukaryotic promoter identification were proposed. Nevertheless, the performances of these approaches are still unsatisfactory due to the variety nature of promoters. To extract more discriminative features and accurately identify eukaryotic promoters, here, we develop an effective hybrid deep learning model HDLMepi, which is able to characterize the original promoter sequences and the structural profiles of promoters simultaneously. We integrate the method we name PromoterClCce which characterizes the original promoter sequences and extracts sequence features, with an approach DSPN, which we design to model the structural profile of promoters and extract structure features, in HDLMepi for precisely eukaryotic promoter identification. We apply HDLMepi on both human and plants datasets and the experimental results demonstrate it is effective in promoter features extraction and can improve the performance of promoter identification significantly. HDLMepi is also open to add new features or new models and can be applied to other biology functional sequences.