Happiness prediction based on large-scale online data and machine learning models is an emerging research topic that underpins a range of issues, from personal growth to social stability. Many advanced machine learning (ML) models with explanations are used for happiness online assessment while maintaining high accuracy of results. However, expert feedback and sociological theory may be absent from these models, which limits the association between prediction results and the right reasons for why they occurred. Sociological studies have shown that primary and secondary relations are inherent in happiness factors, which can be used as domain knowledge to guide model training. Inspired by such insights, this article attempts to provide new insights into the explanation consistency from an empirical study perspective. Then this article studies how to represent and introduce domain knowledge constraints to make ML models more trustworthy. We achieve this by 1) proving that multiple prediction models with additive factor attributions will have the desirable property of primary and secondary relations consistency; and 2) showing that factor relations with quantity can be represented as an importance distribution for encoding domain knowledge. Factor explanation difference is penalized by the Wasserstein distance among prediction models. Experimental results using two online datasets show that domain knowledge of stable factor relations exists. Using this knowledge not only improves happiness prediction accuracy but also reveals more significant happiness factors for assisting decisions.

Predicting stock price movements is a high-stakes task that demands explainability for human decision-makers. A key shortcoming in current methods is treating sub-predictions independently, without learning from accumulated experiences. We propose a novel triplet network for contrastive learning to enhance the explainability of stock movement prediction by considering instances of "integrated textual information and quantitative indicators". We refer to the target past-l-day tweet-price time series as the "anchor instance". Each anchor instance is paired with a "positive instance" characterized by highly correlated return trends yet significant differences across the entire feature space, and a "negative instance" that exhibits similar return trends along with high proximity in the feature space. The model is designed with the objective of (1) minimizing the cross entropy loss between input logits and target, (2) minimizing the distance between the anchor instances and positive instances, and (3) maximizing the distance between the anchor instances and negative instances. Our framework's effectiveness is demonstrated through extensive testing, showing superior performance on stock prediction benchmarks.

The prediction of stock price movements is challenging due to the inherently dynamic and complex characteristics of financial markets. A current research gap is the lack of exploration into the complex interrelationships inherent in stock price dynamics, often analyzing predictions in isolation with an implicit presumption that solely the historical data of a given stock influences its future trend. However, stock prices are impacted by a diverse array of driving factors that extend beyond the traditionally examined historical prices, encompassing influences such as inter-stock correlations. In this paper, we present a predictive approach using a dynamic dual-graph neural network. The network combines textual data and quantitative metrics to capture multiple dynamic relationships. Specifically, We have developed a price relationship graph (PRG) and a semantic relationship graph (SRG), which are later integrated using a graph attention neural network. The effectiveness of our neural architecture is validated through extensive testing on two benchmark datasets for stock movement prediction, illustrating its superior performance compared to other graph-based networks for stock market prediction.

Traditionally, individual financial risk tolerance information is gathered via questionnaires or similar structured psychometric tools. Our abundant digital footprint, as an unstructured alternative, is less investigated. Leveraging such information can potentially support large-scale and cost-efficient financial services. Therefore, I explore the possibility of building a computational model that distills risk tolerance information from user texts in this study, and discuss the design principles discovered from empirical results and their implications. Specifically, a new quaternary classification task is defined for text mining-based risk profiling. Experiments show that pre-trained large language models set a baseline micro-F1 of circa 0.34. Using a convolutional neural network (CNN), the reported system achieves a micro-F1 of circa 0.51, which significantly outperforms the baselines, and is a circa 4% further improvement over the standard CNN configurations (micro-F1 of circa 0.47). Textual feature richness and supervised learning are found to be the key contributors to model performances, while other machine learning strategies suggested by previous research (data augmentation and multi-tasking) are less effective. The findings confirm user texts to be a useful risk profiling resource and provide several insights on this task.

Neural network-based methods, especially deep learning, have been a burgeoning area in AI research and have been successful in tackling the expanding data volume as we move into a digital age. Today, the neural network-based methods are not only used for low-level cognitive tasks, such as recognizing objects and spotting keywords, but they have also been deployed in various industrial information systems to assist high-level decision-making. In natural language processing, there have been two milestones for the past decade: one is word2vec [1], a group of neural models that learn word embeddings (vector representations of words) from large datasets; and one is the most recent GPT-based models [2], which combine reinforcement learning with a generative transformer in order to enable multi-round end-to-end conversations. While producing highly accurate predictions on datasets and generating human-like utterances, those neural network-based artifacts provide little understanding of the internal features and representations of the data. Many problems and concerns subsequently emerge from this black-box issue. Because some of the problems and concerns are also relevant in the context of sentiment analysis.

The goal of building intelligent dialogue systems has largely been separately pursued under two paradigms: task-oriented dialogue (TOD) systems, which perform task-specific functions, and open-domain dialogue (ODD) systems, which focus on non-goal-oriented chitchat. The two dialogue modes can potentially be intertwined together seamlessly in the same conversation, as easily done by a friendly human assistant. Such ability is desirable in conversational agents, as the integration makes them more accessible and useful. Our paper addresses this problem of fusing TODs and ODDs in multi-turn dialogues. Based on the popular TOD dataset MultiWOZ, we build a new dataset FusedChat, by rewriting the existing TOD turns and adding new ODD turns. This procedure constructs conversation sessions containing exchanges from both dialogue modes. It features inter-mode contextual dependency, i.e., the dialogue turns from the two modes depend on each other. Rich dependency patterns such as co-reference and ellipsis are included. The new dataset, with 60k new human-written ODD turns and 5k re-written TOD turns, offers a benchmark to test a dialogue model's ability to perform inter-mode conversations. This is a more challenging task since the model has to determine the appropriate dialogue mode and generate the response based on the inter-mode context. However, such models would better mimic human-level conversation capabilities. We evaluate two baseline models on this task, including the classification-based two-stage models and the two-in-one fused models. We publicly release FusedChat and the baselines to propel future work on inter-mode dialogue systems.

In recent years, AI research has demonstrated enormous potential for the benefit of humanity and society. While often better than its human counterparts in classification and pattern recognition tasks, however, AI still struggles with complex tasks that require commonsense reasoning such as natural language understanding. In this context, the key limitations of current AI models are: dependency, reproducibility, trustworthiness, interpretability, and explainability. In this work, we propose a commonsense-based neurosymbolic framework that aims to overcome these issues in the context of sentiment analysis. In particular, we employ unsupervised and reproducible subsymbolic techniques such as auto-regressive language models and kernel methods to build trustworthy symbolic representations that convert natural language to a sort of protolanguage and, hence, extract polarity from text in a completely interpretable and explainable manner.

Deep learning has unlocked new paths towards the emulation of the peculiarly-human capability of learning from examples. While this kind of bottom-up learning works well for tasks such as image classification or object detection, it is not as effective when it comes to natural language processing. Communication is much more than learning a sequence of letters and words: it requires a basic understanding of the world and social norms, cultural awareness, commonsense knowledge, etc.; all things that we mostly learn in a top-down manner. In this work, we integrate top-down and bottom-up learning via an ensemble of symbolic and subsymbolic AI tools, which we apply to the interesting problem of polarity detection from text. In particular, we integrate logical reasoning within deep learning architectures to build a new version of SenticNet, a commonsense knowledge base for sentiment analysis.

The recent dominance of machine learning-based natural language processing methods has fostered the culture of overemphasizing model accuracies rather than studying the reasons behind their errors.Interpretability, however, is a critical requirement for many downstream AI and NLP applications, e.g., in finance, healthcare, and autonomous driving.This study, instead of proposing any "new model", investigates the error patterns of some widely acknowledged sentiment analysis methods in the finance domain.We discover that (1) those methods belonging to the same clusters are prone to similar error patterns, and (2) there are six types of linguistic features that are pervasive in the common errors.These findings provide important clues and practical considerations for improving sentiment analysis models for financial applications.

This work tackles the seldom discussed task of predicting chaotic time series generated by dynamic systems with evolving parameters. Representative chaotic time series produced by different system dimensions are introduced with a critical parameter linearly depending on time. The evolving character of systems are qualitatively studied by phase portraits. We assess the predictability of different fuzzy neural network (FNN) architectures on several evolving chaotic time series. Experiments illustrate that FNN models can generally better approximate evolving chaotic systems comparing to the autoregression method as a benchmark. The main contribution of our work is that we found out certain FNN types, e.g., NEFCON and DENFIS, are more robust to changing system parameters. In spite of the performance, some FNN models are more vulnerable and incline to be destabilized by high order chaotic systems. This work also casts light on composing FNN structures to capture evolving characters of chaotic time series in the future.

In this paper, we propose to use contexts of superpixels as a prior to improve semantic segmentation by the CRF framework. A graphical model is constructed on over-segmented images. Our main contribution is to take the concept of “superpixel embedding” into consideration, which is formalized as a potential item for optimizing the energy of the whole graph. We also introduce two ways of calculating this embedding potential. Experiments on several popular datasets, e.g., MRSC-21 and PASCAL VOC, illustrate that our approach enhances the performance of a previously proposed segmentation model without embedding. The accuracy results are comparable to some fully supervised methods.