The recently developed matrix based Renyi's entropy enables measurement of information in data simply using the eigenspectrum of symmetric positive semi definite (PSD) matrices in reproducing kernel Hilbert space, without estimation of the underlying data distribution. This intriguing property makes the new information measurement widely adopted in multiple statistical inference and learning tasks. However, the computation of such quantity involves the trace operator on a PSD matrix $G$ to power $\alpha$(i.e., $tr(G^\alpha)$), with a normal complexity of nearly $O(n^3)$, which severely hampers its practical usage when the number of samples (i.e., $n$) is large. In this work, we present computationally efficient approximations to this new entropy functional that can reduce its complexity to even significantly less than $O(n^2)$. To this end, we first develop randomized approximations to $\tr(\G^\alpha)$ that transform the trace estimation into matrix-vector multiplications problem. We extend such strategy for arbitrary values of $\alpha$ (integer or non-integer). We then establish the connection between the matrix-based Renyi's entropy and PSD matrix approximation, which enables us to exploit both clustering and block low-rank structure of $\G$ to further reduce the computational cost. We theoretically provide approximation accuracy guarantees and illustrate the properties of different approximations. Large-scale experimental evaluations on both synthetic and real-world data corroborate our theoretical findings, showing promising speedup with negligible loss in accuracy.
Graph convolutional networks (GCNs) have been widely adopted for graph representation learning and achieved impressive performance. For larger graphs stored separately on different clients, distributed GCN training algorithms were proposed to improve efficiency and scalability. However, existing methods directly exchange node features between different clients, which results in data privacy leakage. Federated learning was incorporated in graph learning to tackle data privacy, while they suffer from severe performance drop due to non-iid data distribution. Besides, these approaches generally involve heavy communication and memory overhead during the training process. In light of these problems, we propose a Privacy-Preserving Subgraph sampling based distributed GCN training method (PPSGCN), which preserves data privacy and significantly cuts back on communication and memory overhead. Specifically, PPSGCN employs a star-topology client-server system. We firstly sample a local node subset in each client to form a global subgraph, which greatly reduces communication and memory costs. We then conduct local computation on each client with features or gradients of the sampled nodes. Finally, all clients securely communicate with the central server with homomorphic encryption to combine local results while preserving data privacy. Compared with federated graph learning methods, our PPSGCN model is trained on a global graph to avoid the negative impact of local data distribution. We prove that our PPSGCN algorithm would converge to a local optimum with probability 1. Experiment results on three prevalent benchmarks demonstrate that our algorithm significantly reduces communication and memory overhead while maintaining desirable performance. Further studies not only demonstrate the fast convergence of PPSGCN, but discuss the trade-off between communication and local computation cost as well.
Political stance detection has become an important task due to the increasingly polarized political ideologies. Most existing works focus on identifying perspectives in news articles or social media posts, while social entities, such as individuals and organizations, produce these texts and actually take stances. In this paper, we propose the novel task of entity stance prediction, which aims to predict entities' stances given their social and political context. Specifically, we retrieve facts from Wikipedia about social entities regarding contemporary U.S. politics. We then annotate social entities' stances towards political ideologies with the help of domain experts. After defining the task of entity stance prediction, we propose a graph-based solution, which constructs a heterogeneous information network from collected facts and adopts gated relational graph convolutional networks for representation learning. Our model is then trained with a combination of supervised, self-supervised and unsupervised loss functions, which are motivated by multiple social and political phenomenons. We conduct extensive experiments to compare our method with existing text and graph analysis baselines. Our model achieves highest stance detection accuracy and yields inspiring insights regarding social entity stances. We further conduct ablation study and parameter analysis to study the mechanism and effectiveness of our proposed approach.
Identifying political perspective in news media has become an important task due to the rapid growth of political commentary and the increasingly polarized ideologies. Previous approaches only focus on leveraging the semantic information and leaves out the rich social and political context that helps individuals understand political stances. In this paper, we propose a perspective detection method that incorporates external knowledge of real-world politics. Specifically, we construct a contemporary political knowledge graph with 1,071 entities and 10,703 triples. We then build a heterogeneous information network for each news document that jointly models article semantics and external knowledge in knowledge graphs. Finally, we apply gated relational graph convolutional networks and conduct political perspective detection as graph-level classification. Extensive experiments show that our method achieves the best performance and outperforms state-of-the-art methods by 5.49%. Numerous ablation studies further bear out the necessity of external knowledge and the effectiveness of our graph-based approach.
Zero-shot object detection (ZSD), the task that extends conventional detection models to detecting objects from unseen categories, has emerged as a new challenge in computer vision. Most existing approaches tackle the ZSD task with a strict mapping-transfer strategy, which may lead to suboptimal ZSD results: 1) the learning process of those models ignores the available unseen class information, and thus can be easily biased towards the seen categories; 2) the original visual feature space is not well-structured and lack of discriminative information. To address these issues, we develop a novel Semantics-Guided Contrastive Network for ZSD, named ContrastZSD, a detection framework that first brings contrastive learning mechanism into the realm of zero-shot detection. Particularly, ContrastZSD incorporates two semantics-guided contrastive learning subnets that contrast between region-category and region-region pairs respectively. The pairwise contrastive tasks take advantage of additional supervision signals derived from both ground truth label and pre-defined class similarity distribution. Under the guidance of those explicit semantic supervision, the model can learn more knowledge about unseen categories to avoid the bias problem to seen concepts, while optimizing the data structure of visual features to be more discriminative for better visual-semantic alignment. Extensive experiments are conducted on two popular benchmarks for ZSD, i.e., PASCAL VOC and MS COCO. Results show that our method outperforms the previous state-of-the-art on both ZSD and generalized ZSD tasks.
A predicate head is a verbal expression that plays a role as the structural center of a sentence. Identifying predicate heads is critical to understanding a sentence. It plays the leading role in organizing the relevant syntactic elements in a sentence, including subject elements, adverbial elements, etc. For some languages, such as English, word morphologies are valuable for identifying predicate heads. However, Chinese offers no morphological information to indicate words` grammatical roles. A Chinese sentence often contains several verbal expressions; identifying the expression that plays the role of the predicate head is not an easy task. Furthermore, Chinese sentences are inattentive to structure and provide no delimitation between words. Therefore, identifying Chinese predicate heads involves significant challenges. In Chinese information extraction, little work has been performed in predicate head recognition. No generally accepted evaluation dataset supports work in this important area. This paper presents the first attempt to develop an annotation guideline for Chinese predicate heads and their relevant syntactic elements. This annotation guideline emphasizes the role of the predicate as the structural center of a sentence. The design of relevant syntactic element annotation also follows this principle. Many considerations are proposed to achieve this goal, e.g., patterns of predicate heads, a flattened annotation structure, and a simpler syntactic unit type. Based on the proposed annotation guideline, more than 1,500 documents were manually annotated. The corpus will be available online for public access. With this guideline and annotated corpus, our goal is to broadly impact and advance the research in the area of Chinese information extraction and to provide the research community with a critical resource that has been lacking for a long time.
Entity alignment (EA) aims to discover the equivalent entities in different knowledge graphs (KGs). It is a pivotal step for integrating KGs to increase knowledge coverage and quality. Recent years have witnessed a rapid increase of EA frameworks. However, state-of-the-art solutions tend to rely on labeled data for model training. Additionally, they work under the closed-domain setting and cannot deal with entities that are unmatchable. To address these deficiencies, we offer an unsupervised framework that performs entity alignment in the open world. Specifically, we first mine useful features from the side information of KGs. Then, we devise an unmatchable entity prediction module to filter out unmatchable entities and produce preliminary alignment results. These preliminary results are regarded as the pseudo-labeled data and forwarded to the progressive learning framework to generate structural representations, which are integrated with the side information to provide a more comprehensive view for alignment. Finally, the progressive learning framework gradually improves the quality of structural embeddings and enhances the alignment performance by enriching the pseudo-labeled data with alignment results from the previous round. Our solution does not require labeled data and can effectively filter out unmatchable entities. Comprehensive experimental evaluations validate its superiority.
Recognizing named entities (NEs) is commonly conducted as a classification problem that predicts a class tag for an NE candidate in a sentence. In shallow structures, categorized features are weighted to support the prediction. Recent developments in neural networks have adopted deep structures that map categorized features into continuous representations. This approach unfolds a dense space saturated with high-order abstract semantic information, where the prediction is based on distributed feature representations. In this paper, the regression operation is introduced to locate NEs in a sentence. In this approach, a deep network is first designed to transform an input sentence into recurrent feature maps. Bounding boxes are generated from the feature maps, where a box is an abstract representation of an NE candidate. In addition to the class tag, each bounding box has two parameters denoting the start position and the length of an NE candidate. In the training process, the location offset between a bounding box and a true NE are learned to minimize the location loss. Based on this motivation, a multiobjective learning framework is designed to simultaneously locate entities and predict the class probability. By sharing parameters for locating and predicting, the framework can take full advantage of annotated data and enable more potent nonlinear function approximators to enhance model discriminability. Experiments demonstrate state-of-the-art performance for nested named entities\footnote{Our codes will be available at: \url{https://github.com/wuyuefei3/BR}}.
Textbook Question Answering (TQA) is a task that one should answer a diagram/non-diagram question given a large multi-modal context consisting of abundant essays and diagrams. We argue that the explainability of this task should place students as a key aspect to be considered. To address this issue, we devise a novel architecture towards span-level eXplanations of the TQA (XTQA) based on our proposed coarse-to-fine grained algorithm, which can provide not only the answers but also the span-level evidences to choose them for students. This algorithm first coarsely chooses top $M$ paragraphs relevant to questions using the TF-IDF method, and then chooses top $K$ evidence spans finely from all candidate spans within these paragraphs by computing the information gain of each span to questions. Experimental results shows that XTQA significantly improves the state-of-the-art performance compared with baselines. The source code is available at https://github.com/keep-smile-001/opentqa
Recognizing named entities (NEs) is commonly conducted as a classification problem that predicts a class tag for an NE candidate in a sentence. In shallow structures, categorized features are weighted to support the prediction. Recent developments in neural networks have adopted deep structures that map categorized features into continuous representations. This approach unfolds a dense space saturated with high-order abstract semantic information, where the prediction is based on distributed feature representations. In this paper, the regression operation is introduced to locate NEs in a sentence. In this approach, a deep network is first designed to transform an input sentence into recurrent feature maps. Bounding boxes are generated from the feature maps, where a box is an abstract representation of an NE candidate. In addition to the class tag, each bounding box has two parameters denoting the start position and the length of an NE candidate. In the training process, the location offset between a bounding box and a true NE are learned to minimize the location loss. Based on this motivation, a multiobjective learning framework is designed to simultaneously locate entities and predict the class probability. By sharing parameters for locating and predicting, the framework can take full advantage of annotated data and enable more potent nonlinear function approximators to enhance model discriminability. Experiments demonstrate state-of-the-art performance for nested named entities\footnote{Our codes will be available at: \url{https://github.com/wuyuefei3/BR}}.