Temporal Knowledge Graphs (TKGs) are being widely explored to predict the future for they record multi-relational knowledge and the happening time of real-life facts. Existing works learn sequential patterns to infer the future from past facts in TKGs for predictions. Although achieving promising results, they are restricted by the sequential modeling in both efficiency and effectiveness. To resolve these limitations, we propose our efficient and effective non-sequential relational modeling (NoSeq). NoSeq works non-sequentially for temporal patterns where it transforms the happening time into time intervals. Time intervals are the period of time between happening time and prediction time which state the temporal distance clearly. Both time intervals and relations are represented using embeddings and merged non-sequentially into entity embeddings for future predictions. We evaluate NoSeq on four datasets from the perspective of effectiveness, efficiency, sensitivity, and the ability to transfer. Consistent better performances verify our idea.

Temporal Knowledge Graphs (TKGs) represent real-life facts using entities, relational types, and timestamps where relational types state the semantic scenario of facts. Current methods learn embeddings by merging facts of multiple types (e.g. sport and family) for predictions. Such embeddings associate well with relations of multiple types. However, the prediction needs only information of a single type, i.e. embeddings contain irrelevant relational types. In this paper, we explore whether embeddings with irrelevant information confuse predictions and propose RefE to improve embeddings by refining meanings. RefE enhances the ability for predicting links of a specific type while maintaining associations of multiple relational types. RefE consists of general learning and embedding refining modules. General learning embeds facts of multiple types to represent general meanings, and embedding refining emphasizes facts of the single type that matches the prediction. RefE uses both general and refined embeddings for predictions. Experimental results on four datasets verify the effectiveness of RefE.

Temporal Knowledge Graph (TKG) reasoning aims to predict facts at specific query timestamps using historical data. The main challenge is accurately modeling historical information for future queries. Previous approaches have primarily focused on global patterns and recent trends, but often overlooked query-relevant historical data. Although some recent methods have addressed this gap, they still struggle to capture implicit information and properly account for temporal dynamics. To overcome these limitations, we propose the Query-Aware Temporal Aggregation Network (QATAN), which adopts a novel “first evolution, then aggregation” strategy. QATAN effectively models the natural temporal evolution of facts and adaptively aggregates entity and relation embeddings based on query-relevant historical information through a query-aware temporal attention mechanism. Empirical evaluations demonstrate that QATAN outperforms state-of-the-art models on multiple benchmark datasets.

Knowledge graph completion (KGC) aims to infer missing information in incomplete knowledge graphs (KGs). Most previous works only consider the transductive scenario where entities are existing in KGs, which cannot work effectively for the inductive scenario containing emerging entities. Recently some graph neural network-based methods have been proposed for inductive KGC by aggregating neighborhood information to capture some uncertainty semantics from the neighboring auxiliary triples. But these methods ignore the more general relational semantics underlying all the known triples that can provide richer information to represent emerging entities so as to satisfy the inductive scenario. In this paper, we propose a novel model called CFAG, which utilizes two granularity levels of relational semantics in a coarse-grained aggregator (CG-AGG) and a fine-grained generative adversarial net (FG-GAN), for inductive KGC. The CG-AGG firstly generates entity representations with multiple semantics through a hypergraph neural network-based global aggregator and a graph neural network-based local aggregator, and the FG-GAN further enhances entity representations with specific semantics through conditional generative adversarial nets. Experimental results on benchmark datasets show that our model outperforms state-of-the-art models for inductive KGC.

Graph contrastive learning (GCL) has recently emerged as an effective learning paradigm to alleviate the reliance on labelling information for graph representation learning. The core of GCL is to maximise the mutual information between the representation of a node and its contextual representation (i.e., the corresponding instance with similar semantic information) summarised from the contextual scope (e.g., the whole graph or 1-hop neighbourhood). This scheme distils valuable self-supervision signals for GCL training. However, existing GCL methods still suffer from limitations, such as the incapacity or inconvenience in choosing a suitable contextual scope for different datasets and building biased contrastiveness. To address aforementioned problems, we present a simple self-supervised learning method termed Unifying Graph Contrastive Learning with Flexible Contextual Scopes (UGCL for short). Our algorithm builds flexible contextual representations with tunable contextual scopes by controlling the power of an adjacency matrix. Additionally, our method ensures contrastiveness is built within connected components to reduce the bias of contextual representations. Based on representations from both local and contextual scopes, UGCL optimises a very simple contrastive loss function for graph representation learning. Essentially, the architecture of UGCL can be considered as a general framework to unify existing GCL methods. We have conducted intensive experiments and achieved new state-of-the-art performance in six out of eight benchmark datasets compared with self-supervised graph representation learning baselines. Our code has been open sourced1.1https://github.com/zyzisastudyreallyhardguy/UGCL

With the vigorous development of dialogue system in natural language processing fields, dialogue summarization has attracted the attention of more scholars, which aims to extract brief introduction and hit points from dialogue information for readers. As the participation of multiple roles and the trans-formation of perspectives in dialogue, one of the most difficult problems, i.e. factual consistency, is raised in the generation of dialogue summaries. It means that the summaries are usually factual wrong although with high matching metric by many methods. Previous methods improve factual consistency between source and target by incorporating knowledge. However, the role of knowledge for dialogue summarization models lacks convincing evidence. In this paper, we propose a Fact-Augmentation (FA) Mechanism based Dialogue Summarization model, called FA-DS model for dialogue summarization. Our FA-DS integrates the fact graph extracted from the dialogue into the summaries generation process, and augment the gain of the factual information through the fact-aware score penalty item. Experiments on the large-scale dialogue dataset SAMSum demonstrate that our fact-augmentation mechanism can improve the quality and factual consistency of dialogue summarization.

Knowledge graph embedding is an effective way to represent knowledge graph, which greatly enhance the performances on knowledge graph completion tasks, e.g., entity or relation prediction. For knowledge graph embedding models, designing a powerful loss framework is crucial to the discrimination between correct and incorrect triplets. Margin-based ranking loss is a commonly used negative sampling framework to make a suitable margin between the scores of positive and negative triples. However, this loss can not ensure ideal low scores for the positive triplets and high scores for the negative triplets, which is not beneficial for knowledge completion tasks. In this paper, we present a double limit scoring loss to separately set upper bound for correct triplets and lower bound for incorrect triplets, which provides more effective and flexible optimization for knowledge graph embedding. Upon the presented loss framework, we present several knowledge graph embedding models including TransE-SS, TransH-SS, TransD-SS, ProjE-SS and ComplEx-SS. The experimental results on link prediction and triplet classification show that our proposed models have the significant improvement compared to state-of-the-art baselines.

Group activities have become an essential part of people’s daily life, which stimulates the requirement for intensive research on the group recommendation task, i.e., recommending items to a group of users. Most existing works focus on aggregating users’ interests within the group to learn group preference. These methods are faced with two problems. First, these methods only model the user preference inside a single group while ignoring the collaborative relations among users and items across different groups. Second, they assume that group preference is an aggregation of user interests, and factually a group may pursue some targets not derived from users’ interests. Thus they are insufficient to model the general group preferences which are independent of existing user interests. To address the above issues, we propose a novel dual channel Hypergraph Convolutional network for group Recommendation (HCR), which consists of member-level preference network and group-level preference network. In the member-level preference network, in order to capture cross-group collaborative connections among users and items, we devise a member-level hypergraph convolutional network to learn group members’ personal preferences. In the group-level preference network, the group’s general preference is captured by a group-level graph convolutional network based on group similarity. We evaluate our model on two real-world datasets and the experimental results show that the proposed model significantly and consistently outperforms state-of-the-art group recommendation techniques.

Image captioning is one of the most challenging tasks in AI, which aims to automatically generate textual sentences for an image. Recent methods for image captioning follow encoder-decoder framework that transforms the sequence of salient regions in an image into natural language descriptions. However, these models usually lack the comprehensive undemanding of the contextual interactions reflected on various visual relationships between objects. In this paper, we explore explicit and implicit visual relationships to enrich region-level representations for image captioning. Explicitly, we build semantic graph over object pairs and exploit gated graph convolutional networks (Gated GCN) to selectively aggregate local neighbors’ information. Implicitly, we draw global interactions among the detected objects through region-based bidirectional encoder representations from transformers (Region BERT) without extra relational annotations. To evaluate the effectiveness and superiority of our proposed method, we conduct extensive experiments on Microsoft COCO bench-mark and achieve remarkable improvements compared with strong baselines.

Sequential recommendation that aims to predict user preference with historical user interactions becomes one of the most popular tasks in the recommendation area. The existing methods concentrated on user's sequential features among exposed items have achieved good performance. However, they only rely on single item prediction optimization to learn data representation, which ignores the association between context data and sequence data. In this paper, we propose a novel self-supervised learning based sequential recommendation network (SSLRN), which contrastively learns data correlation to promote data representation of users and items. We design two auxiliary contrastive learning tasks to regularize user and item representation based on mutual information maximization (MIM). In particular, the item contrastive learning captures sequential contrast feature with sequence-item MIM, and the user contrastive learning regularizes user latent representation with user-item MIM. We evaluate our model on five real-world datasets and the experimental results show that the proposed framework significantly and consistently outperforms state-of-the-art sequential recommendation techniques.

Image captioning is a challenging task that combines computer vision and natural language processing for generating a textual description of the content within an image. Recently, Transformer-based encoder-decoder architectures have shown great success in image captioning, where multi-head attention mechanism is utilized to capture the contextual interactions between object regions. However, such methods regard region features as a bag of tokens without considering the directional relationships between them, making it hard to understand the relative position between objects in the image and generate correct captions effectively. In this paper, we propose a novel Direction Relation Transformer to improve the orientation perception between visual features by incorporating the relative direction embedding into multi-head attention, termed DRT. We first generate the relative direction matrix according to the positional information of the object regions, and then explore three forms of direction-aware multi-head attention to integrate the direction embedding into Transformer architecture. We conduct experiments on challenging Microsoft COCO image captioning benchmark. The quantitative and qualitative results demonstrate that, by integrating the relative directional relation, our proposed approach achieves significant improvements over all evaluation metrics compared with baseline model, e.g., DRT improves task-specific metric CIDEr score from 129.7% to 133.2% on the offline ''Karpathy'' test split.

Recurrent neural network(RNN) has achieved remarkable performances in complex reasoning on knowledge bases, which usually takes as inputs vector embeddings of relations along a path between an entity pair. However, it is insufficient to extract local correlations of a path due to RNN is better at capturing global sequential information of a path. In this paper, we take full advantages of convolutional neural network that can effectively extract local features, and propose a convolutional-based RNN architecture denoted as C-RNN to perform reasoning. C-RNN first utilizes CNN to extract local high-level correlation features of a path, and then feeds the correlation features into recurrent neural network to model the path representation. Our C-RNN architecture is adaptable to obtain not only local features but also global sequential features of a path. Based on C-RNN architecture, we devise two models, the unidirectional C-RNN and bidirectional C-RNN. We empirically evaluate them on a large-scale FreeBase+ClueWeb prediction task. Experimental results show that C-RNN models achieve state-of-the-art predictive performance.

Knowledge-graph-aware recommendation systems have increasingly attracted attention in both industry and academic recently. Many existing knowledge-aware recommendation methods have achieved better performance, which usually perform recommendation by reasoning on the paths between users and items in knowledge graphs. However, they ignore the users' personal clicked history sequences that can better reflect users' preferences within a period of time for recommendation. In this paper, we propose a knowledge-aware attentional reasoning network KARN that incorporates the users' clicked history sequences and path connectivity between users and items for recommendation. The proposed KARN not only develops an attention-based RNN to capture the user's history interests from the user's clicked history sequences, but also a hierarchical attentional neural network to reason on paths between users and items for inferring the potential user intents on items. Based on both user's history interest and potential intent, KARN can predict the clicking probability of the user with respective to a candidate item. We conduct experiment on Amazon review dataset, and the experimental results demonstrate the superiority and effectiveness of our proposed KARN model.

Joint extraction of entities and relations is an important task in natural language processing (NLP), which aims to capture all relational triplets from plain texts. This is a big challenge due to some of the triplets extracted from one sentence may have overlapping entities. Most existing methods perform entity recognition followed by relation detection between every possible entity pairs, which usually suffers from numerous redundant operations. In this paper, we propose a relation-specific attention network (RSAN) to handle the issue. Our RSAN utilizes relation-aware attention mechanism to construct specific sentence representations for each relation, and then performs sequence labeling to extract its corresponding head and tail entities. Experiments on two public datasets show that our model can effectively extract overlapping triplets and achieve state-of-the-art performance.

Recent neural sequence to sequence models have provided feasible solutions for abstractive summarization. However, such models are still hard to tackle long text dependency in the summarization task. A high-quality summarization system usually depends on strong encoder which can refine important information from long input texts so that the decoder can generate salient summaries from the encoder's memory. In this paper, we propose an aggregation mechanism based on the Transformer model to address the challenge of long text representation. Our model can review history information to make encoder hold more memory capacity. Empirically, we apply our aggregation mechanism to the Transformer model and experiment on CNN/DailyMail dataset to achieve higher quality summaries compared to several strong baseline models on the ROUGE metrics.

Fake News Detection on social network is still a challenging task that requires to integrate different types of information, e.g., source post, comments, and related users to verify the given news. However, previous solutions extract features from different aspects respectively, ignore the inherent relational and logical information among these features. In this paper, we propose IARNet, an Information Aggregating and Reasoning Network over heterogeneous graph for fake news detection, which exploits the interaction between information to aggregate multi-type information and grasps the inherent relationship simultaneously. Firstly, we construct a heterogeneous graph which takes source post, comments, and users as nodes and the interaction between them as edges. Then, a two-level attention mechanism is applied at the node level and type level. Specifically, the node-level attention aims to learn the importance between a node and its specific edge based neighbors, while the type-level attention aims to learn the importance of different types of edges. With the two-level attention mechanism, IARNet can aggregate multi-type information in a hierarchical manner and the information can reason over heterogeneous graph for the facticity of the news. Experimental result shows that our method outperforms the state-of-the-art competitors on real-world datasets with GloVe embeddings. We also demonstrate that using BERT representations further substantially boosts the performance. Our code is available at https://github.com/serryuer/IARNet.

With the rapid information explosion of news, making personalized news recommendation for users becomes an increasingly challenging problem. Many existing recommendation methods that regard the recommendation procedure as the static process, have achieved better recommendation performance. However, they usually fail with the dynamic diversity of news and user’s interests, or ignore the importance of sequential information of user’s clicking selection. In this paper, taking full advantages of convolution neural network (CNN), recurrent neural network (RNN) and attention mechanism, we propose a deep attention neural network DAN for news recommendation. Our DAN model presents to use attention-based parallel CNN for aggregating user’s interest features and attention-based RNN for capturing richer hidden sequential features of user’s clicks, and combines these features for new recommendation. We conduct experiment on real-world news data sets, and the experimental results demonstrate the superiority and effectiveness of our proposed DAN model.

Multilingual knowledge graphs constructed by entity alignment are the indispensable resources for numerous AI-related applications. Most existing entity alignment methods only use the triplet-based knowledge to find the aligned entities across multilingual knowledge graphs, they usually ignore the neighborhood subgraph knowledge of entities that implies more richer alignment information for aligning entities. In this paper, we incorporate neighborhood subgraph-level information of entities, and propose a neighborhood-aware attentional representation method NAEA for multilingual knowledge graphs. NAEA devises an attention mechanism to learn neighbor-level representation by aggregating neighbors' representations with a weighted combination. The attention mechanism enables entities not only capture different impacts of their neighbors on themselves, but also attend over their neighbors' feature representations with different importance. We evaluate our model on two real-world datasets DBP15K and DWY100K, and the experimental results show that the proposed model NAEA significantly and consistently outperforms state-of-the-art entity alignment models.

In knowledge graph embedding models, the margin-based ranking loss as the common loss function is usually used to encourage discrimination between golden triplets and incorrect triplets, which has proved effective in many translation-based models for knowledge graph embedding. However, we find that the loss function cannot ensure the fact that the scoring of correct triplets must be low enough to fulfill the translation. In this paper, we present a limit-based scoring loss to provide lower scoring of a golden triplet, and then to extend two basic translation models TransE and TransH, separately to TransE-RS and TransH-RS by combining limit-based scoring loss with margin-based ranking loss. Both the presented models have low complexities of parameters benefiting for application on large scale graphs. In experiments, we evaluate our models on two typical tasks including triplet classification and link prediction, and also analyze the scoring distributions of positive and negative triplets by different models. Experimental results show that the introduced limit-based scoring loss is effective to improve the capacities of knowledge graph embedding.

Sentiment word identification (SWI) is of high relevance to sentiment analysis technologies and applications. Currently most SWI methods heavily rely on sentiment seed words that have limited sentiment information. Even though there emerge non-seed approaches based on sentiment labels of documents, but in which the context information has not been fully considered. In this paper, based on matrix factorization with co-occurrence neighbor regularization which is derived from context, we propose a novel non-seed model called CONR for SWI. Instead of seed words, CONR exploits two important factors: sentiment matching and sentiment consistency for sentiment word identification. Experimental results on four publicly available datasets show that CONR can outperform the state of-the-art methods.

A new kernel-based learning algorithm called kernel affine subspace nearest point (KASNP) approach is proposed in this paper. Inspired by the geometrical explanation of support vector machines (SVMs) and its nearest point problem in convex hulls, we extend the convex hull of each class to its corresponding affine subspace in high dimensional space induced by kernel. In two class affine subspaces, KASNP finds the nearest points and then constructs a separating hyperplane, which bisects the line segment joining them. The nearest point problem of KASNP is only an unconstrained optimal problem whose solution can be directly computed. Compared with SVM, KASNP avoids solving convex quadratic programming. Experiments on two-spiral dataset, two UCI credit datasets, and face recognition datasets show that our proposed KASNP is effective for data classification.