Incorporating anticipation embedding into reinforcement learning framework for multi-hop knowledge graph question answering

Document-level relation extraction aims to infer complex semantic relations among entities in an entire document. Compared with the sentence-level relation extraction, document-level relational facts are expressed by multiple mentions across the sentences in a long-distance, requiring excellent reasoning. In this paper, we propose Dual-Channel and Hierarchical Graph Convolutional Networks (DHGCN), which constructs three graphs in token-level, mention-level, and entity-level to model complex interactions among different semantic representations across the document. Based on the multi-level graphs, we apply the Graph Convolutional Network (GCN) for each level to aggregate the relevant information scattered throughout the document for better inferring the implicit relations. Moreover, we propose a dual-channel encoder to capture structural and contextual information simultaneously, which also supplies the contextual representation for the higher layer to avoid losing low-dimension information. Our DHGCN yields significant improvements over the state-of-the-art methods by 2.75, 5.5, and 3.5 F1 on DocRED, CDR, and GDA, respectively, which are popular document-level relation extraction datasets. Furthermore, to demonstrate the effectiveness of our method, we evaluate DHGCN on a fine gained clinical document-level dataset Symptom-Acupoint Relation (SAR) proposed by ourselves and available at https://github.com/QiSun123/SAR. The experimental results illustrate that DHGCN is able to infer more valuable relations among entities in the document. 

Dual-Channel and Hierarchical Graph Convolutional Networks for document-level relation extraction

/pdf/distantly-supervised-relation-extraction-via-recursive-27o08uvd.pdf

Distantly Supervised Relation Extraction via Recursive Hierarchy-Interactive Attention and Entity-Order Perception

Text-enhanced and implicit reasoning methods are proposed for answering questions over incomplete knowledge graph (KG), whereas prior studies either rely on external resources or lack necessary interpretability. This article desires to extend the line of reinforcement learning (RL) methods for better interpretability and dynamically augment original KG action space with additional actions. To this end, we propose a RL framework along with a dynamic completion mechanism, namely Dynamic Completion Reasoning Network (DCRN). DCRN consists of an action space completion module and a policy network. The action space completion module exploits three sub-modules (relation selector, relation pruner and tail entity predictor) to enrich options for decision making. The policy network calculates probability distribution over joint action space and selects promising next-step actions. Simultaneously, we employ the beam search-based action selection strategy to alleviate delayed and sparse rewards. Extensive experiments conducted on WebQSP, CWQ and MetaQA demonstrate the effectiveness of DCRN. Specifically, under 50% KG setting, the [email protected] performance improvements of DCRN on MetaQA-1H and MetaQA-3H are 2.94% and 1.18% respectively. Moreover, under 30% and 10% KG settings, DCRN prevails over all baselines by 0.9% and 1.5% on WebQSP, indicating the robustness to sparse KGs. 

Reinforcement learning with dynamic completion for answering multi-hop questions over incomplete knowledge graph

We investigate distantly supervised relation extraction with knowledge-guided latent graphs and an iterative graph learner. For the relation extraction tasks, we assume that the input sentences contain latent graphs with useful structural information between mentions and relations of the distantly supervised data. We first embed input sentences with default initialized graphs to utilize this information. We also used the pre-trained Knowledge Bases (KB) to guide the latent space of a Variational Graph Auto-Encoders(VGAE) module. Then the VGAE applies the re-parameterization mechanism to reconstruct the initial graph, injecting extra knowledge into the initial latent graph and reducing its unsteadiness. Subsequently, we optimize the latent graph structures and their corresponding node embeddings simultaneously by the iterative graph learner, and the better latent graphs can improve the downstream relation extraction task. Experiment results on three datasets (NYT10, WIKIDISTANT and GIDS) show that latent graph learning helps our model perform better than previous works. 

Distantly supervised relation extraction with KB-enhanced reconstructed latent iterative graph networks

Question answering over knowledge graph (KGQA), which automatically answers natural language questions by querying the facts in knowledge graph (KG), has drawn significant attention in recent years. In this paper, we focus on single-relation questions, which can be answered through a single fact in KG. This task is a non-trivial problem since capturing the meaning of questions and selecting the golden fact from billions of facts in KG are both challengeable. We propose a pipeline framework for KGQA, which consists of three cascaded components: (1) an entity detection model, which can label the entity mention in the question; (2) a novel entity linking model, which considers the contextual information of candidate entities in KG and builds a question pattern classifier according to the correlations between question patterns and relation types to mitigate entity ambiguity problem; and (3) a simple yet effective relation detection model, which is used to match the semantic similarity between the question and relation candidates. Substantial experiments on the SimpleQuestions benchmark dataset show that our proposed method could achieve better performance than many existing state-of-the-art methods on accuracy, top-N recall and other evaluation metrics.

Simple Question Answering over Knowledge Graph Enhanced by Question Pattern Classification

To find relational facts of interest from plain texts, distantly supervised relation extraction (DSRE) has drawn significant attention. Recent works exploit relation hierarchies to mine more clues for long-tail relations and achieve good performance. However, they ignore or underutilize the correlation of relations in the hierarchical structure. Empirically, the correlation facilitates knowledge transfer between different relations to further handle long-tail relations and improves inter-relational discrimination. In this paper, we devise an end-to-end network to model the correlation of relations from two perspectives. Globally, we construct an undirected connected graph according to the relation hierarchies, and employ Graph Attention Networks (GATs) to aggregate node information and generate correlation-aware Global Hierarchy Embeddings (GHE). Locally, we assume that along the relation hierarchies, the classification results of adjacent levels should be highly interdependent, and introduce a constraint called Local Probability Constraints (LPC) to take it into account. LPC is then combined with a branch network for both sentence-level and bag-level classification. Experimental results on the popular New York Times (NYT) dataset show that, our model GHE-LPC outperforms other state-of-the-art baselines in terms of AUC, Top-N precision, accuracy of Hits@K, etc.

Distantly Supervised Relation Extraction using Global Hierarchy Embeddings and Local Probability Constraints

Distantly supervised relation extraction has drawn significant attention recently. However, almost all prior works ignore the fact that, in a sentence, the appearance order of two entities contributes to the understanding of its semantics. Furthermore, they leverage relation hierarchies but don't fully exploit the heuristic effect between relation levels, i.e., higher-level relations can give useful information to the lower ones. In this paper, we design a novel Recursive Hierarchy-Interactive Attention network (RHIA), which uses the hierarchical structure of the relation to model the interactive information between the relation levels to further handle long-tail relations. It generates relation-augmented sentence representations along hierarchical relation chains in a recursive structure. Besides, we introduce a newfangled training objective, called Entity-Order Perception (EOP), to make the sentence encoder retain more entity appearance information. Substantial experiments on the popular New York Times (NYT) dataset are conducted. Compared to prior baselines, our approach achieves state-of-the-art performance in terms of precision-recall (P-R) curves, AUC, Top-N precision and other evaluation metrics.

Distantly Supervised Relation Extraction via Recursive Hierarchy-Interactive Attention and Entity-Order Perception.

The invention discloses an intelligent question-answering method and system The method comprises the following steps: constructing a knowledge graph and a corpus; obtaining a question input by a user; and searching a question answer based on the knowledge graph subgraph search method and the question input by the user, and if the answer is successfully retrieved, outputting the answer The intelligent question-answering method and system provided by the invention are realized based on knowledge graph and sub-graph search, have relatively high identification accuracy, can answer complex questions, and are high in question answering capability

Cui Hai

Papers

Simple Question Answering over Knowledge Graph Enhanced by Question Pattern Classification

Distantly Supervised Relation Extraction using Global Hierarchy Embeddings and Local Probability Constraints

Distantly Supervised Relation Extraction via Recursive Hierarchy-Interactive Attention and Entity-Order Perception.

Intelligent question-answering method and system