Multi-hop Reasoning Research Articles

Heuristic Heterogeneous Graph Reasoning Networks for Fact Verification.

Existing studies on table-based fact verification generally capture linguistic evidence from claim-table subgraphs or logical evidence from program-table subgraphs independently. However, there is insufficient association interaction between the two types of evidence, which makes it difficult to obtain valuable consistency features between them. In this work, we propose heuristic heterogeneous graph reasoning networks (H2GRN) to capture the shared consistent evidence by strengthening associations between linguistic and logical evidence from two perspectives of graph construction and reasoning mechanism. Specifically, 1) to enhance the close connectivity of the two subgraphs, rather than simply connecting two subgraphs by the nodes with the same content (the constructed graph in this way has severe sparsity), we construct a heuristic heterogeneous graph, which relies on claim semantics as heuristic knowledge to guide the connections of the program-table subgraph, and in turn expands the connectivity of the claim-table subgraph through logical information of programs as heuristic knowledge; and 2) to establish adequate association interaction between linguistic evidence and logical evidence, we design multiview reasoning networks. In detail, we propose local-view multihop knowledge reasoning (MKR) networks to enable the current node to establish association not only with one-hop neighbors, but also with multihop neighbors, to capture context-richer evidence information. We execute MKR on heuristic claim-table and program-table subgraphs to learn context-richer linguistic evidence and logical evidence, respectively. Meanwhile, we develop global-view graph dual-attention networks (DAN) that execute on the entire heuristic heterogeneous graph, reinforcing global-level significant consistency evidence. Finally, the consistency fusion layer is devised to weaken the disagreement between the three types of evidence to assist in capturing consistent shared evidence for verifying claims. Experiments on TABFACT and FEVEROUS demonstrate the effectiveness of H2GRN.

Adversarial Entity Graph Convolutional Networks for multi-hop inference question answering

Multi-hop reasoning is critical for natural language understanding but poses challenges for current models, requiring models capable of aggregating and reasoning across multiple sources of information. We propose an Adversarial Entity Graph Convolutional Network (AEGCN) to improve multi-hop inference performance. Unlike previous GNNs-based models, AEGCN places a greater emphasis on the construction of rich entity graph which focuses on identifying the related entities from support document and connecting these entities with innovative edge relationships. The entity graph built by AEGCN preserves both the semantic information and structure of the original text. Further, adversarial training is adopted to generate challenging embeddings for the entity graph, increasing the model’s robustness against the interference. The experiments evaluated on WIKIHOP and MEDHOP dataset indicate that AEGCN achieves 6.8 % and 7.7 % accuracy improvement over baseline model respectively, confirming the model’s advanced capability in multi-hop reasoning tasks.

GS-CBR-KBQA: Graph-structured case-based reasoning for knowledge base question answering

Knowledge Base Question Answering (KBQA) task is an important research direction in natural language processing. Due to the flexibility and ambiguity of natural language, users’ questions often have more complex query types and richer semantic information. To address this issue, this paper proposes the GS-CBR-KBQA model, a Case-Based Reasoning model tailored for KBQA to improve the semantic parsing accuracy and interpretability of natural language questions. The model integrates Knowledge-oriented Programming Language (KoPL) reasoning graphs with query information, employing a Graph Auto-Encoder and the RoBERTa pretrained language model for a highly effective case retrieval. This integration leads to a more robust knowledge retrieval and application approach, particularly innovative in capturing the relationships within KoPL graphs. The model addresses explicitly complex questions such as multi-hop reasoning and questions involving intricate entity relationships. Finally, our extensive experiments show that the model performs excellently in accuracy and F1 metrics on benchmark datasets such as WebQSP and ComplexWebQuestions, particularly in complex question-answering. The code of our model is available at https://anonymous.4open.science/r/GS-CBR-KBQA.

Deconstructing reasoning paths and attending to semantic guidance for document-level relation extraction

Extracting relations from a document is more challenging than from a sentence, due to the involvement of more entities and complex contextual information. To capture long-distance contextual dependencies, graph networks are widely used in document-level relation extraction. However, lengthy reasoning paths, built by most models based on graph networks to capture dependency, will result in the gradual accumulation of noise information as these paths extend. Furthermore, semantic information that is not explicitly present will be easily overlooked due to the utilization of heuristic rules for constructing document structural feature-based static graphs. To tackle these issues, we propose a novel framework to extract document-level relations by Deconstructing Reasoning Paths and Attending to Semantic Guidance, namely DRPASG. The model progressively deconstructs reasoning paths and dynamically aggregates multi-level representations of entities. Furthermore, a unique attention regularizer is constructed to guide the model in focusing on uncaptured semantic correlations, facilitating the transmission and interaction of both structural and semantic information within the document. Experimental results on the document-level dataset DocRED indicate that the proposed model outperforms previous state-of-the-art methods. Extensive analysis validates the effectiveness of our method in performing multi-hop reasoning.

Nonfactoid Question Answering as Query-Focused Summarization With Graph-Enhanced Multihop Inference.

Nonfactoid question answering (QA) is one of the most extensive yet challenging applications and research areas in natural language processing (NLP). Existing methods fall short of handling the long-distance and complex semantic relations between the question and the document sentences. In this work, we propose a novel query-focused summarization method, namely a graph-enhanced multihop query-focused summarizer (GMQS), to tackle the nonfactoid QA problem. Specifically, we leverage graph-enhanced reasoning techniques to elaborate the multihop inference process in nonfactoid QA. Three types of graphs with different semantic relations, namely semantic relevance, topical coherence, and coreference linking, are constructed for explicitly capturing the question-document and sentence-sentence interrelationships. Relational graph attention network (RGAT) is then developed to aggregate the multirelational information accordingly. In addition, the proposed method can be adapted to both extractive and abstractive applications as well as be mutually enhanced by joint learning. Experimental results show that the proposed method consistently outperforms both existing extractive and abstractive methods on two nonfactoid QA datasets, WikiHow and PubMedQA, and possesses the capability of performing explainable multihop reasoning.

KnowledgeNavigator: leveraging large language models for enhanced reasoning over knowledge graph

Large language models have achieved outstanding performance on various downstream tasks with their advanced understanding of natural language and zero-shot capability. However, they struggle with knowledge constraints, particularly in tasks requiring complex reasoning or extended logical sequences. These limitations can affect their performance in question answering by leading to inaccuracies and hallucinations. This paper proposes a novel framework called KnowledgeNavigator that leverages large language models on knowledge graphs to achieve accurate and interpretable multi-hop reasoning. Especially with an analysis-retrieval-reasoning process, KnowledgeNavigator searches the optimal path iteratively to retrieve external knowledge and guide the reasoning to reliable answers. KnowledgeNavigator treats knowledge graphs and large language models as flexible components that can be switched between different tasks without additional costs. Experiments on three benchmarks demonstrate that KnowledgeNavigator significantly improves the performance of large language models in question answering and outperforms all large language models-based baselines.

HierarT: Multi-hop temporal knowledge graph forecasting with hierarchical reinforcement learning

Temporal knowledge graph (TKG) multi-hop reasoning is a dominant approach that infers the target entity by walking along the path connecting entities and relations. However, in most TKGs, there are multiple relations related to an identical entity and multiple tail entities for an identical pair of head entity and relation. This characteristic leads to a significant amount of redundant data in the binary action space which is a collection of binary tuples composed of relations and entities, thereby impeding model inference, i.e., action space explosion. We propose a method to address this issue that dismantles the reasoning process into a relation level for relation reasoning and an entity level for entity reasoning. In addition, hybrid time encoding is proposed to enhance the utilization of the timestamp information in the reasoning process. Moreover, K-means-based reward shaping is proposed to alleviate the issue of sparse reward matrices by comparing the pre-clustered labels of both the predicted entity and the target entity. A text transformer module is used to deal with the limitation of a single information modality by integrating the text information of TKG into model reasoning. Our method is evaluated using four benchmark datasets, and the results verify its superiority over state-of-the-art baselines.

ChatGPT: The End of Online Exam Integrity?

This study addresses the significant challenge posed by the use of Large Language Models (LLMs) such as ChatGPT on the integrity of online examinations, focusing on how these models can undermine academic honesty by demonstrating their latent and advanced reasoning capabilities. An iterative self-reflective strategy was developed for invoking critical thinking and higher-order reasoning in LLMs when responding to complex multimodal exam questions involving both visual and textual data. The proposed strategy was demonstrated and evaluated on real exam questions by subject experts and the performance of ChatGPT (GPT-4) with vision was estimated on an additional dataset of 600 text descriptions of multimodal exam questions. The results indicate that the proposed self-reflective strategy can invoke latent multi-hop reasoning capabilities within LLMs, effectively steering them towards correct answers by integrating critical thinking from each modality into the final response. Meanwhile, ChatGPT demonstrated considerable proficiency in being able to answer multimodal exam questions across 12 subjects. These findings challenge prior assertions about the limitations of LLMs in multimodal reasoning and emphasise the need for robust online exam security measures such as advanced proctoring systems and more sophisticated multimodal exam questions to mitigate potential academic misconduct enabled by AI technologies.

Multi-hop temporal knowledge graph reasoning with multi-agent reinforcement learning

Knowledge graph (KG) is a key component of artificial intelligence. In recent years, many large-scale knowledge graphs have been produced and put into practical applications. At present, researchers have proposed many methods to reason facts that do not exist in knowledge graphs using the existing information. However, most traditional reasoning methods lack interpretability, and cannot get the reasoning paths. Therefore, the multi-hop path reasoning method of knowledge graph has gradually become a hot spot. This method finds the next relations through the reinforcement learning agent, gets the whole path, and scores the path. Although these multi-hop path reasoning methods make the reasoning method interpretable, these multi-hop path reasoning methods focus on the choice of relations, ignoring the importance of entities in the path reasoning process. Moreover, with the development of the Temporal Knowledge Graph (TKG), traditional multi-hop path reasoning methods cannot effectively process time information. To solve these problems, a multi-hop path reasoning method of temporal knowledge graph based on multi-agent reinforcement learning is proposed, which is named MA-TPath. MA-TPath uses two agents to perform relation selection and entity selection iteratively. Meanwhile, considering the diversity of temporal reasoning paths, we propose a new type of reward function. In MA-TPath, two agents employ the Long Short-Term Memory Networks (LSTM) to capture current results from the environment and output the corresponding action vectors to the environment through activation functions. Experimentally, MA-TPath outperforms all existing models in 4 out of 4 indicators on the YAGO dataset, 3 out of 4 indicators on the GDETL-5 dataset and the ICEWS-250 dataset, and 1 out of 4 indicators on the ICEWS05–15 dataset. Analysis of the temporal reasoning path indicates that MA-TPath not only surpasses other state-of-the-art reasoning methods over four public temporal knowledge graph datasets but provides rationality for results.

Advancing document-level relation extraction with a syntax-enhanced multi-hop reasoning network

Document-level relation extraction aims to uncover relations between entities by harnessing the intricate information spread throughout a document. Previous research involved constructing discrete syntactic matrices to capture syntactic relationships within documents. However, these methods are significantly influenced by dependency parsing errors, leaving much of the latent syntactic information untapped. Moreover, prior research has mainly focused on modeling two-hop reasoning between entity pairs, which has limited applicability in scenarios requiring multi-hop reasoning. To tackle these challenges, a syntax-enhanced multi-hop reasoning network (SEMHRN) is proposed. Specifically, the approach begins by using a dependency probability matrix that incorporates richer grammatical information instead of a sparse syntactic parsing matrix to build the syntactic graph. This effectively reduces syntactic parsing errors and enhances the model’s robustness. To fully leverage dependency information, dependency-type-aware attention is introduced to refine edge weights based on connecting edge types. Additionally, a part-of-speech prediction task is included to regularize word embeddings. Unrelated entity pairs can disrupt the model’s focus, reducing its efficiency. To concentrate the model’s attention on related entity pairs, these related pairs are extracted, and a multi-hop reasoning graph attention network is employed to capture the multi-hop dependencies among them. Experimental results on three public document-level relation extraction datasets validate that SEMHRN achieves a competitive F1 score compared to the current state-of-the-art methods.

Hierarchical Knowledge-Enhancement Framework for multi-hop knowledge graph reasoning

Multi-hop reasoning has gained significant attention in the area of knowledge graph completion, intending to predict missing facts through existing knowledge. However, due to sparsity knowledge and lack of reachable paths for the model to make sensible decisions, most multi-hop reasoning models suffer from degradation performance on the sparse knowledge graphs (KGs). Faced with this challenge, we propose the Hierarchical Knowledge-Enhancement Framework (Hi-KnowE), which improves reasoning performance over sparse KGs by conducting hierarchical dynamic path completion under various background knowledge guidance. Hi-KnowE is based on hierarchical reinforcement learning, breaking the task into high-level and low-level processes. Firstly, we apply high-level policy to reason relations and the relation action space is enlarged under rule guidance. Secondly, tail entities are obtained by low-level policy and semantic knowledge is used to refine the entity action space. The sparsity problem is alleviated by generating reliable information. Moreover, we introduce rule-based inner attention, adopting the overall semantics of relevant rules to assist agent reasoning. This strategy makes the path for the agent more logical and flexible. We further evaluate Hi-KnowE on four sparse datasets extracted from Wikidata and FreeBase. The results outperform the baseline models, demonstrating improvements of up to 1.5% and 1.9% in MRR and Hit@10.

Advancing Spatial Reasoning in Large Language Models: An In-Depth Evaluation and Enhancement Using the StepGame Benchmark

Artificial intelligence (AI) has made remarkable progress across various domains, with large language models like ChatGPT gaining substantial attention for their human-like text-generation capabilities. Despite these achievements, improving spatial reasoning remains a significant challenge for these models. Benchmarks like StepGame evaluate AI spatial reasoning, where ChatGPT has shown unsatisfactory performance. However, the presence of template errors in the benchmark has an impact on the evaluation results. Thus there is potential for ChatGPT to perform better if these template errors are addressed, leading to more accurate assessments of its spatial reasoning capabilities. In this study, we refine the StepGame benchmark, providing a more accurate dataset for model evaluation. We analyze GPT’s spatial reasoning performance on the rectified benchmark, identifying proficiency in mapping natural language text to spatial relations but limitations in multi-hop reasoning. We provide a flawless solution to the benchmark by combining template-to-relation mapping with logic-based reasoning. This combination demonstrates proficiency in performing qualitative reasoning on StepGame without encountering any errors. We then address the limitations of GPT models in spatial reasoning. To improve spatial reasoning, we deploy Chain-of-Thought and Tree-of-thoughts prompting strategies, offering insights into GPT’s cognitive process. Our investigation not only sheds light on model deficiencies but also proposes enhancements, contributing to the advancement of AI with more robust spatial reasoning capabilities.

Tree-of-Reasoning Question Decomposition for Complex Question Answering with Large Language Models

Large language models (LLMs) have recently demonstrated remarkable performance across various Natual Language Processing tasks. In the field of multi-hop reasoning, the Chain-of-thought (CoT) prompt method has emerged as a paradigm, using curated stepwise reasoning demonstrations to enhance LLM's ability to reason and produce coherent rational pathways. To ensure the accuracy, reliability, and traceability of the generated answers, many studies have incorporated information retrieval (IR) to provide LLMs with external knowledge. However, existing CoT with IR methods decomposes questions into sub-questions based on a single compositionality type, which limits their effectiveness for questions involving multiple compositionality types. Additionally, these methods suffer from inefficient retrieval, as complex questions often contain abundant information, leading to the retrieval of irrelevant information inconsistent with the query's intent. In this work, we propose a novel question decomposition framework called TRQA for multi-hop question answering, which addresses these limitations. Our framework introduces a reasoning tree (RT) to represent the structure of complex questions. It consists of four components: the Reasoning Tree Constructor (RTC), the Question Generator (QG), the Retrieval and LLM Interaction Module (RAIL), and the Answer Aggregation Module (AAM). Specifically, the RTC predicts diverse sub-question structures to construct the reasoning tree, allowing a more comprehensive representation of complex questions. The QG generates sub-questions for leaf-node in the reasoning tree, and we explore two methods for QG: prompt-based and T5-based approaches. The IR module retrieves documents aligned with sub-questions, while the LLM formulates answers based on the retrieved information. Finally, the AAM aggregates answers along the reason tree, producing a definitive response from bottom to top.

Learning to walk with logical embedding for knowledge reasoning

The path-based model has remarkably succeeded in the knowledge graph (KG) multi-hop reasoning task. It employs all available resources to accomplish various complex path reasoning tasks and continuously explores new graph paths. However, existing multi-hop reasoning methods rely heavily on the high reward, which is fed back to the model when the agent searches for the target. In contrast, most previous methods focused on efficiently querying the correct answer while disregarding the logic and validity of the entire reasoning chain. It contradicts the intention of various complex reasoning tasks in real-world scenarios. Unreasonable paths will cause the selection to deviate from normal cognition, resulting in invalid path resource information. Therefore, we must be able to complete specific tasks via these reliable paths. In order to address these issues, we proposed a Reinforcement Learning-Based Knowledge Reasoning Model with Logical Embedding (RKLE) to enhance the interpretability of the reasoning chain. RKLE assembles the logical structure (query structure) with additional nodes in the current step and develops the logical reward shaping to assist the agent in selecting a more reasonable path. Finally, experimental results on several benchmarks demonstrate that our approach can search for the correct answers more efficiently than existing path-based methods and that the corresponding reasoning chain is interpretable.

A Relation Embedding Assistance Networks for Multi-hop Question Answering

Multi-hop Knowledge Graph Question Answering aims at finding an entity to answer natural language questions from knowledge graphs. When humans perform multi-hop reasoning, people tend to focus on specific relations across different hops and confirm the next entity. Therefore, most algorithms choose the wrong specific relation, which makes the system deviate from the correct reasoning path. The specific relation at each hop plays an important role in multi-hop question answering. Existing work mainly relies on the question representation as relation information, which cannot accurately calculate the specific relation distribution. In this article, we propose an interpretable assistance framework that fully utilizes the relation embeddings to assist in calculating relation distributions at each hop. Moreover, we employ the fusion attention mechanism to ensure the integrity of relation information and hence to enrich the relation embeddings. The experimental results on three English datasets and one Chinese dataset demonstrate that our method significantly outperforms all baselines. The source code of REAN will be available at https://github.com/2399240664/REAN

DSAMR: Dual-Stream Attention Multi-hop Reasoning for knowledge-based visual question answering

Knowledge-based visual question answering aims to associate external knowledge facts for answering questions about images. Most existing methods emphasize high-order associations between knowledge facts and questions, and fail to consider the negative effects of unnecessary knowledge facts in multi-hop reasoning. In this paper, we propose a Dual-Stream Attention Multi-hop Reasoning (DSAMR) architecture that constructs two different attention streams to mitigate unnecessary knowledge facts. This dual-stream mechanism enables the model to reduce the attention weights on unnecessary knowledge while gathering essential knowledge by learning the implicit correlations between knowledge facts and questions. In addition, we designed a hypergraph knowledge extraction module in the architecture to extract optimal knowledge facts by evaluating the relevance of each knowledge fact to the question. The experimental results demonstrate the effectiveness of our method not only on the knowledge-based visual question answering dataset KVQA, but also on the multi-hop question answering dataset PathQuestion.

TodBR: Target-Oriented Dialog with Bidirectional Reasoning on Knowledge Graph

Target-oriented dialog explores how a dialog agent connects two topics cooperatively and coherently, which aims to generate a “bridging” utterance connecting the new topic to the previous conversation turn. The central focus of this task entails multi-hop reasoning on a knowledge graph (KG) to achieve the desired target. However, current target-oriented dialog approaches suffer from inefficiencies in reasoning and the inability to locate pertinent key information without bidirectional reason. To address these limitations, we present a bidirectional reasoning model for target-oriented dialog implemented on a commonsense knowledge graph. Furthermore, we introduce an automated technique for constructing dialog subgraphs, which aids in acquiring multi-hop reasoning capabilities. Our experiments demonstrate that our proposed method attains superior performance in reaching the target while providing more coherent responses.

Multi-Hop Reasoning With Relation Based Node Quality Evaluation for Sparse Medical Knowledge Graph

Multi-Hop Reasoning With Relation Based Node Quality Evaluation for Sparse Medical Knowledge Graph

Multi-Hop Arabic LLM Reasoning in Complex QA

The introduction of Large Language Models (LLMs), and generative AI has significantly transformed the field of natural language processing. These models have exhibited profound reasoning capabilities, marking considerable progress across diverse general knowledge reasoning tasks. Consequently, the deployment of LLMs in domain-specific contexts has become a prime objective for governments and corporations eager to leverage the generative AI revolution. However, the Arabic language has notably lagged in attention and development compared to other languages in this arena. This research endeavors to delve into various facets of Arabic closed-domain question and answering systems that emulate the reasoning requirements of private enterprise data. Our study focuses on the practical deployment of Arabic LLMs in targeted applications, specifically utilizing the ACQAD (Arabic Complex Question Answering Dataset), which exhibits multi-hop reasoning. Different strategies are experimented using Long Context Window (LCW) and Retrieval Augmented Generation (RAG). Results showed that decomposing complex questions using Chain-of-Thought reasoning considerably improved the performance from 75% to 92% using LCW, but at much higher token cost compared to RAG. Trade-of between cost and performance showed that 80% accuracy can be attained using only 30% of the cost using RAG Sentence - level embeddings. Microsoft E5 embedding model is used and OpenAI GPT4-turbo LLM which proved superior reasoning performance compared to other Arabic LLMs

Multi-hop path reasoning over sparse temporal knowledge graphs based on path completion and reward shaping

Multi-hop path reasoning plays a crucial role in temporal knowledge graph reasoning, aiming to infer deep complex relationships and obtain interpretable reasoning results. Previous reasoning models primarily focus on designing temporal knowledge graphs with rich paths between entities but struggle to handle path sparsity, known as sparse temporal knowledge graphs. Sparse temporal knowledge graphs store only essential knowledge information, resulting in missing connection paths between certain entities and encountering issues of sparse rewards and information scarcity. To tackle these challenges, this paper proposes a multi-hop path reasoning model over sparse temporal knowledge graph based on path completion and reward shaping (STKGR-PR). STKGR-PR dynamically completes missing paths using a temporal embedding model to alleviate path sparsity. To tackle the issue of sparse rewards caused by reduced hit rates, we propose semantic composition-based reasoning path embeddings derived from relation embeddings and timestamp embeddings for reward shaping. Considering the limited information contained in sparse temporal knowledge graphs, we incorporate time vectors into the embedding and multi-hop path reasoning models to enhance the accuracy of reasoning paths and results. Experimental results conducted on six sparse temporal datasets sampled from ICEWS14 and ICEWS515–05 demonstrate that STKGR-PR outperforms state-of-the-art multi-hop path reasoning models over temporal knowledge graphs across all evaluation metrics, while ensuring interpretability.