AORO: Auto-Optimizing Reasoning Order for Multi-Hop Question Answering

Multi-hop question answering (QA) requires a model to retrieve and integrate information from multiple passages to answer a question. Rapid progress has been made on multi-hop QA systems with regard to standard evaluation metrics, including EM and F1. However, by simply evaluating the correctness of the answers, it is unclear to what extent these systems have learned the ability to perform multi-hop reasoning. In this paper, we propose an additional sub-question evaluation for the multi-hop QA dataset HotpotQA, in order to shed some light on explaining the reasoning process of QA systems in answering complex questions. We adopt a neural decomposition model to generate sub-questions for a multi-hop question, followed by extracting the corresponding sub-answers. Contrary to our expectation, multiple state-of-the-art multi-hop QA models fail to answer a large portion of sub-questions, although the corresponding multi-hop questions are correctly answered. Our work takes a step forward towards building a more explainable multi-hop QA system.

Multi-hop question generation (MQG) aims to generate complex questions which require reasoning over multiple pieces of information of the input passage. Most existing work on MQG has focused on exploring graph-based networks to equip the traditional Sequence-to-sequence framework with reasoning ability. However, these models do not take full advantage of the constraint between questions and answers. Furthermore, studies on multi-hop question answering (QA) suggest that Transformers can replace the graph structure for multi-hop reasoning. Therefore, in this work, we propose a novel framework, QA4QG, a QA-augmented BART-based framework for MQG. It augments the standard BART model with an additional multi-hop QA module to further constrain the generated question. Our results on the HotpotQA dataset show that QA4QG outperforms all state-of-the-art models, with an increase of 8 BLEU-4 and 8 ROUGE points compared to the best results previously reported. Our work suggests the advantage of introducing pre-trained language models and QA module for the MQG task.

Knowledge editing aims to update Large Language Models (LLMs) with new information without costly retraining.However, consistently reflecting these updates in complex multihop Question Answering (QA), which demands reasoning over interconnected facts, is challenging.Many existing methods overlook the interplay with pre-existing knowledge, leading to inconsistent edit propagation.To overcome this, we introduce StepKE (Stepwise Knowledge Editing for Multi-hop QA), a novel framework for robustly integrating edited and existing knowledge for coherent multi-hop reasoning.StepKE uniquely decomposes multihop questions into sequential single-hop subquestions, retrieving relevant facts (both edited and pre-existing) from an external knowledge graph for each step.It employs context-aware prompting with prior reasoning history and finetuning for precise edit propagation.This systematic integration enables effective stepwise reasoning.Experiments show StepKE generates significantly more accurate and consistent responses than baselines, showcasing strong knowledge editing and integration in multi-hop QA.

Multi-hop questions are associated with a series of justifications, and one needs to obtain the answers by following the reasoning path (RP) that orders the justifications adequately. So reasoning path retrieval becomes a critical preliminary stage for multi-hop Question Answering (QA). Within the RP, two fundamental challenges emerge for better performance: (i) what the order of the justifications in the RP should be, and (ii) what if the wrong justification has been in the path. In this paper, we propose Reasoning Path Augmentation (RPA), which uses reasoning path reordering and augmentation to handle the above two challenges, respectively. Reasoning path reordering restructures the reasoning by targeting the easier justification first but difficult one later, in which the difficulty is determined by the overlap between query and justifications since the higher overlap means more lexical relevance and easier searchable. Reasoning path augmentation automatically generates artificial RPs, in which the distracted justifications are inserted to aid the model recover from the wrong justification. We build RPA with a naive pre-trained model and evaluate RPA on the QASC and MultiRC datasets. The evaluation results demonstrate that RPA outperforms previously published reasoning path retrieval methods, showing the effectiveness of the proposed methods. Moreover, we present detailed experiments on how the orders of justifications and the percent of augmented paths affect the question- answering performance, revealing the importance of polishing RPs and the necessity of augmentation.

In today’s fast-paced industry, professionals face the challenge of summarizing a large number of documents and extracting vital information from them on a daily basis. These metrics are frequently hidden away in tables and/or their nested hyperlinks. To address this challenge, the approach of Table Question Answering (QA) has been developed to extract the relevant information. However, traditional Table QA training tasks that provide a table and an answer(s) from a gold cell coordinate(s) for a question may not always ensure extracting the accurate answer(s). Recent advancements in Large Language Models (LLMs) have opened up new possibilities for extracting information from tabular data using prompts. In this paper, we introduce the Multi-hop Few-shot Open Rich Table QA (MFORT-QA) approach, which consists of two major steps. The first step involves Few-Shot Learning (FSL), where relevant tables and associated contexts of hyperlinks are retrieved based on a given question. The retrieved content is then used to construct few-shot prompts as inputs to an LLM, such as ChatGPT. To tackle the challenge of answering complex questions, the second step leverages Chain-of-thought (CoT) prompting to decompose the complex question into a sequential chain of questions and reasoning thoughts in a multi-hop manner. Retrieval-Augmented Generation (RAG) enhances this process by retrieving relevant tables and contexts of hyperlinks that are relevant to the resulting reasoning thoughts and questions. These additional contexts are then used to supplement the prompt used in the first step, resulting in more accurate answers from an LLM. Empirical results from OTT-QA demonstrate that our abstractive QA approach significantly improves the accuracy of extractive Table QA methods.

Question generation (QG) attempts to solve the inverse of question answering (QA) problem by generating a natural language question given a document and an answer. While sequence to sequence neural models surpass rule-based systems for QG, they are limited in their capacity to focus on more than one supporting fact. For QG, we often require multiple supporting facts to generate high-quality questions. Inspired by recent works on multi-hop reasoning in QA, we take up Multi-hop question generation, which aims at generating relevant questions based on supporting facts in the context. We employ multitask learning with the auxiliary task of answer aware supporting fact prediction to guide the question generator. In addition, we also proposed a question-aware reward function in a Reinforcement Learning (RL) framework to maximize the utilization of the supporting facts. We demonstrate the effectiveness of our approach through experiments on the multi-hop question answering dataset, HotPotQA. Empirical evaluation shows our model to outperform the single-hop neural question generation models on both automatic evaluation metrics such as BLEU, METEOR, and ROUGE and human evaluation metrics for quality and coverage of the generated questions.

Multimodal Knowledge Editing (MKE) extends traditional knowledge editing to settings involving both textual and visual modalities. However, existing MKE benchmarks primarily assess final answer correctness, neglecting the quality of intermediate reasoning and robustness to visually rephrased inputs. To address this limitation, we introduce MMQAKE, the first benchmark for multimodal multihop question answering with knowledge editing. MMQAKE evaluates: (1) a model’s ability to reason over 2–5-hop factual chains that span both text and images, including performance at each intermediate step; (2) robustness to visually rephrased inputs in multihop questions. Our evaluation shows that current MKE methods often struggle to consistently update and reason over multimodal reasoning chains following knowledge edits. To overcome these challenges, we propose Hybrid-DMKG, a hybrid reasoning framework built on a dynamic multimodal knowledge graph (DMKG) to enable accurate multihop reasoning over updated multimodal knowledge. Hybrid-DMKG first uses a large language model to decompose multimodal multihop questions into sequential sub-questions, then applies a multimodal retrieval model to locate updated facts by jointly encoding each sub-question with candidate entities and their associated images. For answer inference, a hybrid reasoning module operates over the DMKG via two parallel paths: (1) relation-linking prediction; (2) RAG Reasoning with large vision-language models. A background-reflective decision module then aggregates evidence from both paths to select the most credible answer. Experimental results on MMQAKE show that Hybrid-DMKG significantly outperforms existing MKE approaches, achieving higher accuracy and improved robustness to knowledge updates.

Given unstructured text, Large Language Models (LLMs) are adept at answering simple (single-hop) questions. However, as the complexity of the questions increase, the performance of LLMs degrade. We believe this is due to the overhead associated with understanding the complex question followed by filtering and aggregating unstructured information in the raw text. Recent methods try to reduce this burden by integrating structured knowledge triples into the raw text, aiming to provide a structured overview that simplifies information processing. However, this simplistic approach is query-agnostic and the extracted facts are ambiguous as they lack context. To address these drawbacks and to enable LLMs to answer complex (multi-hop) questions with ease, we propose to use a knowledge graph (KG) that is context-aware and is distilled to contain query-relevant information. The use of our compressed distilled KG as input to the LLM results in our method utilizing up to 67% fewer tokens to represent the query relevant information present in the supporting documents, compared to the state-of-the-art (SoTA) method. Our experiments show consistent improvements over the SoTA across several metrics (EM, F1, BERTScore, and Human Eval) on two popular benchmark datasets (HotpotQA and MuSiQue).

Large Language Models (LLMs) have shown proficiency in question-answering tasks but often struggle to integrate real-time knowledge, leading to potentially outdated or inaccurate responses. This problem becomes even more challenging when dealing with multi-hop questions, since they require LLMs to update and integrate multiple knowledge pieces relevant to the questions. To tackle the problem, we propose the Retrieval-Augmented model Editing (RAE) framework for multi-hop question answering. RAE first retrieves edited facts and then refines the language model through in-context learning. Specifically, our retrieval approach, based on mutual information maximization, leverages the reasoning abilities of LLMs to identify chain facts that traditional similarity-based searches might miss. In addition, our framework includes a pruning strategy to eliminate redundant information from the retrieved facts, which enhances the editing accuracy and mitigates the hallucination problem. Our framework is supported by theoretical justification for its fact retrieval efficacy. Finally, comprehensive evaluation across various LLMs validates RAE's ability in providing accurate answers with updated knowledge. Our code is available at: https://github.com/sycny/RAE.

Most log-based anomaly detectors assume logs are stable, though in reality they are often unstable due to software or environmental changes. Anomaly detection on unstable logs (ULAD) is therefore a more realistic, yet under-investigated challenge. Current approaches predominantly employ machine learning (ML) models, which often require extensive labeled data for training. To mitigate data insufficiency, we propose FlexLog , a novel hybrid approach for ULAD that combines ML models — decision tree, k-nearest neighbors, and a feedforward neural network — with a Large Language Model (Mistral) through ensemble learning. FlexLog also incorporates a cache and retrieval-augmented generation (RAG) to further enhance efficiency and effectiveness. To evaluate FlexLog , we configured four datasets for ULAD, namely ADFA-U, LOGEVOL-U, SynHDFS-U, and SYNEVOL-U. FlexLog outperforms all baselines by at least 1.2 percentage points (pp) in F1 score while using much less labeled data (62.87 pp reduction). When trained on the same amount of data as the baselines, FlexLog achieves up to a 13 pp increase in F1 score on ADFA-U across varying training dataset sizes. Additionally, FlexLog maintains inference time under one second per log sequence, making it suitable for most applications, except latency-sensitive systems. Further analysis reveals the positive impact of FlexLog ’s key components: cache, RAG, and ensemble learning.

Explainability and interpretability are two important concepts, the absence of which can and should impede the application of well-performing neural networks to real-world problems. At the same time, they are difficult to incorporate into the large, black-box models that achieve state-of-the-art results in a multitude of NLP tasks. Bidirectional Encoder Representations from Transformers (BERT) is one such black-box model. It has become a staple architecture to solve many different NLP tasks and has inspired a number of related Transformer models. Understanding how these models draw conclusions is crucial for both their improvement and application. We contribute to this challenge by presenting VisBERT, a tool for visualizing the contextual token representations within BERT for the task of (multi-hop) Question Answering. Instead of analyzing attention weights, we focus on the hidden states resulting from each encoder block within the BERT model. This way we can observe how the semantic representations are transformed throughout the layers of the model. VisBERT enables users to get insights about the model’s internal state and to explore its inference steps or potential shortcomings. The tool allows us to identify distinct phases in BERT’s transformations that are similar to a traditional NLP pipeline and offer insights during failed predictions.

Most Knowledge Graph-based Question Answering (KGQA) systems rely on training data to reach their optimal performance. However, acquiring training data for supervised systems is both time-consuming and resource-intensive. To address this, in this paper, we propose <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Tree-KGQA</b> , an unsupervised KGQA system leveraging pre-trained language models and tree-based algorithms. Entity and relation linking are essential components of any KGQA system. We employ several pre-trained language models in the entity linking task to recognize the entities mentioned in the question and obtain the contextual representation for indexing. Furthermore, for relation linking we incorporate a pre-trained language model previously trained for language inference task. Finally, we introduce a novel algorithm for extracting the answer entities from a KG, where we construct a forest of interpretations and introduce tree-walking and tree disambiguation techniques. Our algorithm uses the linked relation and predicts the tree branches that eventually lead to the potential answer entities. The proposed method achieves 4.5% and 7.1% gains in F1 score in entity linking tasks on LC-QuAD 2.0 and LC-QuAD 2.0 (KBpearl) datasets, respectively, and a 5.4% increase in the relation linking task on LC-QuAD 2.0 (KBpearl). The comprehensive evaluations demonstrate that our unsupervised KGQA approach outperforms other supervised state-of-the-art methods on the WebQSP-WD test set (1.4% increase in F1 score) - without training on the target dataset.

Coffee bean production can encounter challenges due to fluctuations in global coffee prices, impacting the economic stability of some countries that heavily depend on coffee production. The primary objective is to evaluate how effectively various pre-trained models can predict coffee types using advanced deep learning techniques. The selection of an optimal pre-trained model is crucial, given the growing popularity of specialty coffee and the necessity for precise classification. We conducted a comprehensive comparison of several pre-trained models, including AlexNet, LeNet, HRNet, Google Net, Mobile V2 Net, ResNet (50), VGG, Efficient, Darknet, and DenseNet, utilizing a coffee-type dataset. By leveraging transfer learning and fine-tuning, we assess the generalization capabilities of the models for the coffee classification task. Our findings emphasize the substantial impact of the pre-trained model choice on the model's performance, with certain models demonstrating higher accuracy and faster convergence than conventional alternatives. This study offers a thorough evaluation of pre-trained architectural models regarding their effectiveness in coffee classification. Through the evaluation of result metrics, including sensitivity (1.0000), specificity (0.9917), precision (0.9924), negative predictive value (1.0000), accuracy (1.0000), and F1 score (0.9962), our analysis provides nuanced insights into the intricate landscape of pre-trained models.

Concept mapping is a knowledge representation method used to represent and understand concepts, entities, and the relationships between them, which are referred to as hyponymy–hypernymy semantic relations. These relations are primarily used to describe the hierarchical and categorical relationships between different concepts or entities. The detection of hyponymy–hypernymy semantic relations is an important task in the field of natural language processing, crucial to many downstream tasks such as information extraction, automatic reasoning, and personalized recommendations. These tasks often require understanding the semantic relations between concepts or entities in text for more accurate analysis and reasoning. Currently, algorithms for identifying and detecting hyponymy–hypernymy semantic relations face two main challenges: first, candidate hyponymy–hypernymy relation tuples do not exist in the same contextual sentence, failing to meet the co-occurrence requirement; second, distributed algorithms have issues with lexical memory. To address these issues, this paper proposes an improved algorithm for detecting hyponymy–hypernymy relations based on heterogeneous graph neural networks, aiming to detect hyponymy–hypernymy relations in various candidate word sets and then construct a hierarchical system. To meet the co-occurrence requirement of candidate word pairs, an open-source large model is utilized to generate contextual sentences for the candidate pairs. Sub-word features are adopted to capture the intrinsic semantic connections between nested phrases, thus alleviating the issue of reversed predictions for nested phrases. The representation of relation nodes is modified by encoding relation definitions through a pre-trained model, enabling the model to understand the semantic relations between concept nodes. To address the problem of overfitting in traditional graph attention networks, the calculation order of adjacency node aggregation is changed in the heterogeneous graph to capture dynamic attention features. In addition, a pipeline for hierarchical system construction is designed and implemented, combining the divide-and-conquer approach with loop detection algorithms. Compared to baseline metrics, the proposed method achieves a 4.14% improvement in accuracy and a 0.62% increase in F1 score on the EVALution dataset; a 4.89% increase in accuracy and a 0.71% improvement in F1 score on the Bansal dataset; and a 1.05% increase in accuracy, a 2.21% increase in recall, and a 1.79% improvement in F1 score on a self-annotated Chinese dictionary dataset.

The goal of Target Sentiment Analysis (TSA) is to predict the users’ sentiment towards specific targets from review sentences. However, the predicting results may not perform well due to the sparsity of training data. Data augmentation is a fruitful technology to alleviate the influence of imperfect training data, which obtains additional data by transforming the original samples. Unfortunately, there is hardly a particular data augmentation approach for TSA. To address this problem, in this paper, we propose a low-cost and effective data augmentation method called SubCrime, which constructs auxiliary sentences in two steps: Substitute and disCriminate. The former aims to substitute reasonable targets for the observed sentences through the masked language model, while the latter discriminates the restructured sentences via the constrained objective. SubCrime does not require extra knowledge and tedious manual annotation. We design SubCrime to answer the key counterfactual question: “If the review target in the sentence changed, would its sentiment be different ?”. Experiments show SubCrime improves on average 2 to 4 points in F1 scores on four datasets compared to methods without enhancement. Moreover, SubCrime also outperforms other data augmentation methods widely used in other Natural Language Processing (NLP) tasks.KeywordsTarget Sentiment AnalysisCounterfactualData augmentation