Hallucination Diversity-Aware Active Learning for Text Summarization

Large Language Models (LLMs) have shown propensity to generate hallucinated outputs, i.e., texts that are factually incorrect or unsupported. Existing methods for alleviating hallucinations typically require costly human annotations to identify and correct hallucinations in LLM outputs. Moreover, most of these methods focus on a specific type of hallucination, e.g., entity or token errors, which limits their effectiveness in addressing various types of hallucinations exhibited in LLM outputs. To our best knowledge, in this paper we propose the first active learning framework to alleviate LLM hallucinations, reducing costly human annotations of hallucination needed. By measuring fine-grained hallucinations from errors in semantic frame, discourse and content verifiability in text summarization, we propose HAllucination Diversity-Aware Sampling (HADAS) to select diverse hallucinations for annotations in active learning for LLM finetuning. Extensive experiments on three datasets and different backbone models demonstrate advantages of our method in effectively and efficiently mitigating LLM hallucinations.

Energy Model-based Accurate Shapley Value Estimation for Interpretable Deep Learning Predictive Modelling

As a favorable tool for explainable artificial intelligence (XAI), Shapley value has been widely used to interpret deep learning based predictive models. However, accurate and efficient estimation of Shapley value is a difficult task since the computation load grows exponentially with the increase of input features. Most existing accelerated Shapley value estimation methods have to compromise on estimation accuracy with efficiency. In this article, we present EmSHAP(Energy model-based Shapley value estimation), which can effectively approximate the expectation of Shapley contribution function/deep learning model under arbitrary subset of features given the rest. In order to determine the proposal conditional distribution in the energy model, a gated recurrent unit(GRU) is introduced by mapping the input features onto a hidden space, so that the impact of input feature orderings can be eliminated. In addition, a dynamic masking scheme is proposed to improve the generalization ability. It is proved in Theorems 1, 2 and 3 that EmSHAP achieves tighter error bound than state-of-the-art methods like KernelSHAP and VAEAC, leading to higher estimation accuracy. Finally, case studies on a medical application and an industrial application show that the proposed Shapley value-based explainable framework exhibits enhanced estimation accuracy without compromise on efficiency.

Which LLM to Play? Convergence-Aware Online Model Selection with Time-Increasing Bandits

Web-based applications such as chatbots, search engines and news recommendations continue to grow in scale and complexity with the recent surge in the adoption of LLMs. Online model selection has thus garnered increasing attention due to the need to choose the best model among a diverse set while balancing task reward and exploration cost. Organizations faces decisions like whether to employ a costly API-based LLM or a locally finetuned small LLM, weighing cost against performance. Traditional selection methods often evaluate every candidate model before choosing one, which are becoming impractical given the rising costs of training and finetuning LLMs. Moreover, it is undesirable to allocate excessive resources towards exploring poor-performing models. While some recent works leverage online bandit algorithm to manage such exploration-exploitation trade-off in model selection, they tend to overlook the increasing-then-converging trend in model performances as the model is iteratively finetuned, leading to less accurate predictions and suboptimal model selections. In this paper, we propose a time-increasing bandit algorithm TI-UCB, which effectively predicts the increase of model performances due to finetuning and efficiently balances exploration and exploitation in model selection. To further capture the converging points of models, we develop a change detection mechanism by comparing consecutive increase predictions. We theoretically prove that our algorithm achieves a logarithmic regret upper bound in a typical increasing bandit setting, which implies a fast convergence rate. The advantage of our method is also empirically validated through extensive experiments on classification model selection and online selection of LLMs. Our results highlight the importance of utilizing increasing-then-converging pattern for more efficient and economic model selection in the deployment of LLMs.

RetailSynth: Synthetic Data Generation for Retail AI Systems Evaluation

Significant research effort has been devoted in recent years to developing personalized pricing, promotions, and product recommendation algorithms that can leverage rich customer data to learn and earn. Systematic benchmarking and evaluation of these causal learning systems remains a critical challenge, due to the lack of suitable datasets and simulation environments. In this work, we propose a multi-stage model for simulating customer shopping behavior that captures important sources of heterogeneity, including price sensitivity and past experiences. We embedded this model into a working simulation environment -- RetailSynth. RetailSynth was carefully calibrated on publicly available grocery data to create realistic synthetic shopping transactions. Multiple pricing policies were implemented within the simulator and analyzed for impact on revenue, category penetration, and customer retention. Applied researchers can use RetailSynth to validate causal demand models for multi-category retail and to incorporate realistic price sensitivity into emerging benchmarking suites for personalized pricing, promotions, and product recommendations.

M2HGCL: Multi-Scale Meta-Path Integrated Heterogeneous Graph Contrastive Learning

Inspired by the successful application of contrastive learning on graphs, researchers attempt to impose graph contrastive learning approaches on heterogeneous information networks. Orthogonal to homogeneous graphs, the types of nodes and edges in heterogeneous graphs are diverse so that specialized graph contrastive learning methods are required. Most existing methods for heterogeneous graph contrastive learning are implemented by transforming heterogeneous graphs into homogeneous graphs, which may lead to ramifications that the valuable information carried by non-target nodes is undermined thereby exacerbating the performance of contrastive learning models. Additionally, current heterogeneous graph contrastive learning methods are mainly based on initial meta-paths given by the dataset, yet according to our deep-going exploration, we derive empirical conclusions: only initial meta-paths cannot contain sufficiently discriminative information; and various types of meta-paths can effectively promote the performance of heterogeneous graph contrastive learning methods. To this end, we propose a new multi-scale meta-path integrated heterogeneous graph contrastive learning (M2HGCL) model, which discards the conventional heterogeneity-homogeneity transformation and performs the graph contrastive learning in a joint manner. Specifically, we expand the meta-paths and jointly aggregate the direct neighbor information, the initial meta-path neighbor information and the expanded meta-path neighbor information to sufficiently capture discriminative information. A specific positive sampling strategy is further imposed to remedy the intrinsic deficiency of contrastive learning, i.e., the hard negative sample sampling issue. Through extensive experiments on three real-world datasets, we demonstrate that M2HGCL outperforms the current state-of-the-art baseline models.

Contrastive Bootstrapping for Label Refinement

Traditional text classification typically categorizes texts into pre-defined coarse-grained classes, from which the produced models cannot handle the real-world scenario where finer categories emerge periodically for accurate services. In this work, we investigate the setting where fine-grained classification is done only using the annotation of coarse-grained categories and the coarse-to-fine mapping. We propose a lightweight contrastive clustering-based bootstrapping method to iteratively refine the labels of passages. During clustering, it pulls away negative passage-prototype pairs under the guidance of the mapping from both global and local perspectives. Experiments on NYT and 20News show that our method outperforms the state-of-the-art methods by a large margin.

The Closeness of In-Context Learning and Weight Shifting for Softmax Regression

Large language models (LLMs) are known for their exceptional performance in natural language processing, making them highly effective in many human life-related or even job-related tasks. The attention mechanism in the Transformer architecture is a critical component of LLMs, as it allows the model to selectively focus on specific input parts. The softmax unit, which is a key part of the attention mechanism, normalizes the attention scores. Hence, the performance of LLMs in various NLP tasks depends significantly on the crucial role played by the attention mechanism with the softmax unit. In-context learning, as one of the celebrated abilities of recent LLMs, is an important concept in querying LLMs such as ChatGPT. Without further parameter updates, Transformers can learn to predict based on few in-context examples. However, the reason why Transformers becomes in-context learners is not well understood. Recently, several works [ASA+22,GTLV22,ONR+22] have studied the in-context learning from a mathematical perspective based on a linear regression formulation $\min_x\| Ax - b \|_2$, which show Transformers' capability of learning linear functions in context. In this work, we study the in-context learning based on a softmax regression formulation $\min_{x} \| \langle \exp(Ax), {\bf 1}_n \rangle^{-1} \exp(Ax) - b \|_2$ of Transformer's attention mechanism. We show the upper bounds of the data transformations induced by a single self-attention layer and by gradient-descent on a $\ell_2$ regression loss for softmax prediction function, which imply that when training self-attention-only Transformers for fundamental regression tasks, the models learned by gradient-descent and Transformers show great similarity.

DocRED-FE: A Document-Level Fine-Grained Entity And Relation Extraction Dataset

Joint entity and relation extraction (JERE) is one of the most important tasks in information extraction. However, most existing works focus on sentence-level coarse-grained JERE, which have limitations in real-world scenarios. In this paper, we construct a large-scale document-level fine-grained JERE dataset DocRED-FE, which improves DocRED with Fine-Grained Entity Type. Specifically, we redesign a hierarchical entity type schema including 11 coarse-grained types and 119 fine-grained types, and then re-annotate DocRED manually according to this schema. Through comprehensive experiments we find that: (1) DocRED-FE is challenging to existing JERE models; (2) Our fine-grained entity types promote relation classification. We make DocRED-FE with instruction and the code for our baselines publicly available at https://github.com/PKU-TANGENT/DOCRED-FE.

Causal Analysis of Syntactic Agreement Neurons in Multilingual Language Models

Structural probing work has found evidence for latent syntactic information in pre-trained language models. However, much of this analysis has focused on monolingual models, and analyses of multilingual models have employed correlational methods that are confounded by the choice of probing tasks. In this study, we causally probe multilingual language models (XGLM and multilingual BERT) as well as monolingual BERT-based models across various languages; we do this by performing counterfactual perturbations on neuron activations and observing the effect on models' subject-verb agreement probabilities. We observe where in the model and to what extent syntactic agreement is encoded in each language. We find significant neuron overlap across languages in autoregressive multilingual language models, but not masked language models. We also find two distinct layer-wise effect patterns and two distinct sets of neurons used for syntactic agreement, depending on whether the subject and verb are separated by other tokens. Finally, we find that behavioral analyses of language models are likely underestimating how sensitive masked language models are to syntactic information.

Low Resource Style Transfer via Domain Adaptive Meta Learning

Text style transfer (TST) without parallel data has achieved some practical success. However, most of the existing unsupervised text style transfer methods suffer from (i) requiring massive amounts of non-parallel data to guide transferring different text styles. (ii) colossal performance degradation when fine-tuning the model in new domains. In this work, we propose DAML-ATM (Domain Adaptive Meta-Learning with Adversarial Transfer Model), which consists of two parts: DAML and ATM. DAML is a domain adaptive meta-learning approach to learn general knowledge in multiple heterogeneous source domains, capable of adapting to new unseen domains with a small amount of data. Moreover, we propose a new unsupervised TST approach Adversarial Transfer Model (ATM), composed of a sequence-to-sequence pre-trained language model and uses adversarial style training for better content preservation and style transfer. Results on multi-domain datasets demonstrate that our approach generalizes well on unseen low-resource domains, achieving state-of-the-art results against ten strong baselines.