Large language models (LLMs) are empowering decision-making in several applications, including tool or API usage and answering multiple-choice questions (MCQs). However, incorrect outputs pose significant risks in high-stakes domains like healthcare and finance. To quantify LLM uncertainty and thereby mitigate these risks, recent works employ conformal prediction (CP), a model- and distribution-agnostic framework that uses LLM outputs to generate a prediction set containing the true answer with high probability. Leveraging CP, we propose conformal revision of questions (CROQ), which revises the question by narrowing down the available choices to those in the prediction set and asking the LLM the revised question. We expect LLMs to be more accurate on revised questions with fewer choices. Furthermore, we expect CROQ to be effective when the prediction sets from CP are small. Commonly used logit scores often lead to large sets, diminishing CROQ’s effectiveness. To overcome this, we propose CP-OPT, an optimization framework to learn scores that minimize set sizes while maintaining coverage. Our extensive experiments on MMLU, ToolAlpaca, and TruthfulQA datasets with multiple LLMs show that CROQ improves accuracy over the standard inference, with more pronounced gains when paired with CP-OPT.

Dongsheng Wang, Natraj Raman, Mathieu Sibue, Zhiqiang Ma, Petr Babkin, Simerjot Kaur, Yulong Pei, Armineh Nourbakhsh, Xiaomo Liu. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Counterfactual explanations utilize feature perturbations to analyze the outcome of an original decision and recommend an actionable recourse. We argue that it is beneficial to provide several alternative explanations rather than a single point solution and propose a probabilistic paradigm to estimate a diverse set of counterfactuals. Specifically, we treat the perturbations as random variables endowed with prior distribution functions. This allows sampling multiple counterfactuals from the posterior density, with the added benefit of incorporating inductive biases, preserving domain specific constraints and quantifying uncertainty in estimates. More importantly, we leverage Bayesian hierarchical modeling to share information across different subgroups of a population, which can both improve robustness and measure fairness. A gradient based sampler with superior convergence characteristics efficiently computes the posterior samples. Experiments across several datasets demonstrate that the counterfactuals estimated using our approach are valid, sparse, diverse and feasible.

Graph-structured diagrams, such as enterprise ownership charts or management hierarchies, are a challenging medium for deep learning models as they not only require the capacity to model language and spatial relations but also the topology of links between entities and the varying semantics of what those links represent. Devising Question Answering models that automatically process and understand such diagrams have vast applications to many enterprise domains, and can move the state-of-the-art on multimodal document understanding to a new frontier. Curating real-world datasets to train these models can be difficult, due to scarcity and confidentiality of the documents where such diagrams are included. Recently released synthetic datasets are often prone to repetitive structures that can be memorized or tackled using heuristics. In this paper, we present a collection of 10,000 synthetic graphs that faithfully reflect properties of real graphs in four business domains, and are realistically rendered within a PDF document with varying styles and layouts. In addition, we have generated over 130,000 question instances that target complex graphical relationships specific to each domain. We hope this challenge will encourage the development of models capable of robust reasoning about graph structured images, which are ubiquitous in numerous sectors in business and across scientific disciplines.

Raj Shah, Kunal Chawla, Dheeraj Eidnani, Agam Shah, Wendi Du, Sudheer Chava, Natraj Raman, Charese Smiley, Jiaao Chen, Diyi Yang. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022.

Retrieving relevant documents from a corpus is typically based on the semantic similarity between the document content and query text. The inclusion of structural relationship between documents can benefit the retrieval mechanism by addressing semantic gaps. However, incorporating these relationships requires tractable mechanisms that balance structure with semantics and take advantage of the prevalent pre-train/fine-tune paradigm. We propose here a holistic approach to learning document representations by integrating intra-document content with inter-document relations. Our deep metric learning solution analyzes the complex neighborhood structure in the relationship network to efficiently sample similar/dissimilar document pairs and defines a novel quintuplet loss function that simultaneously encourages document pairs that are semantically relevant to be closer and structurally unrelated to be far apart in the representation space. Furthermore, the separation margins between the documents are varied flexibly to encode the heterogeneity in relationship strengths. The model is fully fine-tunable and natively supports query projection during inference. We demonstrate that it outperforms competing methods on multiple datasets for document retrieval tasks.

The automatic extraction of breaking news events from natural language text is a valuable capability for decision support systems. Traditional systems tend to focus on extracting events from a single media source and often ignore cross-media references. Here, we describe a large-scale automated system for extracting natural disasters and critical events from both newswire text and social media. We outline a comprehensive architecture that can identify, categorize and summarize seven different event types - namely floods, storms, fires, armed conflict, terrorism, infrastructure breakdown, and labour unavailability. The system comprises fourteen modules and is equipped with a novel coreference mechanism, capable of linking events extracted from the two complementary data sources. Additionally, the system is easily extensible to accommodate new event types. Our experimental evaluation demonstrates the effectiveness of the system.

We present a contrastive study of document-level event classification of a range of seven different event types, namely floods, storms, fires, armed conflict, terrorism, infrastructure breakdown and labour unavailability from English-language news. Our study compares different supervised classification approaches, namely Support Vector Machine (SVM), Random Forest (RF), Convolutional Neural Network (CNN) and Hierarchical Attention Network (HAN). While past systems for Topic Detection and Tracking (TDT) and event extraction have proposed different machine learning models, to date SVMs, RFs, CNNs and HANs have not been compared on this task. Our classifiers are also informed by word embeddings trained on large amounts of high-quality agency news, which leads to improvements compared to the use of pre-trained embedding vectors. We report a detailed quantitative error analysis.

Conditional Random Fields (CRF), a structured prediction method, combines probabilistic graphical models and discriminative classification techniques in order to predict class labels in sequence recognition problems. Its extension the Hidden Conditional Random Fields (HCRF) uses hidden state variables in order to capture intermediate structures. The number of hidden states in an HCRF must be specified a priori. This number is often not known in advance. A non-parametric extension to the HCRF, with the number of hidden states automatically inferred from data, is proposed here. This is a significant advantage over the classical HCRF since it avoids ad hoc model selection procedures. Further, the training and inference procedure is fully Bayesian eliminating the over fitting problem associated with frequentist methods. In particular, our construction is based on scale mixtures of Gaussians as priors over the HCRF parameters and makes use of Hierarchical Dirichlet Process (HDP) and Laplace distribution. The proposed inference procedure uses elliptical slice sampling, a Markov Chain Monte Carlo (MCMC) method, in order to sample optimal and sparse posterior HCRF parameters. The above technique is applied for classifying human actions that occur in depth image sequences - a challenging computer vision problem. Experiments with real world video datasets confirm the efficacy of our classification approach.

We classify human actions occurring in videos, using the skeletal joint positions extracted from a depth image sequence as features. Each action class is represented by a non-parametric Hidden Markov Model (NP-HMM) and the model parameters are learnt in a discriminative way. Specifically, we use a Bayesian framework based on Hierarchical Dirichlet Process (HDP) to automatically infer the cardinality of hidden states and formulate a discriminative function based on distance between Gaussian distributions to improve classification performance. We use elliptical slice sampling to efficiently sample parameters from the complex posterior distribution induced by our discriminative likelihood function. We illustrate our classification results for action class models trained using this technique.