One of the ultimate quests of question answering (QA) is to deploy a system that can answer any type of question from the users, and refrain from answering when it does not know the answer. While recent advancements in scaling large language models (LLMs) brought significant improvements on various QA datasets, it remains difficult for a single model to generalize across question types that require distinct reasoning abilities. In this paper, we first provide empirical evidence that state-of-the-art LLMs such as Codex suffer from poor generalizability on question types beyond those seen in the prompt. To address this, we propose a Mixture-of-Prompt-Experts (MOPE) system that ensembles multiple specialized LLMs. We first implement each specialized model based on the same backbone model (Codex) but with prompts optimized for different reasoning categories including factual, multihop, mathematical, and commonsense reasoning. By strategically selecting the best specialized model for each given question, our MOPE system significantly outperforms any single specialized model on a collection of 12 QA datasets from four reasoning types. Moreover, the attribution and agreement among specialized expert models offer greater interpretability, allowing for better selective question answering. Our human study further confirms that presenting the expert predictions and answer selection process helps annotators more accurately decide when to trust the system's output. We release all code and data to facilitate future work.
For humans and computers, the first step in answering an open-domain question is retrieving a set of relevant documents from a large corpus. However, the strategies that computers use fundamentally differ from those of humans. To better understand these differences, we design a gamified interface for data collection -- Cheater's Bowl -- where a human answers complex questions with access to both traditional and modern search tools. We collect a dataset of human search sessions, analyze human search strategies, and compare them to state-of-the-art multi-hop QA models. Humans query logically, apply dynamic search chains, and use world knowledge to boost searching. We demonstrate how human queries can improve the accuracy of existing systems and propose improving the future design of QA models.
Large language models (LLMs) show impressive abilities via few-shot prompting. Commercialized APIs such as OpenAI GPT-3 further increase their use in real-world language applications. However, existing research focuses on models' accuracy on standard benchmarks and largely ignores their reliability, which is crucial for avoiding catastrophic real-world harms. While reliability is a broad and vaguely defined term, this work decomposes reliability into four facets: generalizability, fairness, calibration, and factuality. We establish simple and effective prompts to demonstrate GPT-3's reliability in these four aspects: 1) generalize out-of-domain, 2) balance demographic distribution to reduce social biases, 3) calibrate language model probabilities, and 4) update the LLM's knowledge. We find that by employing appropriate prompts, GPT-3 outperforms smaller-scale supervised models by large margins on all these facets. We release all processed datasets, evaluation scripts, and model predictions to facilitate future analysis. Our findings not only shed new insights on the reliability of prompting LLMs, but more importantly, our prompting strategies can help practitioners more reliably use large language models like GPT-3.
Question Answering (QA) systems require a large amount of annotated data which is costly and time-consuming to gather. Converting datasets of existing QA benchmarks are challenging due to different formats and complexities. To address these issues, we propose an algorithm to automatically generate shorter questions resembling day-to-day human communication in the Natural Questions (NQ) dataset from longer trivia questions in Quizbowl (QB) dataset by leveraging conversion in style among the datasets. This provides an automated way to generate more data for our QA systems. To ensure quality as well as quantity of data, we detect and remove ill-formed questions using a neural classifier. We demonstrate that in a low resource setting, using the generated data improves the QA performance over the baseline system on both NQ and QB data. Our algorithm improves the scalability of training data while maintaining quality of data for QA systems.
Model calibration aims to adjust (calibrate) models' confidence so that they match expected accuracy. We argue that the traditional evaluation of calibration (expected calibration error; ECE) does not reflect usefulness of the model confidence. For example, after conventional temperature scaling, confidence scores become similar for all predictions, which makes it hard for users to distinguish correct predictions from wrong ones, even though it achieves low ECE. Building on those observations, we propose a new calibration metric, MacroCE, that better captures whether the model assigns low confidence to wrong predictions and high confidence to correct predictions. We examine various conventional calibration methods including temperature scaling, feature-based classifier, neural answer reranking, and label smoothing, all of which do not bring significant gains under our new MacroCE metric. Towards more effective calibration, we propose a new calibration method based on the model's prediction consistency along the training trajectory. This new method, which we name as consistency calibration, shows promise for better calibration.
This paper develops automatic song translation (AST) for tonal languages and addresses the unique challenge of aligning words' tones with melody of a song in addition to conveying the original meaning. We propose three criteria for effective AST -- preserving meaning, singability and intelligibility -- and design metrics for these criteria. We develop a new benchmark for English--Mandarin song translation and develop an unsupervised AST system, Guided AliGnment for Automatic Song Translation (GagaST), which combines pre-training with three decoding constraints. Both automatic and human evaluations show GagaST successfully balances semantics and singability.
Pretrained multilingual models enable zero-shot learning even for unseen languages, and that performance can be further improved via adaptation prior to finetuning. However, it is unclear how the number of pretraining languages influences a model's zero-shot learning for languages unseen during pretraining. To fill this gap, we ask the following research questions: (1) How does the number of pretraining languages influence zero-shot performance on unseen target languages? (2) Does the answer to that question change with model adaptation? (3) Do the findings for our first question change if the languages used for pretraining are all related? Our experiments on pretraining with related languages indicate that choosing a diverse set of languages is crucial. Without model adaptation, surprisingly, increasing the number of pretraining languages yields better results up to adding related languages, after which performance plateaus. In contrast, with model adaptation via continued pretraining, pretraining on a larger number of languages often gives further improvement, suggesting that model adaptation is crucial to exploit additional pretraining languages.
Open-domain question answering answers a question based on evidence retrieved from a large corpus. State-of-the-art neural approaches require intermediate evidence annotations for training. However, such intermediate annotations are expensive, and methods that rely on them cannot transfer to the more common setting, where only question-answer pairs are available. This paper investigates whether models can learn to find evidence from a large corpus, with only distant supervision from answer labels for model training, thereby generating no additional annotation cost. We introduce a novel approach (DistDR) that iteratively improves over a weak retriever by alternately finding evidence from the up-to-date model and encouraging the model to learn the most likely evidence. Without using any evidence labels, DistDR is on par with fully-supervised state-of-the-art methods on both multi-hop and single-hop QA benchmarks. Our analysis confirms that DistDR finds more accurate evidence over iterations, which leads to model improvements.
A flaw in QA evaluation is that annotations often only provide one gold answer. Thus, model predictions semantically equivalent to the answer but superficially different are considered incorrect. This work explores mining alias entities from knowledge bases and using them as additional gold answers (i.e., equivalent answers). We incorporate answers for two settings: evaluation with additional answers and model training with equivalent answers. We analyse three QA benchmarks: Natural Questions, TriviaQA, and SQuAD. Answer expansion increases the exact match score on all datasets for evaluation, while incorporating it helps model training over real-world datasets. We ensure the additional answers are valid through a human post hoc evaluation.
Topic model evaluation, like evaluation of other unsupervised methods, can be contentious. However, the field has coalesced around automated estimates of topic coherence, which rely on the frequency of word co-occurrences in a reference corpus. Recent models relying on neural components surpass classical topic models according to these metrics. At the same time, unlike classical models, the practice of neural topic model evaluation suffers from a validation gap: automatic coherence for neural models has not been validated using human experimentation. In addition, as we show via a meta-analysis of topic modeling literature, there is a substantial standardization gap in the use of automated topic modeling benchmarks. We address both the standardization gap and the validation gap. Using two of the most widely used topic model evaluation datasets, we assess a dominant classical model and two state-of-the-art neural models in a systematic, clearly documented, reproducible way. We use automatic coherence along with the two most widely accepted human judgment tasks, namely, topic rating and word intrusion. Automated evaluation will declare one model significantly different from another when corresponding human evaluations do not, calling into question the validity of fully automatic evaluations independent of human judgments.