Despite improvements in performances on different natural language generation tasks, deep neural models are prone to hallucinating facts that are incorrect or nonexistent. Different hypotheses are proposed and examined separately for different tasks, but no systematic explanations are available across these tasks. In this study, we draw connections between hallucinations and predictive uncertainty in conditional language generation. We investigate their relationship in both image captioning and data-to-text generation and propose a simple extension to beam search to reduce hallucination. Our analysis shows that higher predictive uncertainty corresponds to a higher chance of hallucination. Epistemic uncertainty is more indicative of hallucination than aleatoric or total uncertainties. It helps to achieve better results of trading performance in standard metric for less hallucination with the proposed beam search variant.
Creating presentation materials requires complex multimodal reasoning skills to summarize key concepts and arrange them in a logical and visually pleasing manner. Can machines learn to emulate this laborious process? We present a novel task and approach for document-to-slide generation. Solving this involves document summarization, image and text retrieval, slide structure and layout prediction to arrange key elements in a form suitable for presentation. We propose a hierarchical sequence-to-sequence approach to tackle our task in an end-to-end manner. Our approach exploits the inherent structures within documents and slides and incorporates paraphrasing and layout prediction modules to generate slides. To help accelerate research in this domain, we release a dataset about 6K paired documents and slide decks used in our experiments. We show that our approach outperforms strong baselines and produces slides with rich content and aligned imagery.
Responsible development of technology involves applications being inclusive of the diverse set of users they hope to support. An important part of this is understanding the many ways to refer to a person and being able to fluently change between the different forms as needed. We perform a case study on the singular they, a common way to promote gender inclusion in English. We define a re-writing task, create an evaluation benchmark, and show how a model can be trained to produce gender-neutral English with <1% word error rate with no human-labeled data. We discuss the practical applications and ethical considerations of the task, providing direction for future work into inclusive natural language systems.
Recent advances in language and vision push forward the research of captioning a single image to describing visual differences between image pairs. Suppose there are two images, I_1 and I_2, and the task is to generate a description W_{1,2} comparing them, existing methods directly model { I_1, I_2 } -> W_{1,2} mapping without the semantic understanding of individuals. In this paper, we introduce a Learning-to-Compare (L2C) model, which learns to understand the semantic structures of these two images and compare them while learning to describe each one. We demonstrate that L2C benefits from a comparison between explicit semantic representations and single-image captions, and generalizes better on the new testing image pairs. It outperforms the baseline on both automatic evaluation and human evaluation for the Birds-to-Words dataset.
Broader transparency in descriptions of and communication regarding AI systems is widely considered desirable. This is particularly the case in discussions of fairness and accountability in systems exposed to the general public. However, previous work has suggested that a trade-off exists between greater system transparency and user confusion, where `too much information' clouds a reader's understanding of what a system description means. Unfortunately, transparency is a nebulous concept, difficult to both define and quantify. In this work we address these two issues by proposing a framework for quantifying transparency in system descriptions and apply it to analyze the trade-off between transparency and end-user confusion using NLP conference abstracts.
With the growing adoption of text generation models in today's society, users are increasingly exposed to machine-generated text. This in turn can leave users vulnerable to the generation of harmful information such as conspiracy theories. While the propagation of conspiracy theories through social media has been studied, previous work has not evaluated their diffusion through text generation. In this work, we investigate the propensity for language models to generate conspiracy theory text. Our study focuses on testing these models for the elicitation of conspiracy theories and comparing these generations to human-written theories from Reddit. We also introduce a new dataset consisting of conspiracy theory topics, machine-generated conspiracy theories, and human-written conspiracy theories. Our experiments show that many well-known conspiracy theory topics are deeply rooted in the pre-trained language models, and can become more prevalent through different model settings.
The growth of social media has encouraged the written use of African American Vernacular English (AAVE), which has traditionally been used only in oral contexts. However, NLP models have historically been developed using dominant English varieties, such as Standard American English (SAE), due to text corpora availability. We investigate the performance of GPT-2 on AAVE text by creating a dataset of intent-equivalent parallel AAVE/SAE tweet pairs, thereby isolating syntactic structure and AAVE- or SAE-specific language for each pair. We evaluate each sample and its GPT-2 generated text with pretrained sentiment classifiers and find that while AAVE text results in more classifications of negative sentiment than SAE, the use of GPT-2 generally increases occurrences of positive sentiment for both. Additionally, we conduct human evaluation of AAVE and SAE text generated with GPT-2 to compare contextual rigor and overall quality.
Obtaining training data for Multi-hop Question Answering (QA) is extremely time-consuming and resource-intensive. To address this, we propose the problem of \textit{unsupervised} multi-hop QA, assuming that no human-labeled multi-hop question-answer pairs are available. We propose MQA-QG, an unsupervised question answering framework that can generate human-like multi-hop training pairs from both homogeneous and heterogeneous data sources. Our model generates questions by first selecting or generating relevant information from each data source and then integrating the multiple information to form a multi-hop question. We find that we can train a competent multi-hop QA model with only generated data. The F1 gap between the unsupervised and fully-supervised models is less than 20 in both the HotpotQA and the HybridQA dataset. Further experiments reveal that an unsupervised pretraining with the QA data generated by our model would greatly reduce the demand for human-annotated training data for multi-hop QA.
Vision-and-Language Navigation (VLN) is a natural language grounding task where an agent learns to follow language instructions and navigate to specified destinations in real-world environments. A key challenge is to recognize and stop at the correct location, especially for complicated outdoor environments. Existing methods treat the STOP action equally as other actions, which results in undesirable behaviors that the agent often fails to stop at the destination even though it might be on the right path. Therefore, we propose Learning to Stop (L2Stop), a simple yet effective policy module that differentiates STOP and other actions. Our approach achieves the new state of the art on a challenging urban VLN dataset Touchdown, outperforming the baseline by 6.89% (absolute improvement) on Success weighted by Edit Distance (SED).