Large language models (LLMs) are capable of generating multiple responses to a single prompt, yet little effort has been expended to help end-users or system designers make use of this capability. In this paper, we explore how to present many LLM responses at once. We design five features, which include both pre-existing and novel methods for computing similarities and differences across textual documents, as well as how to render their outputs. We report on a controlled user study (n=24) and eight case studies evaluating these features and how they support users in different tasks. We find that the features support a wide variety of sensemaking tasks and even make tasks previously considered to be too difficult by our participants now tractable. Finally, we present design guidelines to inform future explorations of new LLM interfaces.
While alignment algorithms are now commonly used to tune pre-trained language models towards a user's preferences, we lack explanations for the underlying mechanisms in which models become ``aligned'', thus making it difficult to explain phenomena like jailbreaks. In this work we study a popular algorithm, direct preference optimization (DPO), and the mechanisms by which it reduces toxicity. Namely, we first study how toxicity is represented and elicited in a pre-trained language model, GPT2-medium. We then apply DPO with a carefully crafted pairwise dataset to reduce toxicity. We examine how the resulting model averts toxic outputs, and find that capabilities learned from pre-training are not removed, but rather bypassed. We use this insight to demonstrate a simple method to un-align the model, reverting it back to its toxic behavior.
Relational databases play an important role in this Big Data era. However, it is challenging for non-experts to fully unleash the analytical power of relational databases, since they are not familiar with database languages such as SQL. Many techniques have been proposed to automatically generate SQL from natural language, but they suffer from two issues: (1) they still make many mistakes, particularly for complex queries, and (2) they do not provide a flexible way for non-expert users to validate and refine the incorrect queries. To address these issues, we introduce a new interaction mechanism that allows users directly edit a step-by-step explanation of an incorrect SQL to fix SQL errors. Experiments on the Spider benchmark show that our approach outperforms three SOTA approaches by at least 31.6% in terms of execution accuracy. A user study with 24 participants further shows that our approach helped users solve significantly more SQL tasks with less time and higher confidence, demonstrating its potential to expand access to databases, particularly for non-experts.
The standard task-oriented dialogue pipeline uses intent classification and slot-filling to interpret user utterances. While this approach can handle a wide range of queries, it does not extract the information needed to handle more complex queries that contain relationships between slots. We propose integration of relation extraction into this pipeline as an effective way to expand the capabilities of dialogue systems. We evaluate our approach by using an internal dataset with slot and relation annotations spanning three domains. Finally, we show how slot-filling annotation schemes can be simplified once the expressive power of relation annotations is available, reducing the number of slots while still capturing the user's intended meaning.
We use paraphrases as a unique source of data to analyze contextualized embeddings, with a particular focus on BERT. Because paraphrases naturally encode consistent word and phrase semantics, they provide a unique lens for investigating properties of embeddings. Using the Paraphrase Database's alignments, we study words within paraphrases as well as phrase representations. We find that contextual embeddings effectively handle polysemous words, but give synonyms surprisingly different representations in many cases. We confirm previous findings that BERT is sensitive to word order, but find slightly different patterns than prior work in terms of the level of contextualization across BERT's layers.
For each goal-oriented dialog task of interest, large amounts of data need to be collected for end-to-end learning of a neural dialog system. Collecting that data is a costly and time-consuming process. Instead, we show that we can use only a small amount of data, supplemented with data from a related dialog task. Naively learning from related data fails to improve performance as the related data can be inconsistent with the target task. We describe a meta-learning based method that selectively learns from the related dialog task data. Our approach leads to significant accuracy improvements in an example dialog task.
Many statistical models have high accuracy on test benchmarks, but are not explainable, struggle in low-resource scenarios, cannot be reused for multiple tasks, and cannot easily integrate domain expertise. These factors limit their use, particularly in settings such as mental health, where it is difficult to annotate datasets and model outputs have significant impact. We introduce a micromodel architecture to address these challenges. Our approach allows researchers to build interpretable representations that embed domain knowledge and provide explanations throughout the model's decision process. We demonstrate the idea on multiple mental health tasks: depression classification, PTSD classification, and suicidal risk assessment. Our systems consistently produce strong results, even in low-resource scenarios, and are more interpretable than alternative methods.
A growing number of people engage in online health forums, making it important to understand the quality of the advice they receive. In this paper, we explore the role of expertise in responses provided to help-seeking posts regarding mental health. We study the differences between (1) interactions with peers; and (2) interactions with self-identified mental health professionals. First, we show that a classifier can distinguish between these two groups, indicating that their language use does in fact differ. To understand this difference, we perform several analyses addressing engagement aspects, including whether their comments engage the support-seeker further as well as linguistic aspects, such as dominant language and linguistic style matching. Our work contributes toward the developing efforts of understanding how health experts engage with health information- and support-seekers in social networks. More broadly, it is a step toward a deeper understanding of the styles of interactions that cultivate supportive engagement in online communities.
Extensive work has argued in favour of paying crowd workers a wage that is at least equivalent to the U.S. federal minimum wage. Meanwhile, research on collecting high quality annotations suggests using a qualification that requires workers to have previously completed a certain number of tasks. If most requesters who pay fairly require workers to have completed a large number of tasks already then workers need to complete a substantial amount of poorly paid work before they can earn a fair wage. Through analysis of worker discussions and guidance for researchers, we estimate that workers spend approximately 2.25 months of full time effort on poorly paid tasks in order to get the qualifications needed for better paid tasks. We discuss alternatives to this qualification and conduct a study of the correlation between qualifications and work quality on two NLP tasks. We find that it is possible to reduce the burden on workers while still collecting high quality data.
Natural language processing methods have been applied in a variety of music studies, drawing the connection between music and language. In this paper, we expand those approaches by investigating \textit{chord embeddings}, which we apply in two case studies to address two key questions: (1) what musical information do chord embeddings capture?; and (2) how might musical applications benefit from them? In our analysis, we show that they capture similarities between chords that adhere to important relationships described in music theory. In the first case study, we demonstrate that using chord embeddings in a next chord prediction task yields predictions that more closely match those by experienced musicians. In the second case study, we show the potential benefits of using the representations in tasks related to musical stylometrics.