Large Language Models (LLMs) are so powerful that they sometimes learn correlations between labels and features that are irrelevant to the task, leading to poor generalization on out-of-distribution data. We propose explanation-based finetuning as a novel and general approach to mitigate LLMs' reliance on spurious correlations. Unlike standard finetuning where the model only predicts the answer given the input, we finetune the model to additionally generate a free-text explanation supporting its answer. To evaluate our method, we finetune the model on artificially constructed training sets containing different types of spurious cues, and test it on a test set without these cues. Compared to standard finetuning, our method makes models remarkably more robust against spurious cues in terms of accuracy drop across four classification tasks: ComVE (+1.2), CREAK (+9.1), e-SNLI (+15.4), and SBIC (+6.5). Moreover, our method works equally well with explanations generated by the model, implying its applicability to more datasets without human-written explanations.
Dungeons & Dragons (D&D) is a tabletop roleplaying game with complex natural language interactions between players and hidden state information. Recent work has shown that large language models (LLMs) that have access to state information can generate higher quality game turns than LLMs that use dialog history alone. However, previous work used game state information that was heuristically created and was not a true gold standard game state. We present FIREBALL, a large dataset containing nearly 25,000 unique sessions from real D&D gameplay on Discord with true game state info. We recorded game play sessions of players who used the Avrae bot, which was developed to aid people in playing D&D online, capturing language, game commands and underlying game state information. We demonstrate that FIREBALL can improve natural language generation (NLG) by using Avrae state information, improving both automated metrics and human judgments of quality. Additionally, we show that LLMs can generate executable Avrae commands, particularly after finetuning.
Recent work has shown that prompting language models with code-like representations of natural language leads to performance improvements on structured reasoning tasks. However, such tasks comprise only a small subset of all natural language tasks. In our work, we seek to answer whether or not code-prompting is the preferred way of interacting with language models in general. We compare code and text prompts across three popular GPT models (davinci, code-davinci-002, and text-davinci-002) on a broader selection of tasks (e.g., QA, sentiment, summarization) and find that with few exceptions, code prompts do not consistently outperform text prompts. Furthermore, we show that the style of code prompt has a large effect on performance for some but not all tasks and that fine-tuning on text instructions leads to better relative performance of code prompts.
Existing question answering (QA) systems owe much of their success to large, high-quality training data. Such annotation efforts are costly, and the difficulty compounds in the cross-lingual setting. Therefore, prior cross-lingual QA work has focused on releasing evaluation datasets, and then applying zero-shot methods as baselines. In this work, we propose a synthetic data generation method for cross-lingual QA which leverages indirect supervision from existing parallel corpora. Our method termed PAXQA ({P}rojecting {a}nnotations for cross-lingual ({x}) QA) decomposes cross-lingual QA into two stages. In the first stage, we apply a question generation (QG) model to the English side. In the second stage, we apply annotation projection to translate both the questions and answers. To better translate questions, we propose a novel use of lexically-constrained machine translation, in which constrained entities are extracted from the parallel bitexts. We release cross-lingual QA datasets across 4 languages, totaling 662K QA examples. We then show that extractive QA models fine-tuned on these datasets outperform both zero-shot and prior synthetic data generation models, showing the sufficient quality of our generations. We find that the largest performance gains are for cross-lingual directions with non-English questions and English contexts. Ablation studies show that our dataset generation method is relatively robust to noise from automatic word alignments.
Schema induction builds a graph representation explaining how events unfold in a scenario. Existing approaches have been based on information retrieval (IR) and information extraction(IE), often with limited human curation. We demonstrate a human-in-the-loop schema induction system powered by GPT-3. We first describe the different modules of our system, including prompting to generate schematic elements, manual edit of those elements, and conversion of those into a schema graph. By qualitatively comparing our system to previous ones, we show that our system not only transfers to new domains more easily than previous approaches, but also reduces efforts of human curation thanks to our interactive interface.
While Chain-of-Thought (CoT) prompting boosts Language Models' (LM) performance on a gamut of complex reasoning tasks, the generated reasoning chain does not necessarily reflect how the model arrives at the answer (aka. faithfulness). We propose Faithful CoT, a faithful-by-construction framework that decomposes a reasoning task into two stages: Translation (Natural Language query $\rightarrow$ symbolic reasoning chain) and Problem Solving (reasoning chain $\rightarrow$ answer), using an LM and a deterministic solver respectively. We demonstrate the efficacy of our approach on 10 reasoning datasets from 4 diverse domains. It outperforms traditional CoT prompting on 9 out of the 10 datasets, with an average accuracy gain of 4.4 on Math Word Problems, 1.9 on Planning, 4.0 on Multi-hop Question Answering (QA), and 18.1 on Logical Inference, under greedy decoding. Together with self-consistency decoding, we achieve new state-of-the-art few-shot performance on 7 out of the 10 datasets, showing a strong synergy between faithfulness and accuracy.
Entities and events have long been regarded as the crux of machine reasoning. Specifically, procedural texts have received increasing attention due to the dynamic nature of involved entities and events. Existing work has exclusively focused on entity state tracking (e.g., the temperature of a pan) or counterfactual event reasoning (e.g., how likely am I to burn myself by touching the pan), while these two tasks are tightly intertwined. In this work, we propose CREPE, the first benchmark on causal reasoning about event plausibility based on entity states. We experiment with strong large language models and show that most models including GPT3 perform close to chance of .30 F1, lagging far behind the human performance of .87 F1. Inspired by the finding that structured representations such as programming languages benefits event reasoning as a prompt to code language models such as Codex, we creatively inject the causal relations between entities and events through intermediate variables and boost the performance to .67 to .72 F1. Our proposed event representation not only allows for knowledge injection, but also marks the first successful attempt of chain-of-thought reasoning with code language models.