We develop an end-to-end model for learning to follow language instructions with compositional policies. Our model combines large language models with pretrained compositional value functions to generate policies for goal-reaching tasks specified in natural language. We evaluate our method in the BabyAI environment and demonstrate compositional generalization to novel combinations of task attributes. Notably our method generalizes to held-out combinations of attributes, and in some cases can accomplish those tasks with no additional learning samples.
ML ID: 416
Extracting the common sense knowledge present in Large Language Models (LLMs) offers a path to designing intelligent, embodied agents. Related works have queried LLMs with a wide-range of contextual information, such as goals, sensor observations and scene descriptions, to generate high-level action plans for specific tasks; however these approaches often involve human intervention or additional machinery to enable sensor-motor interactions. In this work, we propose a prompting-based strategy for extracting executable plans from an LLM, which leverages a novel and readily-accessible source of information: precondition errors. Our approach assumes that actions are only afforded execution in certain contexts, i.e., implicit preconditions must be met for an action to execute (e.g., a door must be unlocked to open it), and that the embodied agent has the ability to determine if the action is/is not executable in the current context (e.g., detect if a precondition error is present). When an agent is unable to execute an action, our approach re-prompts the LLM with precondition error information to extract an executable corrective action to achieve the intended goal in the current context. We evaluate our approach in the VirtualHome simulation environment on 88 different tasks and 7 scenes. We evaluate different prompt templates and compare to methods that naively resample actions from the LLM. Our approach, using precondition errors, improves executability and semantic correctness of plans, while also reducing the number of re-prompts required when querying actions.
ML ID: 415
Automatically fixing software bugs is a challenging task. While recent work showed that natural language context is useful in guiding bug-fixing models, the approach required prompting developers to provide this context, which was simulated through commit messages written after the bug-fixing code changes were made. We instead propose using bug report discussions, which are available before the task is performed and are also naturally occurring, avoiding the need for any additional information from developers. For this, we augment standard bug-fixing datasets with bug report discussions. Using these newly compiled datasets, we demonstrate that various forms of natural language context derived from such discussions can aid bug-fixing, even leading to improved performance over using commit messages corresponding to the oracle bug-fixing commits.
ML ID: 414
Visual question answering (VQA) has recently emerged as a challenging multi-modal task and has gained popularity. The goal is to answer questions that query information associated with the visual content in the given image. Since the required information could be from both inside and outside the image, common types of visual features, such as object and attribute detection, fail to provide enough materials for answering the questions. External information, such as captions, explanations, encyclopedia articles, and commonsense databases, can help VQA systems comprehensively understand the image, reason following the right path, and access external facts. Specifically, they provide concise descriptions of the image, precise reasons for the correct answer, and factual knowledge beyond the image. In this dissertation, we present our work on generating image captions that are targeted to help answer a specific visual question. We use explanations to recognize the critical objects to prevent the VQA models from taking language prior shortcuts. We introduce an approach that generates textual explanations and utilizes them to determine which answer is mostly supported. At last, we explore retrieving and exploiting external knowledge beyond the visual content, which is indispensable, to help answer knowledge-based visual questions.
ML ID: 413
For the majority of the machine learning community, the expensive nature of collecting high-quality human-annotated data and the inability to efficiently finetune very large state-of-the-art pretrained models on limited compute are major bottlenecks for building models for new tasks. We propose a zero-shot simple approach for one such task, Video Moment Retrieval (VMR), that does not perform any additional finetuning and simply repurposes off-the-shelf models trained on other tasks. Our three-step approach consists of moment proposal, moment-query matching and postprocessing, all using only off-the-shelf models. On the QVHighlights (Lei et al., 2021) benchmark for VMR, we vastly improve performance of previous zero-shot approaches by at least 2.5x on all metrics and reduce the gap between zero-shot and state-of-the-art supervised by over 74%. Further, we also show that our zero-shot approach beats non-pretrained supervised models on the Recall metrics and comes very close on mAP metrics; and that it also performs better than the best pretrained supervised model on shorter moments. Finally, we ablate and analyze our results and propose interesting future directions.
ML ID: 412
Most Outside-Knowledge Visual Question Answering (OK-VQA) systems employ a two-stage framework that first retrieves external knowledge given the visual question and then predicts the answer based on the retrieved content. However, the retrieved knowledge is often inadequate. Retrievals are frequently too general and fail to cover specific knowledge needed to answer the question. Also, the naturally available supervision (whether the passage contains the correct answer) is weak and does not guarantee question relevancy. To address these issues, we propose an Entity-Focused Retrieval (EnFoRe) model that provides stronger supervision during training and recognizes question-relevant entities to help retrieve more specific knowledge. Experiments show that our EnFoRe model achieves superior retrieval performance on OK-VQA, the currently largest outside-knowledge VQA dataset. We also combine the retrieved knowledge with state-of-the-art VQA models, and achieve a new state-of-the-art performance on OK-VQA.
ML ID: 411
Building intelligent agents that can help humans accomplish everyday tasks, such as a personal robot at home or a robot in a work environment, is a long-standing goal of artificial intelligence. One of the requirements for such general-purpose agents is the ability to teach them new tasks or skills relatively easily. Common approaches to teaching agents new skills include reinforcement learning (RL) and imitation learning (IL). However, specifying the task to the learning agent, i.e. designing effective reward functions for reinforcement learning and providing demonstrations for imitation learning, are often cumbersome and time-consuming. Further, designing reward functions and providing a set of demonstrations that sufficiently disambiguates the desired task may not be particularly accessible for end users without a technical background.
In this dissertation, we explore using natural language as an auxiliary signal to aid task specification, which reduces the burden on the end user. To make reward design easier, we propose a novel framework that is used to generate language-based rewards in addition to the extrinsic rewards from the environment for faster policy training using RL. We show that using our framework, very simple extrinsic rewards along with a natural language description of the task are sufficient to teach new tasks to the learning agent. To ameliorate the problem of providing demonstrations, we propose a new setting that enables an agent to learn a new task without demonstrations in an IL setting, given a demonstration from a related task and a natural language description of the difference between the desired task and the demonstrated task. The techniques we develop for this setting would enable teaching multiple related tasks to learning agents by providing a small set of demonstrations and several natural language descriptions, thereby reducing the burden of providing demonstrations for each task.
The primary contributions of this dissertation include novel problem settings, benchmarks, and algorithms that allow using natural language as an auxiliary modality for task specification in RL and IL. We believe this dissertation will serve as a foundation for future research along these lines, to make progress toward having intelligent agents that can conveniently be taught new tasks by end users.
ML ID: 410
Software projects are continually evolving, as developers incorporate changes to refactor code, support new functionality, and fix bugs. To uphold software quality amidst constant changes and also facilitate prompt implementation of critical changes, it is desirable to have automated tools for supporting and driving software evolution. In this thesis, we explore tasks and data and design machine learning approaches which leverage natural language to serve this purpose.
When developers make code changes, they sometimes fail to update the accompanying natural language comments documenting various aspects of the code, which can lead to confusion and vulnerability to bugs. We present our work on alerting developers of inconsistent comments upon code changes and suggesting updates by learning to correlate comments and code.
When a bug is reported, developers engage in a dialogue to collaboratively understand it and ultimately resolve it. While the solution is likely formulated within the discussion, it is often buried in a large amount of text, making it difficult to comprehend, which delays its implementation through the necessary repository changes. To guide developers in more easily absorbing information relevant towards making these changes and consequently expedite bug resolution, we investigate generating a concise natural language description of the solution by synthesizing relevant content as it emerges in the discussion. We benchmark models for generating solution descriptions and design a classifier for determining when sufficient context for generating an informative description becomes available. We investigate approaches for real-time generation, entailing separately trained and jointly trained classification and generation models. Furthermore, we also study techniques for deriving natural language context from bug report discussions and generated solution descriptions to guide models in generating suggested bug-resolving code changes.
ML ID: 409
Answering questions in narratives about why events happened often requires commonsense knowledge external to the text. What aspects of this knowledge are available in large language models? What aspects can be made accessible via external commonsense resources? We study these questions in the context of answering questions in the TELLMEWHY dataset using COMET as a source of relevant commonsense relations. We analyze the effects of model size (T5 variants and GPT-3) along with methods of injecting knowledge (COMET) into these models. Results show that the largest models, as expected, yield substantial improvements over base models and injecting external knowledge helps models of all sizes. We also find that the format in which knowledge is provided is critical, and that smaller models benefit more from larger amounts of knowledge. Finally, we develop an ontology of knowledge types and analyze the relative coverage of the models across these categories.
ML ID: 407
We propose the task of updated headline generation, in which a system generates a headline for an updated article, considering both the previous article and headline. The system must identify the novel information in the article update, and modify the existing headline accordingly. We create data for this task using the NewsEdits corpus (Spangher and May, 2021) by automatically identifying contiguous article versions that are likely to require a substantive headline update. We find that models conditioned on the prior headline and body re-visions produce headlines judged by humans to be as factual as gold headlines while making fewer unnecessary edits compared to a standard headline generation model. Our experiments establish benchmarks for this new contextual summarization task.
ML ID: 405
When a software bug is reported, developers engage in a discussion to collaboratively resolve it. While the solution is likely formulated within the discussion, it is often buried in a large amount of text, making it difficult to comprehend and delaying its implementation. To expedite bug resolution, we propose generating a concise natural language description of the solution by synthesizing relevant content within the discussion, which encompasses both natural language and source code. We build a corpus for this task using a novel technique for obtaining noisy supervision from repository changes linked to bug reports, with which we establish benchmarks. We also design two systems for generating a description during an ongoing discussion by classifying when sufficient context for performing the task emerges in real-time. With automated and human evaluation, we find this task to form an ideal testbed for complex reasoning in long, bimodal dialogue context.
ML ID: 404
There has been a growing interest in developing machine learning (ML) models for code summarization tasks, e.g., comment generation and method naming. Despite substantial increase in the effectiveness of ML models, the evaluation methodologies, i.e., the way people split datasets into training, validation, and test sets, were not well studied. Specifically, no prior work on code summarization considered the timestamps of code and comments during eval- uation. This may lead to evaluations that are inconsistent with the intended use cases. In this paper, we introduce the time-segmented evaluation methodology, which is novel to the code summarization research community, and compare it with the mixed-project and cross-project methodologies that have been commonly used. Each methodology can be mapped to some use cases, and the time-segmented methodology should be adopted in the evaluation of ML models for code summarization. To assess the impact of methodologies, we collect a dataset of (code, comment) pairs with timestamps to train and evaluate several recent ML models for code summarization. Our experiments show that different methodologies lead to conflicting evaluation results. We invite the community to expand the set of methodologies used in evaluations.
ML ID: 403
The problem of knowledge-based visual question answering involves answering questions that require external knowledge in addition to the content of the image. Such knowledge typically comes in a variety of forms, including visual, textual, and commonsense knowledge. The use of more knowledge sources, however, also increases the chance of retrieving more irrelevant or noisy facts, making it difficult to comprehend the facts and find the answer. To address this challenge, we propose Multi-modal Answer Validation using External knowledge (MAVEx), where the idea is to validate a set of promising answer candidates based on answer-specific knowledge retrieval. This is in contrast to existing approaches that search for the answer in a vast collection of often irrelevant facts. Our approach aims to learn which knowledge source should be trusted for each answer candidate and how to validate the candidate using that source. We consider a multi-modal setting, relying on both textual and visual knowledge resources, including images searched using Google, sentences from Wikipedia articles, and concepts from ConceptNet. Our experiments with OK-VQA, a challenging knowledge-based VQA dataset, demonstrate that MAVEx achieves new state-of-the-art results.
ML ID: 401
Characterizing the patterns of errors that a system makes helps researchers focus future development on increasing its accuracy and robustness. We propose a novel form of ”meta learning” that automatically learns interpretable rules that characterize the types of errors that a system makes, and demonstrate these rules’ ability to help understand and improve two NLP systems. Our approach works by collecting error cases on validation data, extracting meta-features describing these samples, and finally learning rules that characterize errors using these features. We apply our approach to VilBERT, for Visual Question Answering, and RoBERTa, for Common Sense Question Answering. Our system learns interpretable rules that provide insights into systemic errors these systems make on the given tasks. Using these insights, we are also able to “close the loop” and modestly improve performance of these systems.
ML ID: 400
