Claire Bonial

Claire Bonial, Claire Benet Post, Laura Michaelis et al.

Usage-based theories of grammars posit that creative productivity of the structures of language is both bolstered and constrained by two distinct frequency signals: entrenchment, stemming from high frequency usage, and preemption, stemming from having never observed a particular linguistic structure in a context where one might expect that structure to appear. Large Language Models are also usage-based, in the sense that the structures of language are learned through exposure to vast amounts of text. Here, we test whether or not the opposing statistical forces of entrenchment and preemption also encourage and constrain linguistic productivity in LLMs. We demonstrate across model architectures that larger models recognize and can reproduce with nonce words constructional productivity (entrenchment) in cases of coercion, wherein the broader constructional context coerces an atypical interpretation of a lexical item. However, we also show that even the largest models do not extend negative evidence to novel language, and statistical preemption does not enable models to avoid overgeneralization of patterns that are semantically felicitous, but never observed in data.

Stephanie M. Lukin, Kimberly A. Pollard, Claire Bonial et al.

Human-guided robotic exploration is a useful approach to gathering information at remote locations, especially those that might be too risky, inhospitable, or inaccessible for humans. Maintaining common ground between the remotely-located partners is a challenge, one that can be facilitated by multi-modal communication. In this paper, we explore how participants utilized multiple modalities to investigate a remote location with the help of a robotic partner. Participants issued spoken natural language instructions and received from the robot: text-based feedback, continuous 2D LIDAR mapping, and upon-request static photographs. We noticed that different strategies were adopted in terms of use of the modalities, and hypothesize that these differences may be correlated with success at several exploration sub-tasks. We found that requesting photos may have improved the identification and counting of some key entities (doorways in particular) and that this strategy did not hinder the amount of overall area exploration. Future work with larger samples may reveal the effects of more nuanced photo and dialogue strategies, which can inform the training of robotic agents. Additionally, we announce the release of our unique multi-modal corpus of human-robot communication in an exploration context: SCOUT, the Situated Corpus on Understanding Transactions.

Stephanie M. Lukin, Claire Bonial, Matthew Marge et al.

We introduce the Situated Corpus Of Understanding Transactions (SCOUT), a multi-modal collection of human-robot dialogue in the task domain of collaborative exploration. The corpus was constructed from multiple Wizard-of-Oz experiments where human participants gave verbal instructions to a remotely-located robot to move and gather information about its surroundings. SCOUT contains 89,056 utterances and 310,095 words from 278 dialogues averaging 320 utterances per dialogue. The dialogues are aligned with the multi-modal data streams available during the experiments: 5,785 images and 30 maps. The corpus has been annotated with Abstract Meaning Representation and Dialogue-AMR to identify the speaker's intent and meaning within an utterance, and with Transactional Units and Relations to track relationships between utterances to reveal patterns of the Dialogue Structure. We describe how the corpus and its annotations have been used to develop autonomous human-robot systems and enable research in open questions of how humans speak to robots. We release this corpus to accelerate progress in autonomous, situated, human-robot dialogue, especially in the context of navigation tasks where details about the environment need to be discovered.

Claire Bonial, Stephanie M. Lukin, Mitchell Abrams et al.

In this paper, we describe the development of symbolic representations annotated on human-robot dialogue data to make dimensions of meaning accessible to autonomous systems participating in collaborative, natural language dialogue, and to enable common ground with human partners. A particular challenge for establishing common ground arises in remote dialogue (occurring in disaster relief or search-and-rescue tasks), where a human and robot are engaged in a joint navigation and exploration task of an unfamiliar environment, but where the robot cannot immediately share high quality visual information due to limited communication constraints. Engaging in a dialogue provides an effective way to communicate, while on-demand or lower-quality visual information can be supplemented for establishing common ground. Within this paradigm, we capture propositional semantics and the illocutionary force of a single utterance within the dialogue through our Dialogue-AMR annotation, an augmentation of Abstract Meaning Representation. We then capture patterns in how different utterances within and across speaker floors relate to one another in our development of a multi-floor Dialogue Structure annotation schema. Finally, we begin to annotate and analyze the ways in which the visual modalities provide contextual information to the dialogue for overcoming disparities in the collaborators' understanding of the environment. We conclude by discussing the use-cases, architectures, and systems we have implemented from our annotations that enable physical robots to autonomously engage with humans in bi-directional dialogue and navigation.

Harish Tayyar Madabushi, Melissa Torgbi, Claire Bonial

In this position paper we raise critical awareness of a realistic view of LLM capabilities that eschews extreme alternative views that LLMs are either 'stochastic parrots' or in possession of 'emergent' advanced reasoning capabilities, which, due to their unpredictable emergence, constitute an existential threat. Our middle-ground view is that LLMs extrapolate from priors from their training data while using context to guide the model to the appropriate priors; we call this "context-directed extrapolation." Specifically, this context direction is achieved through examples in base models, leading to in-context learning, while instruction tuning allows LLMs to perform similarly based on prompts rather than explicit examples. Under this view, substantiated though existing literature, while reasoning capabilities go well beyond stochastic parroting, such capabilities are predictable, controllable, not indicative of advanced reasoning akin to high-level cognitive capabilities in humans, and not infinitely scalable with additional training. As a result, fears of uncontrollable emergence of agency are allayed, while research advances are appropriately refocused on the processes of context-directed extrapolation and how this interacts with training data to produce valuable capabilities in LLMs. Future work can therefore explore alternative augmenting techniques that do not rely on inherent advanced reasoning in LLMs.

Wesley Scivetti, Melissa Torgbi, Austin Blodgett et al.

The web-scale of pretraining data has created an important evaluation challenge: to disentangle linguistic competence on cases well-represented in pretraining data from generalization to out-of-domain language, specifically the dynamic, real-world instances less common in pretraining data. To this end, we construct a diagnostic evaluation to systematically assess natural language understanding in LLMs by leveraging Construction Grammar (CxG). CxG provides a psycholinguistically grounded framework for testing generalization, as it explicitly links syntactic forms to abstract, non-lexical meanings. Our novel inference evaluation dataset consists of English phrasal constructions, for which speakers are known to be able to abstract over commonplace instantiations in order to understand and produce creative instantiations. Our evaluation dataset uses CxG to evaluate two central questions: first, if models can 'understand' the semantics of sentences for instances that are likely to appear in pretraining data less often, but are intuitive and easy for people to understand. Second, if LLMs can deploy the appropriate constructional semantics given constructions that are syntactically identical but with divergent meanings. Our results demonstrate that state-of-the-art models, including GPT-o1, exhibit a performance drop of over 40% on our second task, revealing a failure to generalize over syntactically identical forms to arrive at distinct constructional meanings in the way humans do. We make our novel dataset and associated experimental data, including prompts and model responses, publicly available.

Claire Bonial, Julia Bonn, Harish Tayyar Madabushi

In this chapter, we argue for the benefits of understanding multiword expressions from the perspective of usage-based, construction grammar approaches. We begin with a historical overview of how construction grammar was developed in order to account for idiomatic expressions using the same grammatical machinery as the non-idiomatic structures of language. We cover a comprehensive description of constructions, which are pairings of meaning with form of any size (morpheme, word, phrase), as well as how constructional approaches treat the acquisition and generalization of constructions. We describe a successful case study leveraging constructional templates for representing multiword expressions in English PropBank. Because constructions can be at any level or unit of form, we then illustrate the benefit of a constructional representation of multi-meaningful morphosyntactic unit constructions in Arapaho, a highly polysynthetic and agglutinating language. We include a second case study leveraging constructional templates for representing these multi-morphemic expressions in Uniform Meaning Representation. Finally, we demonstrate the similarities and differences between a usage-based explanation of a speaker learning a novel multiword expression, such as "dancing with deer," and that of a large language model. We present experiments showing that both models and speakers can generalize the meaning of novel multiword expressions based on a single exposure of usage. However, only speakers can reason over the combination of two such expressions, as this requires comparison of the novel forms to a speaker's lifetime of stored constructional exemplars, which are rich with cross-modal details.

Rupak Sarkar, Neha Srikanth, Taylor Hudson et al.

While it is commonly accepted that maintaining common ground plays a role in conversational success, little prior research exists connecting conversational grounding to success in task-oriented conversations. We study failures of grounding in the Ubuntu IRC dataset, where participants use text-only communication to resolve technical issues. We find that disruptions in conversational flow often stem from a misalignment in common ground, driven by a divergence in beliefs and assumptions held by participants. These disruptions, which we call conversational friction, significantly correlate with task success. We find that although LLMs can identify overt cases of conversational friction, they struggle with subtler and more context-dependent instances requiring pragmatic or domain-specific reasoning.

Tom Mackintosh, Harish Tayyar Madabushi, Claire Bonial

We probe large language models' ability to learn deep form-meaning mappings as defined by construction grammars. We introduce the ConTest-NLI benchmark of 80k sentences covering eight English constructions from highly lexicalized to highly schematic. Our pipeline generates diverse synthetic NLI triples via templating and the application of a model-in-the-loop filter. This provides aspects of human validation to ensure challenge and label reliability. Zero-shot tests on leading LLMs reveal a 24% drop in accuracy between naturalistic (88%) and adversarial data (64%), with schematic patterns proving hardest. Fine-tuning on a subset of ConTest-NLI yields up to 9% improvement, yet our results highlight persistent abstraction gaps in current LLMs and offer a scalable framework for evaluating construction-informed learning.

Mollie Shichman, Claire Bonial, Austin Blodgett et al.

During Human Robot Interactions in disaster relief scenarios, Large Language Models (LLMs) have the potential for substantial physical reasoning to assist in mission objectives. However, these reasoning capabilities are often found only in larger models, which are not currently reasonable to deploy on robotic systems due to size constraints. To meet our problem space requirements, we introduce a dataset and pipeline to create Field Reasoning and Instruction Decoding Agent (FRIDA) models. In our pipeline, domain experts and linguists combine their knowledge to make high-quality, few-shot prompts used to generate synthetic data for fine-tuning. We hand-curate datasets for this few-shot prompting and for evaluation to improve LLM reasoning on both general and disaster-specific objects. We concurrently run an ablation study to understand which kinds of synthetic data most affect performance. We fine-tune several small instruction-tuned models and find that ablated FRIDA models only trained on objects' physical state and function data outperformed both the FRIDA models trained on all synthetic data and the base models in our evaluation. We demonstrate that the FRIDA pipeline is capable of instilling physical common sense with minimal data.

Navita Goyal, Eleftheria Briakou, Amanda Liu et al.

NLP systems have shown impressive performance at answering questions by retrieving relevant context. However, with the increasingly large models, it is impossible and often undesirable to constrain models' knowledge or reasoning to only the retrieved context. This leads to a mismatch between the information that the models access to derive the answer and the information that is available to the user to assess the model predicted answer. In this work, we study how users interact with QA systems in the absence of sufficient information to assess their predictions. Further, we ask whether adding the requisite background helps mitigate users' over-reliance on predictions. Our study reveals that users rely on model predictions even in the absence of sufficient information needed to assess the model's correctness. Providing the relevant background, however, helps users better catch model errors, reducing over-reliance on incorrect predictions. On the flip side, background information also increases users' confidence in their accurate as well as inaccurate judgments. Our work highlights that supporting users' verification of QA predictions is an important, yet challenging, problem.

Stephanie M. Lukin, Claire Bonial, Clare R. Voss

We describe the task of Visual Understanding and Narration, in which a robot (or agent) generates text for the images that it collects when navigating its environment, by answering open-ended questions, such as 'what happens, or might have happened, here?'

Matthew Marge, Claire Bonial, Stephanie Lukin et al.

We describe a multi-phased Wizard-of-Oz approach to collecting human-robot dialogue in a collaborative search and navigation task. The data is being used to train an initial automated robot dialogue system to support collaborative exploration tasks. In the first phase, a wizard freely typed robot utterances to human participants. For the second phase, this data was used to design a GUI that includes buttons for the most common communications, and templates for communications with varying parameters. Comparison of the data gathered in these phases show that the GUI enabled a faster pace of dialogue while still maintaining high coverage of suitable responses, enabling more efficient targeted data collection, and improvements in natural language understanding using GUI-collected data. As a promising first step towards interactive learning, this work shows that our approach enables the collection of useful training data for navigation-based HRI tasks.

Stephanie M. Lukin, Kimberly A. Pollard, Claire Bonial et al.

This paper identifies stylistic differences in instruction-giving observed in a corpus of human-robot dialogue. Differences in verbosity and structure (i.e., single-intent vs. multi-intent instructions) arose naturally without restrictions or prior guidance on how users should speak with the robot. Different styles were found to produce different rates of miscommunication, and correlations were found between style differences and individual user variation, trust, and interaction experience with the robot. Understanding potential consequences and factors that influence style can inform design of dialogue systems that are robust to natural variation from human users.

Sungmin Eum, Christopher Reale, Heesung Kwon et al.

Many previous methods have demonstrated the importance of considering semantically relevant objects for carrying out video-based human activity recognition, yet none of the methods have harvested the power of large text corpora to relate the objects and the activities to be transferred into learning a unified deep convolutional neural network. We present a novel activity recognition CNN which co-learns the object recognition task in an end-to-end multitask learning scheme to improve upon the baseline activity recognition performance. We further improve upon the multitask learning approach by exploiting a text-guided semantic space to select the most relevant objects with respect to the target activities. To the best of our knowledge, we are the first to investigate this approach.

Claire Bonial, Matthew Marge, Ron artstein et al.

We describe the adaptation and refinement of a graphical user interface designed to facilitate a Wizard-of-Oz (WoZ) approach to collecting human-robot dialogue data. The data collected will be used to develop a dialogue system for robot navigation. Building on an interface previously used in the development of dialogue systems for virtual agents and video playback, we add templates with open parameters which allow the wizard to quickly produce a wide variety of utterances. Our research demonstrates that this approach to data collection is viable as an intermediate step in developing a dialogue system for physical robots in remote locations from their users - a domain in which the human and robot need to regularly verify and update a shared understanding of the physical environment. We show that our WoZ interface and the fixed set of utterances and templates therein provide for a natural pace of dialogue with good coverage of the navigation domain.

Matthew Marge, Claire Bonial, Brendan Byrne et al.

Our overall program objective is to provide more natural ways for soldiers to interact and communicate with robots, much like how soldiers communicate with other soldiers today. We describe how the Wizard-of-Oz (WOz) method can be applied to multimodal human-robot dialogue in a collaborative exploration task. While the WOz method can help design robot behaviors, traditional approaches place the burden of decisions on a single wizard. In this work, we consider two wizards to stand in for robot navigation and dialogue management software components. The scenario used to elicit data is one in which a human-robot team is tasked with exploring an unknown environment: a human gives verbal instructions from a remote location and the robot follows them, clarifying possible misunderstandings as needed via dialogue. We found the division of labor between wizards to be workable, which holds promise for future software development.