Generating step-by-step "chain-of-thought" rationales improves language model performance on complex reasoning tasks like mathematics or commonsense question-answering. However, inducing language model rationale generation currently requires either constructing massive rationale datasets or sacrificing accuracy by using only few-shot inference. We propose a technique to iteratively leverage a small number of rationale examples and a large dataset without rationales, to bootstrap the ability to perform successively more complex reasoning. This technique, the "Self-Taught Reasoner" (STaR), relies on a simple loop: generate rationales to answer many questions, prompted with a few rationale examples; if the generated answers are wrong, try again to generate a rationale given the correct answer; fine-tune on all the rationales that ultimately yielded correct answers; repeat. We show that STaR significantly improves performance on multiple datasets compared to a model fine-tuned to directly predict final answers, and performs comparably to fine-tuning a 30$\times$ larger state-of-the-art language model on CommensenseQA. Thus, STaR lets a model improve itself by learning from its own generated reasoning.
Distillation efforts have led to language models that are more compact and efficient without serious drops in performance. The standard approach to distillation trains a student model against two objectives: a task-specific objective (e.g., language modeling) and an imitation objective that encourages the hidden states of the student model to be similar to those of the larger teacher model. In this paper, we show that it is beneficial to augment distillation with a third objective that encourages the student to imitate the causal computation process of the teacher through interchange intervention training(IIT). IIT pushes the student model to become a causal abstraction of the teacher model - a simpler model with the same causal structure. IIT is fully differentiable, easily implemented, and combines flexibly with other objectives. Compared with standard distillation of BERT, distillation via IIT results in lower perplexity on Wikipedia (masked language modeling) and marked improvements on the GLUE benchmark (natural language understanding), SQuAD (question answering), and CoNLL-2003 (named entity recognition).
In many areas, we have well-founded insights about causal structure that would be useful to bring into our trained models while still allowing them to learn in a data-driven fashion. To achieve this, we present the new method of interchange intervention training(IIT). In IIT, we (1)align variables in the causal model with representations in the neural model and (2) train a neural model to match the counterfactual behavior of the causal model on a base input when aligned representations in both models are set to be the value they would be for a second source input. IIT is fully differentiable, flexibly combines with other objectives, and guarantees that the target causal model is acausal abstraction of the neural model when its loss is minimized. We evaluate IIT on a structured vision task (MNIST-PVR) and a navigational instruction task (ReaSCAN). We compare IIT against multi-task training objectives and data augmentation. In all our experiments, IIT achieves the best results and produces neural models that are more interpretable in the sense that they realize the target causal model.
To be good conversational partners, natural language processing (NLP) systems should be trained to produce contextually useful utterances. Prior work has investigated training NLP systems with communication-based objectives, where a neural listener stands in as a communication partner. However, these systems commonly suffer from semantic drift where the learned language diverges radically from natural language. We propose a method that uses a population of neural listeners to regularize speaker training. We first show that language drift originates from the poor uncertainty calibration of a neural listener, which makes high-certainty predictions on novel sentences. We explore ensemble- and dropout-based populations of listeners and find that the former results in better uncertainty quantification. We evaluate both population-based objectives on reference games, and show that the ensemble method with better calibration enables the speaker to generate pragmatic utterances while scaling to a large vocabulary and generalizing to new games and listeners.
From photorealistic sketches to schematic diagrams, drawing provides a versatile medium for communicating about the visual world. How do images spanning such a broad range of appearances reliably convey meaning? Do viewers understand drawings based solely on their ability to resemble the entities they refer to (i.e., as images), or do they understand drawings based on shared but arbitrary associations with these entities (i.e., as symbols)? In this paper, we provide evidence for a cognitive account of pictorial meaning in which both visual and social information is integrated to support effective visual communication. To evaluate this account, we used a communication task where pairs of participants used drawings to repeatedly communicate the identity of a target object among multiple distractor objects. We manipulated social cues across three experiments and a full internal replication, finding pairs of participants develop referent-specific and interaction-specific strategies for communicating more efficiently over time, going beyond what could be explained by either task practice or a pure resemblance-based account alone. Using a combination of model-based image analyses and crowdsourced sketch annotations, we further determined that drawings did not drift toward arbitrariness, as predicted by a pure convention-based account, but systematically preserved those visual features that were most distinctive of the target object. Taken together, these findings advance theories of pictorial meaning and have implications for how successful graphical conventions emerge via complex interactions between visual perception, communicative experience, and social context.
Images have become an integral part of online media. This has enhanced self-expression and the dissemination of knowledge, but it poses serious accessibility challenges. Adequate textual descriptions are rare. Captions are more abundant, but they do not consistently provide the needed descriptive details, and systems trained on such texts inherit these shortcomings. To address this, we introduce the publicly available Wikipedia-based corpus Concadia, which consists of 96,918 images with corresponding English-language descriptions, captions, and surrounding context. We use Concadia to further characterize the commonalities and differences between descriptions and captions, and this leads us to the hypothesis that captions, while not substitutes for descriptions, can provide a useful signal for creating effective descriptions. We substantiate this hypothesis by showing that image captioning systems trained on Concadia benefit from having caption embeddings as part of their inputs. These experiments also begin to show how Concadia can be a powerful tool in addressing the underlying accessibility issues posed by image data.
Languages are powerful solutions to coordination problems: they provide stable, shared expectations about how the words we say correspond to the beliefs and intentions in our heads. Yet language use in a variable and non-stationary social environment requires linguistic representations to be flexible: old words acquire new ad hoc or partner-specific meanings on the fly. In this paper, we introduce a hierarchical Bayesian theory of convention formation that aims to reconcile the long-standing tension between these two basic observations. More specifically, we argue that the central computational problem of communication is not simply transmission, as in classical formulations, but learning and adaptation over multiple timescales. Under our account, rapid learning within dyadic interactions allows for coordination on partner-specific common ground, while social conventions are stable priors that have been abstracted away from interactions with multiple partners. We present new empirical data alongside simulations showing how our model provides a cognitive foundation for explaining several phenomena that have posed a challenge for previous accounts: (1) the convergence to more efficient referring expressions across repeated interaction with the same partner, (2) the gradual transfer of partner-specific common ground to novel partners, and (3) the influence of communicative context on which conventions eventually form.
A hallmark of human language is the ability to effectively and efficiently convey contextually relevant information. One theory for how humans reason about language is presented in the Rational Speech Acts (RSA) framework, which captures pragmatic phenomena via a process of recursive social reasoning (Goodman & Frank, 2016). However, RSA represents ideal reasoning in an unconstrained setting. We explore the idea that speakers might learn to amortize the cost of RSA computation over time by directly optimizing for successful communication with an internal listener model. In simulations with grounded neural speakers and listeners across two communication game datasets representing synthetic and human-generated data, we find that our amortized model is able to quickly generate language that is effective and concise across a range of contexts, without the need for explicit pragmatic reasoning.
A key property of linguistic conventions is that they hold over an entire community of speakers, allowing us to communicate efficiently even with people we have never met before. At the same time, much of our language use is partner-specific: we know that words may be understood differently by different people based on local common ground. This poses a challenge for accounts of convention formation. Exactly how do agents make the inferential leap to community-wide expectations while maintaining partner-specific knowledge? We propose a hierarchical Bayesian model of convention to explain how speakers and listeners abstract away meanings that seem to be shared across partners. To evaluate our model's predictions, we conducted an experiment where participants played an extended natural-language communication game with different partners in a small community. We examine several measures of generalization across partners, and find key signatures of local adaptation as well as collective convergence. These results suggest that local partner-specific learning is not only compatible with global convention formation but may facilitate it when coupled with a powerful hierarchical inductive mechanism.
The language we use over the course of conversation changes as we establish common ground and learn what our partner finds meaningful. Here we draw upon recent advances in natural language processing to provide a finer-grained characterization of the dynamics of this learning process. We release an open corpus (>15,000 utterances) of extended dyadic interactions in a classic repeated reference game task where pairs of participants had to coordinate on how to refer to initially difficult-to-describe tangram stimuli. We find that different pairs discover a wide variety of idiosyncratic but efficient and stable solutions to the problem of reference. Furthermore, these conventions are shaped by the communicative context: words that are more discriminative in the initial context (i.e. that are used for one target more than others) are more likely to persist through the final repetition. Finally, we find systematic structure in how a speaker's referring expressions become more efficient over time: syntactic units drop out in clusters following positive feedback from the listener, eventually leaving short labels containing open-class parts of speech. These findings provide a higher resolution look at the quantitative dynamics of ad hoc convention formation and support further development of computational models of learning in communication.