Allyson Ettinger

Peter West, Ximing Lu, Nouha Dziri et al. · allen-ai, cmu

The recent wave of generative AI has sparked unprecedented global attention, with both excitement and concern over potentially superhuman levels of artificial intelligence: models now take only seconds to produce outputs that would challenge or exceed the capabilities even of expert humans. At the same time, models still show basic errors in understanding that would not be expected even in non-expert humans. This presents us with an apparent paradox: how do we reconcile seemingly superhuman capabilities with the persistence of errors that few humans would make? In this work, we posit that this tension reflects a divergence in the configuration of intelligence in today's generative models relative to intelligence in humans. Specifically, we propose and test the Generative AI Paradox hypothesis: generative models, having been trained directly to reproduce expert-like outputs, acquire generative capabilities that are not contingent upon -- and can therefore exceed -- their ability to understand those same types of outputs. This contrasts with humans, for whom basic understanding almost always precedes the ability to generate expert-level outputs. We test this hypothesis through controlled experiments analyzing generation vs. understanding in generative models, across both language and image modalities. Our results show that although models can outperform humans in generation, they consistently fall short of human capabilities in measures of understanding, as well as weaker correlation between generation and understanding performance, and more brittleness to adversarial inputs. Our findings support the hypothesis that models' generative capability may not be contingent upon understanding capability, and call for caution in interpreting artificial intelligence by analogy to human intelligence.

Allyson Ettinger, Jena D. Hwang, Valentina Pyatkin et al. · allen-ai

Large language models (LLMs) show amazing proficiency and fluency in the use of language. Does this mean that they have also acquired insightful linguistic knowledge about the language, to an extent that they can serve as an "expert linguistic annotator"? In this paper, we examine the successes and limitations of the GPT-3, ChatGPT, and GPT-4 models in analysis of sentence meaning structure, focusing on the Abstract Meaning Representation (AMR; Banarescu et al. 2013) parsing formalism, which provides rich graphical representations of sentence meaning structure while abstracting away from surface forms. We compare models' analysis of this semantic structure across two settings: 1) direct production of AMR parses based on zero- and few-shot prompts, and 2) indirect partial reconstruction of AMR via metalinguistic natural language queries (e.g., "Identify the primary event of this sentence, and the predicate corresponding to that event."). Across these settings, we find that models can reliably reproduce the basic format of AMR, and can often capture core event, argument, and modifier structure -- however, model outputs are prone to frequent and major errors, and holistic analysis of parse acceptability shows that even with few-shot demonstrations, models have virtually 0% success in producing fully accurate parses. Eliciting natural language responses produces similar patterns of errors. Overall, our findings indicate that these models out-of-the-box can capture aspects of semantic structure, but there remain key limitations in their ability to support fully accurate semantic analyses or parses.

Chenghao Yang, Allyson Ettinger

Understanding sentence meanings and updating information states appropriately across time -- what we call "situational understanding" (SU) -- is a critical ability for human-like AI agents. SU is essential in particular for chat models, such as ChatGPT, to enable consistent, coherent, and effective dialogue between humans and AI. Previous works have identified certain SU limitations in non-chatbot Large Language models (LLMs), but the extent and causes of these limitations are not well understood, and capabilities of current chat-based models in this domain have not been explored. In this work we tackle these questions, proposing a novel synthetic environment for SU testing which allows us to do controlled and systematic testing of SU in chat-oriented models, through assessment of models' ability to track and enumerate environment states. Our environment also allows for close analysis of dynamics of model performance, to better understand underlying causes for performance patterns. We apply our test to ChatGPT, the state-of-the-art chatbot, and find that despite the fundamental simplicity of the task, the model's performance reflects an inability to retain correct environment states across time. Our follow-up analyses suggest that performance degradation is largely because ChatGPT has non-persistent in-context memory (although it can access the full dialogue history) and it is susceptible to hallucinated updates -- including updates that artificially inflate accuracies. Our findings suggest overall that ChatGPT is not currently equipped for robust tracking of situation states, and that trust in the impressive dialogue performance of ChatGPT comes with risks. We release the codebase for reproducing our test environment, as well as all prompts and API responses from ChatGPT, at https://github.com/yangalan123/SituationalTesting.

Kanishka Misra, Julia Taylor Rayz, Allyson Ettinger

A characteristic feature of human semantic cognition is its ability to not only store and retrieve the properties of concepts observed through experience, but to also facilitate the inheritance of properties (can breathe) from superordinate concepts (animal) to their subordinates (dog) -- i.e. demonstrate property inheritance. In this paper, we present COMPS, a collection of minimal pair sentences that jointly tests pre-trained language models (PLMs) on their ability to attribute properties to concepts and their ability to demonstrate property inheritance behavior. Analyses of 22 different PLMs on COMPS reveal that they can easily distinguish between concepts on the basis of a property when they are trivially different, but find it relatively difficult when concepts are related on the basis of nuanced knowledge representations. Furthermore, we find that PLMs can demonstrate behavior consistent with property inheritance to a great extent, but fail in the presence of distracting information, which decreases the performance of many models, sometimes even below chance. This lack of robustness in demonstrating simple reasoning raises important questions about PLMs' capacity to make correct inferences even when they appear to possess the prerequisite knowledge.

Kanishka Misra, Julia Taylor Rayz, Allyson Ettinger

To what extent can experience from language contribute to our conceptual knowledge? Computational explorations of this question have shed light on the ability of powerful neural language models (LMs) -- informed solely through text input -- to encode and elicit information about concepts and properties. To extend this line of research, we present a framework that uses neural-network language models (LMs) to perform property induction -- a task in which humans generalize novel property knowledge (has sesamoid bones) from one or more concepts (robins) to others (sparrows, canaries). Patterns of property induction observed in humans have shed considerable light on the nature and organization of human conceptual knowledge. Inspired by this insight, we use our framework to explore the property inductions of LMs, and find that they show an inductive preference to generalize novel properties on the basis of category membership, suggesting the presence of a taxonomic bias in their representations.

Team Olmo, Allyson Ettinger, Amanda Bertsch et al. · uw

We introduce Olmo 3, a family of state-of-the-art, fully-open language models at the 7B and 32B parameter scales. Olmo 3 model construction targets long-context reasoning, function calling, coding, instruction following, general chat, and knowledge recall. This release includes the entire model flow, i.e., the full lifecycle of the family of models, including every stage, checkpoint, data point, and dependency used to build it. Our flagship model, Olmo 3 Think 32B, is the strongest fully-open thinking model released to-date.

Jiaxuan Li, Lang Yu, Allyson Ettinger

Current pre-trained language models have enabled remarkable improvements in downstream tasks, but it remains difficult to distinguish effects of statistical correlation from more systematic logical reasoning grounded on understanding of the real world. In this paper we tease these factors apart by leveraging counterfactual conditionals, which force language models to predict unusual consequences based on hypothetical propositions. We introduce a set of tests drawn from psycholinguistic experiments, as well as larger-scale controlled datasets, to probe counterfactual predictions from a variety of popular pre-trained language models. We find that models are consistently able to override real-world knowledge in counterfactual scenarios, and that this effect is more robust in case of stronger baseline world knowledge -- however, we also find that for most models this effect appears largely to be driven by simple lexical cues. When we mitigate effects of both world knowledge and lexical cues to test knowledge of linguistic nuances of counterfactuals, we find that only GPT-3 shows sensitivity to these nuances, though this sensitivity is also non-trivially impacted by lexical associative factors.

Sanghee J. Kim, Lang Yu, Allyson Ettinger

A critical component of competence in language is being able to identify relevant components of an utterance and reply appropriately. In this paper we examine the extent of such dialogue response sensitivity in pre-trained language models, conducting a series of experiments with a particular focus on sensitivity to dynamics involving phenomena of at-issueness and ellipsis. We find that models show clear sensitivity to a distinctive role of embedded clauses, and a general preference for responses that target main clause content of prior utterances. However, the results indicate mixed and generally weak trends with respect to capturing the full range of dynamics involved in targeting at-issue versus not-at-issue content. Additionally, models show fundamental limitations in grasp of the dynamics governing ellipsis, and response selections show clear interference from superficial factors that outweigh the influence of principled discourse constraints.

Yanhong Li, Chenghao Yang, Allyson Ettinger

Recent studies suggest that self-reflective prompting can significantly enhance the reasoning capabilities of Large Language Models (LLMs). However, the use of external feedback as a stop criterion raises doubts about the true extent of LLMs' ability to emulate human-like self-reflection. In this paper, we set out to clarify these capabilities under a more stringent evaluation setting in which we disallow any kind of external feedback. Our findings under this setting show a split: while self-reflection enhances performance in TruthfulQA, it adversely affects results in HotpotQA. We conduct follow-up analyses to clarify the contributing factors in these patterns, and find that the influence of self-reflection is impacted both by reliability of accuracy in models' initial responses, and by overall question difficulty: specifically, self-reflection shows the most benefit when models are less likely to be correct initially, and when overall question difficulty is higher. We also find that self-reflection reduces tendency toward majority voting. Based on our findings, we propose guidelines for decisions on when to implement self-reflection. We release the codebase for reproducing our experiments at https://github.com/yanhong-lbh/LLM-SelfReflection-Eval.

Team OLMo, Pete Walsh, Luca Soldaini et al. · allen-ai, cambridge

We present OLMo 2, the next generation of our fully open language models. OLMo 2 includes a family of dense autoregressive language models at 7B, 13B and 32B scales with fully released artifacts -- model weights, full training data, training code and recipes, training logs and thousands of intermediate checkpoints. In this work, we describe our modified model architecture and training recipe, focusing on techniques for achieving better training stability and improved per-token efficiency. Our updated pretraining data mixture introduces a new, specialized data mix called Dolmino Mix 1124, which significantly improves model capabilities across many downstream task benchmarks when introduced via late-stage curriculum training (i.e. specialized data during the annealing phase of pretraining). Finally, we incorporate best practices from Tülu 3 to develop OLMo 2-Instruct, focusing on permissive data and extending our final-stage reinforcement learning with verifiable rewards (RLVR). Our OLMo 2 base models sit at the Pareto frontier of performance to training compute, often matching or outperforming open-weight only models like Llama 3.1, Qwen 2.5, and Gemma 2 while using fewer FLOPs and with fully transparent training data, code, and recipe. Our fully open OLMo 2-Instruct models are competitive with open-weight only models of comparable size and even some proprietary models like GPT-3.5 Turbo and GPT 4o Mini.

Liwei Jiang, Kavel Rao, Seungju Han et al.

We introduce WildTeaming, an automatic LLM safety red-teaming framework that mines in-the-wild user-chatbot interactions to discover 5.7K unique clusters of novel jailbreak tactics, and then composes multiple tactics for systematic exploration of novel jailbreaks. Compared to prior work that performed red-teaming via recruited human workers, gradient-based optimization, or iterative revision with LLMs, our work investigates jailbreaks from chatbot users who were not specifically instructed to break the system. WildTeaming reveals previously unidentified vulnerabilities of frontier LLMs, resulting in up to 4.6x more diverse and successful adversarial attacks compared to state-of-the-art jailbreak methods. While many datasets exist for jailbreak evaluation, very few open-source datasets exist for jailbreak training, as safety training data has been closed even when model weights are open. With WildTeaming we create WildJailbreak, a large-scale open-source synthetic safety dataset with 262K vanilla (direct request) and adversarial (complex jailbreak) prompt-response pairs. To mitigate exaggerated safety behaviors, WildJailbreak provides two contrastive types of queries: 1) harmful queries (vanilla & adversarial) and 2) benign queries that resemble harmful queries in form but contain no harm. As WildJailbreak considerably upgrades the quality and scale of existing safety resources, it uniquely enables us to examine the scaling effects of data and the interplay of data properties and model capabilities during safety training. Through extensive experiments, we identify the training properties that enable an ideal balance of safety behaviors: appropriate safeguarding without over-refusal, effective handling of vanilla and adversarial queries, and minimal, if any, decrease in general capabilities. All components of WildJailbeak contribute to achieving balanced safety behaviors of models.

Seungju Han, Kavel Rao, Allyson Ettinger et al.

We introduce WildGuard -- an open, light-weight moderation tool for LLM safety that achieves three goals: (1) identifying malicious intent in user prompts, (2) detecting safety risks of model responses, and (3) determining model refusal rate. Together, WildGuard serves the increasing needs for automatic safety moderation and evaluation of LLM interactions, providing a one-stop tool with enhanced accuracy and broad coverage across 13 risk categories. While existing open moderation tools such as Llama-Guard2 score reasonably well in classifying straightforward model interactions, they lag far behind a prompted GPT-4, especially in identifying adversarial jailbreaks and in evaluating models' refusals, a key measure for evaluating safety behaviors in model responses. To address these challenges, we construct WildGuardMix, a large-scale and carefully balanced multi-task safety moderation dataset with 92K labeled examples that cover vanilla (direct) prompts and adversarial jailbreaks, paired with various refusal and compliance responses. WildGuardMix is a combination of WildGuardTrain, the training data of WildGuard, and WildGuardTest, a high-quality human-annotated moderation test set with 5K labeled items covering broad risk scenarios. Through extensive evaluations on WildGuardTest and ten existing public benchmarks, we show that WildGuard establishes state-of-the-art performance in open-source safety moderation across all the three tasks compared to ten strong existing open-source moderation models (e.g., up to 26.4% improvement on refusal detection). Importantly, WildGuard matches and sometimes exceeds GPT-4 performance (e.g., up to 3.9% improvement on prompt harmfulness identification). WildGuard serves as a highly effective safety moderator in an LLM interface, reducing the success rate of jailbreak attacks from 79.8% to 2.4%.

Sean McGregor, Allyson Ettinger, Nick Judd et al.

In August of 2024, 495 hackers generated evaluations in an open-ended bug bounty targeting the Open Language Model (OLMo) from The Allen Institute for AI. A vendor panel staffed by representatives of OLMo's safety program adjudicated changes to OLMo's documentation and awarded cash bounties to participants who successfully demonstrated a need for public disclosure clarifying the intent, capacities, and hazards of model deployment. This paper presents a collection of lessons learned, illustrative of flaw reporting best practices intended to reduce the likelihood of incidents and produce safer large language models (LLMs). These include best practices for safety reporting processes, their artifacts, and safety program staffing.

Kanishka Misra, Allyson Ettinger, Kyle Mahowald

Recent zero-shot evaluations have highlighted important limitations in the abilities of language models (LMs) to perform meaning extraction. However, it is now well known that LMs can demonstrate radical improvements in the presence of experimental contexts such as in-context examples and instructions. How well does this translate to previously studied meaning-sensitive tasks? We present a case-study on the extent to which experimental contexts can improve LMs' robustness in performing property inheritance -- predicting semantic properties of novel concepts, a task that they have been previously shown to fail on. Upon carefully controlling the nature of the in-context examples and the instructions, our work reveals that they can indeed lead to non-trivial property inheritance behavior in LMs. However, this ability is inconsistent: with a minimal reformulation of the task, some LMs were found to pick up on shallow, non-semantic heuristics from their inputs, suggesting that the computational principles of semantic property inference are yet to be mastered by LMs.

Weijia Shi, Akshita Bhagia, Kevin Farhat et al. · allen-ai

We introduce FlexOlmo, a new class of language models (LMs) that supports (1) distributed training without data sharing, where different model parameters are independently trained on closed datasets, and (2) data-flexible inference, where these parameters along with their associated data can be flexibly included or excluded from model inferences with no further training. FlexOlmo employs a mixture-of-experts (MoE) architecture where each expert is trained independently on closed datasets and later integrated through a new domain-informed routing without any joint training. FlexOlmo is trained on FlexMix, a corpus we curate comprising publicly available datasets alongside seven domain-specific sets, representing realistic approximations of closed sets. We evaluate models with up to 37 billion parameters (20 billion active) on 31 diverse downstream tasks. We show that a general expert trained on public data can be effectively combined with independently trained experts from other data owners, leading to an average 41% relative improvement while allowing users to opt out of certain data based on data licensing or permission requirements. Our approach also outperforms prior model merging methods by 10.1% on average and surpasses the standard MoE trained without data restrictions using the same training FLOPs. Altogether, this research presents a solution for both data owners and researchers in regulated industries with sensitive or protected data. FlexOlmo enables benefiting from closed data while respecting data owners' preferences by keeping their data local and supporting fine-grained control of data access during inference.

Nouha Dziri, Ximing Lu, Melanie Sclar et al.

Transformer large language models (LLMs) have sparked admiration for their exceptional performance on tasks that demand intricate multi-step reasoning. Yet, these models simultaneously show failures on surprisingly trivial problems. This begs the question: Are these errors incidental, or do they signal more substantial limitations? In an attempt to demystify transformer LLMs, we investigate the limits of these models across three representative compositional tasks -- multi-digit multiplication, logic grid puzzles, and a classic dynamic programming problem. These tasks require breaking problems down into sub-steps and synthesizing these steps into a precise answer. We formulate compositional tasks as computation graphs to systematically quantify the level of complexity, and break down reasoning steps into intermediate sub-procedures. Our empirical findings suggest that transformer LLMs solve compositional tasks by reducing multi-step compositional reasoning into linearized subgraph matching, without necessarily developing systematic problem-solving skills. To round off our empirical study, we provide theoretical arguments on abstract multi-step reasoning problems that highlight how autoregressive generations' performance can rapidly decay with\,increased\,task\,complexity.

Jiaxuan Li, Lang Yu, Allyson Ettinger

Current pre-trained language models have enabled remarkable improvements in downstream tasks, but it remains difficult to distinguish effects of statistical correlation from more systematic logical reasoning grounded on the understanding of real world. We tease these factors apart by leveraging counterfactual conditionals, which force language models to predict unusual consequences based on hypothetical propositions. We introduce a set of tests from psycholinguistic experiments, as well as larger-scale controlled datasets, to probe counterfactual predictions from five pre-trained language models. We find that models are consistently able to override real-world knowledge in counterfactual scenarios, and that this effect is more robust in case of stronger baseline world knowledge -- however, we also find that for most models this effect appears largely to be driven by simple lexical cues. When we mitigate effects of both world knowledge and lexical cues to test knowledge of linguistic nuances of counterfactuals, we find that only GPT-3 shows sensitivity to these nuances, though this sensitivity is also non-trivially impacted by lexical associative factors.

Qinxuan Wu, Allyson Ettinger

While sentence anomalies have been applied periodically for testing in NLP, we have yet to establish a picture of the precise status of anomaly information in representations from NLP models. In this paper we aim to fill two primary gaps, focusing on the domain of syntactic anomalies. First, we explore fine-grained differences in anomaly encoding by designing probing tasks that vary the hierarchical level at which anomalies occur in a sentence. Second, we test not only models' ability to detect a given anomaly, but also the generality of the detected anomaly signal, by examining transfer between distinct anomaly types. Results suggest that all models encode some information supporting anomaly detection, but detection performance varies between anomalies, and only representations from more recent transformer models show signs of generalized knowledge of anomalies. Follow-up analyses support the notion that these models pick up on a legitimate, general notion of sentence oddity, while coarser-grained word position information is likely also a contributor to the observed anomaly detection.

Lalchand Pandia, Yan Cong, Allyson Ettinger

As pre-trained language models (LMs) continue to dominate NLP, it is increasingly important that we understand the depth of language capabilities in these models. In this paper, we target pre-trained LMs' competence in pragmatics, with a focus on pragmatics relating to discourse connectives. We formulate cloze-style tests using a combination of naturally-occurring data and controlled inputs drawn from psycholinguistics. We focus on testing models' ability to use pragmatic cues to predict discourse connectives, models' ability to understand implicatures relating to connectives, and the extent to which models show humanlike preferences regarding temporal dynamics of connectives. We find that although models predict connectives reasonably well in the context of naturally-occurring data, when we control contexts to isolate high-level pragmatic cues, model sensitivity is much lower. Models also do not show substantial humanlike temporal preferences. Overall, the findings suggest that at present, dominant pre-training paradigms do not result in substantial pragmatic competence in our models.

Lalchand Pandia, Allyson Ettinger

Pre-trained LMs have shown impressive performance on downstream NLP tasks, but we have yet to establish a clear understanding of their sophistication when it comes to processing, retaining, and applying information presented in their input. In this paper we tackle a component of this question by examining robustness of models' ability to deploy relevant context information in the face of distracting content. We present models with cloze tasks requiring use of critical context information, and introduce distracting content to test how robustly the models retain and use that critical information for prediction. We also systematically manipulate the nature of these distractors, to shed light on dynamics of models' use of contextual cues. We find that although models appear in simple contexts to make predictions based on understanding and applying relevant facts from prior context, the presence of distracting but irrelevant content has clear impact in confusing model predictions. In particular, models appear particularly susceptible to factors of semantic similarity and word position. The findings are consistent with the conclusion that LM predictions are driven in large part by superficial contextual cues, rather than by robust representations of context meaning.

Lang Yu, Allyson Ettinger

Pre-trained transformer language models have shown remarkable performance on a variety of NLP tasks. However, recent research has suggested that phrase-level representations in these models reflect heavy influences of lexical content, but lack evidence of sophisticated, compositional phrase information. Here we investigate the impact of fine-tuning on the capacity of contextualized embeddings to capture phrase meaning information beyond lexical content. Specifically, we fine-tune models on an adversarial paraphrase classification task with high lexical overlap, and on a sentiment classification task. After fine-tuning, we analyze phrasal representations in controlled settings following prior work. We find that fine-tuning largely fails to benefit compositionality in these representations, though training on sentiment yields a small, localized benefit for certain models. In follow-up analyses, we identify confounding cues in the paraphrase dataset that may explain the lack of composition benefits from that task, and we discuss potential factors underlying the localized benefits from sentiment training.

Kanishka Misra, Allyson Ettinger, Julia Taylor Rayz

Building on research arguing for the possibility of conceptual and categorical knowledge acquisition through statistics contained in language, we evaluate predictive language models (LMs) -- informed solely by textual input -- on a prevalent phenomenon in cognitive science: typicality. Inspired by experiments that involve language processing and show robust typicality effects in humans, we propose two tests for LMs. Our first test targets whether typicality modulates LM probabilities in assigning taxonomic category memberships to items. The second test investigates sensitivities to typicality in LMs' probabilities when extending new information about items to their categories. Both tests show modest -- but not completely absent -- correspondence between LMs and humans, suggesting that text-based exposure alone is insufficient to acquire typicality knowledge.

Lang Yu, Allyson Ettinger

Deep transformer models have pushed performance on NLP tasks to new limits, suggesting sophisticated treatment of complex linguistic inputs, such as phrases. However, we have limited understanding of how these models handle representation of phrases, and whether this reflects sophisticated composition of phrase meaning like that done by humans. In this paper, we present systematic analysis of phrasal representations in state-of-the-art pre-trained transformers. We use tests leveraging human judgments of phrase similarity and meaning shift, and compare results before and after control of word overlap, to tease apart lexical effects versus composition effects. We find that phrase representation in these models relies heavily on word content, with little evidence of nuanced composition. We also identify variations in phrase representation quality across models, layers, and representation types, and make corresponding recommendations for usage of representations from these models.

Kanishka Misra, Allyson Ettinger, Julia Taylor Rayz

Models trained to estimate word probabilities in context have become ubiquitous in natural language processing. How do these models use lexical cues in context to inform their word probabilities? To answer this question, we present a case study analyzing the pre-trained BERT model with tests informed by semantic priming. Using English lexical stimuli that show priming in humans, we find that BERT too shows "priming," predicting a word with greater probability when the context includes a related word versus an unrelated one. This effect decreases as the amount of information provided by the context increases. Follow-up analysis shows BERT to be increasingly distracted by related prime words as context becomes more informative, assigning lower probabilities to related words. Our findings highlight the importance of considering contextual constraint effects when studying word prediction in these models, and highlight possible parallels with human processing.

Shubham Toshniwal, Sam Wiseman, Allyson Ettinger et al.

Long document coreference resolution remains a challenging task due to the large memory and runtime requirements of current models. Recent work doing incremental coreference resolution using just the global representation of entities shows practical benefits but requires keeping all entities in memory, which can be impractical for long documents. We argue that keeping all entities in memory is unnecessary, and we propose a memory-augmented neural network that tracks only a small bounded number of entities at a time, thus guaranteeing a linear runtime in length of document. We show that (a) the model remains competitive with models with high memory and computational requirements on OntoNotes and LitBank, and (b) the model learns an efficient memory management strategy easily outperforming a rule-based strategy.

Davis Yoshida, Allyson Ettinger, Kevin Gimpel

Fine-tuning a pretrained transformer for a downstream task has become a standard method in NLP in the last few years. While the results from these models are impressive, applying them can be extremely computationally expensive, as is pretraining new models with the latest architectures. We present a novel method for applying pretrained transformer language models which lowers their memory requirement both at training and inference time. An additional benefit is that our method removes the fixed context size constraint that most transformer models have, allowing for more flexible use. When applied to the GPT-2 language model, we find that our method attains better perplexity than an unmodified GPT-2 model on the PG-19 and WikiText-103 corpora, for a given amount of computation or memory.

Shubham Toshniwal, Allyson Ettinger, Kevin Gimpel et al.

We propose PeTra, a memory-augmented neural network designed to track entities in its memory slots. PeTra is trained using sparse annotation from the GAP pronoun resolution dataset and outperforms a prior memory model on the task while using a simpler architecture. We empirically compare key modeling choices, finding that we can simplify several aspects of the design of the memory module while retaining strong performance. To measure the people tracking capability of memory models, we (a) propose a new diagnostic evaluation based on counting the number of unique entities in text, and (b) conduct a small scale human evaluation to compare evidence of people tracking in the memory logs of PeTra relative to a previous approach. PeTra is highly effective in both evaluations, demonstrating its ability to track people in its memory despite being trained with limited annotation.

Josef Klafka, Allyson Ettinger

Although models using contextual word embeddings have achieved state-of-the-art results on a host of NLP tasks, little is known about exactly what information these embeddings encode about the context words that they are understood to reflect. To address this question, we introduce a suite of probing tasks that enable fine-grained testing of contextual embeddings for encoding of information about surrounding words. We apply these tasks to examine the popular BERT, ELMo and GPT contextual encoders, and find that each of our tested information types is indeed encoded as contextual information across tokens, often with near-perfect recoverability-but the encoders vary in which features they distribute to which tokens, how nuanced their distributions are, and how robust the encoding of each feature is to distance. We discuss implications of these results for how different types of models breakdown and prioritize word-level context information when constructing token embeddings.

Allyson Ettinger

Pre-training by language modeling has become a popular and successful approach to NLP tasks, but we have yet to understand exactly what linguistic capacities these pre-training processes confer upon models. In this paper we introduce a suite of diagnostics drawn from human language experiments, which allow us to ask targeted questions about the information used by language models for generating predictions in context. As a case study, we apply these diagnostics to the popular BERT model, finding that it can generally distinguish good from bad completions involving shared category or role reversal, albeit with less sensitivity than humans, and it robustly retrieves noun hypernyms, but it struggles with challenging inferences and role-based event prediction -- and in particular, it shows clear insensitivity to the contextual impacts of negation.

Allyson Ettinger, Ahmed Elgohary, Colin Phillips et al.

An important component of achieving language understanding is mastering the composition of sentence meaning, but an immediate challenge to solving this problem is the opacity of sentence vector representations produced by current neural sentence composition models. We present a method to address this challenge, developing tasks that directly target compositional meaning information in sentence vector representations with a high degree of precision and control. To enable the creation of these controlled tasks, we introduce a specialized sentence generation system that produces large, annotated sentence sets meeting specified syntactic, semantic and lexical constraints. We describe the details of the method and generation system, and then present results of experiments applying our method to probe for compositional information in embeddings from a number of existing sentence composition models. We find that the method is able to extract useful information about the differing capacities of these models, and we discuss the implications of our results with respect to these systems' capturing of sentence information. We make available for public use the datasets used for these experiments, as well as the generation system.

Allyson Ettinger, Sudha Rao, Hal Daumé et al.

This paper presents a summary of the first Workshop on Building Linguistically Generalizable Natural Language Processing Systems, and the associated Build It Break It, The Language Edition shared task. The goal of this workshop was to bring together researchers in NLP and linguistics with a shared task aimed at testing the generalizability of NLP systems beyond the distributions of their training data. We describe the motivation, setup, and participation of the shared task, provide discussion of some highlighted results, and discuss lessons learned.