Large language models (LMs) such as ChatGPT have revolutionized natural language processing and artificial intelligence more broadly. In this work, we discuss our research on understanding and advancing these models, centered around how they use the very large text corpora they are trained on. First, we describe our efforts to understand how these models learn to perform new tasks after training, demonstrating that their so-called in-context learning capabilities are almost entirely determined by what they learn from the training data. Next, we introduce a new class of LMs—nonparametric LMs—that repurpose this training data as a data store from which they retrieve information for improved accuracy and updatability. We discuss our work establishing the foundations of such models, including one of the first broadly used neural retrieval models and an approach that simplifies a traditional, two-stage pipeline into one. We also discuss how nonparametric models open up new avenues for responsible data use, e.g., by segregating permissive and copyrighted text and using them differently. Finally, we envision the next generation of LMs we should build, focusing on efficient scaling, improved factuality, and decentralization.

Tokenization is the first step in modern neural language model pipelines where an input text is converted to a sequence of subword tokens. We introduce from first principles a finite-state transduction framework that can encode all possible tokenizations of a regular language. We then constructively show that Byte-Pair Encoding (BPE) and MaxMatch (WordPiece), two popular tokenization schemes, are also efficiently representable by simple finite-state transducers. For BPE, this is particularly surprising given that it does not tokenize strings from left to right and requires a notion of priority. We also discuss an application of subword-level pattern promotion to guided generation, where the outputs of a language model are constrained to match a specified pattern, and how tokenization-aware promotion offers a theoretical benefit to modeling.

Natural language processing (NLP) models may leak private information in different ways, including membership inference, reconstruction, or attribute inference attacks. Sensitive information may not be explicit in the text, but hidden in underlying writing characteristics. Methods to protect privacy can involve using representations inside models that are demonstrated not to detect sensitive attributes or—for instance, in cases where users might be at risk from an untrustworthy model, the sort of scenario of interest here—changing the raw text before models can have access to it. The goal is to rewrite text to prevent someone from inferring a sensitive attribute (e.g., the gender of the author, or their location by the writing style) while keeping the text useful for its original intention (e.g., the sentiment of a product review). The few works tackling this have focused on generative techniques. However, these often create extensively different texts from the original ones or face problems such as mode collapse. This article explores a novel adaptation of adversarial attack techniques to manipulate a text to deceive a classifier w.r.t. one task (privacy) while keeping the predictions of another classifier trained for another task (utility) unchanged. We propose IDT, a method that analyses predictions made by auxiliary and interpretable models to identify which tokens are important to change for the privacy task, and which ones should be kept for the utility task. We evaluate different datasets for NLP suitable for different tasks. Automatic and human evaluations show that IDT retains the utility of text, while also outperforming existing methods when deceiving a classifier w.r.t. a privacy task.

In many fields, such as language acquisition, neuropsychology of language, the study of aging, and historical linguistics, corpora are used for estimating the diversity of grammatical structures that are produced during a period by an individual, community, or type of speakers. In these cases, treebanks are taken as representative samples of the syntactic structures that might be encountered. Generalizing the potential syntactic diversity from the structures documented in a small corpus requires careful extrapolation whose accuracy is constrained by the limited size of representative sub-corpora. In this article, I demonstrate—both theoretically and empirically—that a grammar’s derivational entropy and the mean length of the utterances (MLU) it generates are fundamentally linked, giving rise to a new measure, the derivational entropy rate. The mean length of utterances becomes the most practical index of syntactic complexity; I demonstrate that MLU is not a mere proxy, but a fundamental measure of syntactic diversity. In combination with the new derivational entropy rate measure, it provides a theory-free assessment of grammatical complexity. The derivational entropy rate indexes the rate at which different grammatical annotation frameworks determine the grammatical complexity of treebanks. I evaluate the Smoothed Induced Treebank Entropy (SITE) as a tool for estimating these measures accurately, even from very small treebanks. I conclude by discussing important implications of these results for both NLP and human language processing.

We present Local Graph-based Dictionary Expansion (LGDE), a method for data-driven discovery of the semantic neighborhood of words using tools from manifold learning and network science. At the heart of LGDE lies the creation of a word similarity graph from the geometry of word embeddings followed by local community detection based on graph diffusion. The diffusion in the local graph manifold allows the exploration of the complex nonlinear geometry of word embeddings to capture word similarities based on paths of semantic association, over and above direct pairwise similarities. Exploiting such semantic neighborhoods enables the expansion of dictionaries of pre-selected keywords, an important step for tasks in information retrieval, such as database queries and online data collection. We validate LGDE on two user-generated English-language corpora and show that LGDE enriches the list of keywords with improved performance relative to methods based on direct word similarities or co-occurrences. We further demonstrate our method through a real-world use case from communication science, where LGDE is evaluated quantitatively on the expansion of a conspiracy-related dictionary from online data collected and analyzed by domain experts. Our empirical results and expert user assessment indicate that LGDE expands the seed dictionary with more useful keywords due to the manifold-learning-based similarity network.

We present a corpus of 8,400 Dutch sentence pairs, intended primarily for the grammatical evaluation of language models. Each pair consists of a grammatical sentence and a minimally different ungrammatical sentence. The corpus covers 84 paradigms, classified into 22 syntactic phenomena. Ten sentence pairs of each paradigm were created by hand, while the remaining 90 were generated semi-automatically and manually validated afterwards. Nine of the 10 hand-crafted sentences of each paradigm are rated for acceptability by at least 30 participants each, and for the same 9 sentences reading times are recorded per word, through self-paced reading. Here, we report on the construction of the dataset, the measured acceptability ratings and reading times, as well as the extent to which a variety of language models can be used to predict both the ground-truth grammaticality and human acceptability ratings.

Accurate temporal expression normalization, the process of assigning a numerical value to a temporal expression, is essential for tasks such as timeline creation and temporal reasoning. While rule-based normalization systems are limited in adaptability across different domains and languages, deep-learning solutions in this area have not been extensively explored. An additional challenge is the scarcity of manually annotated corpora with temporal annotations. To address the adaptability limitations of current systems, we propose a highly adaptable methodology that can be applied to multiple domains and languages. This can be achieved by leveraging a multilingual Pre-trained Language Model (PTLM) with a fill-mask architecture, using a Value Intermediate Representation (VIR) where the temporal expression value format is adjusted to the fill-mask representation. Our approach involves a two-phase training process. Initially, the model is trained with a novel masking policy on a large English biomedical corpus that is automatically annotated with normalized temporal expressions, along with a complementary hand-crafted temporal expressions corpus. This addresses the lack of manually annotated data and helps to achieve sufficient capacity for adaptation to diverse domains or languages. In the second phase, we show how the model can be tailored to different domains and languages using various techniques, showcasing the versatility of the proposed methodology. This approach significantly outperforms existing systems.

While there is a widespread belief that artificial general intelligence—or even superhuman AI—is imminent, complex problems in expert domains are far from being solved. We argue that such problems require human–AI cooperation and that the current state of the art in generative AI is unable to play the role of a reliable partner due to a multitude of shortcomings, including difficulty in keeping track of a complex solution artifact (e.g., a software program), limited support for versatile human preference expression, and lack of adapting to human preference in an interactive setting. To address these challenges, we propose HAI-Co2, a novel human–AI co-construction framework. We take first steps towards a formalization of HAI-Co2 and discuss the difficult open research problems that it faces.

While large language models (LLMs) have demonstrated remarkable capabilities across a range of downstream tasks, a significant concern revolves around their propensity to exhibit hallucinations: LLMs occasionally generate content that diverges from the user input, contradicts previously generated context, or misaligns with established world knowledge. This phenomenon poses a substantial challenge to the reliability of LLMs in real-world scenarios. In this article, we survey recent efforts on the detection, explanation, and mitigation of hallucination, with an emphasis on the unique challenges posed by LLMs. We present taxonomies of the LLM hallucination phenomena and evaluation benchmarks, analyze existing approaches aiming at mitigating LLM hallucination, and discuss potential directions for future research.

The ACL community has very little interest in evaluating the real-world impact of NLP systems. A structured survey of the ACL Anthology shows that perhaps 0.1% of its papers contain such evaluations; furthermore most papers that include impact evaluations present them very sketchily and instead focus on metric evaluations. NLP technology would be more useful and more quickly adopted if we seriously tried to understand and evaluate its real-world impact.