Knowledge claims are abundant in the literature on large language models (LLMs); but can we say that GPT-4 truly “knows” the Earth is round? To address this question, we review standard definitions of knowledge in epistemology and we formalize interpretations applicable to LLMs. In doing so, we identify inconsistencies and gaps in how current NLP research conceptualizes knowledge with respect to epistemological frameworks. Additionally, we conduct a survey of 100 professional philosophers and computer scientists to compare their preferences in knowledge definitions and their views on whether LLMs can really be said to know. Finally, we suggest evaluation protocols for testing knowledge in accordance to the most relevant definitions.

Facts are subject to contingencies and can be true or false in different circumstances. One such contingency is time, wherein some facts mutate over a given period, e.g., the president of a country or the winner of a championship. Trustworthy language models ideally identify mutable facts as such and process them accordingly. We create MuLan, a benchmark for evaluating the ability of English language models to anticipate time-contingency, covering both 1:1 and 1:N relations. We hypothesize that mutable facts are encoded differently than immutable ones, hence being easier to update. In a detailed evaluation of six popular large language models, we consistently find differences in the LLMs’ confidence, representations, and update behavior, depending on the mutability of a fact. Our findings should inform future work on the injection of and induction of time-contingent knowledge to/from LLMs.

In this work, we conduct a detailed analysis on the performance of legal-oriented pre-trained language models (PLMs). We examine the interplay between their original objective, acquired knowledge, and legal language understanding capacities which we define as the upstream, probing, and downstream performance, respectively. We consider not only the models’ size but also the pre-training corpora used as important dimensions in our study. To this end, we release a multinational English legal corpus (LeXFiles) and a legal knowledge probing benchmark (LegalLAMA) to facilitate training and detailed analysis of legal-oriented PLMs. We release two new legal PLMs trained on LeXFiles and evaluate them alongside others on LegalLAMA and LexGLUE. We find that probing performance strongly correlates with upstream performance in related legal topics. On the other hand, downstream performance is mainly driven by the model’s size and prior legal knowledge which can be estimated by upstream and probing performance. Based on these findings, we can conclude that both dimensions are important for those seeking the development of domain-specific PLMs.

In this paper, we introduce a new beam search algorithm that improves the generalization of neural generators to unseen examples, especially in low-resource data-to-text settings. Our algorithm aims to reduce the number of omissions and hallucinations during the decoding process. For this purpose, it relies on two regression models to explicitly characterize factual errors. We explain how to create a new dataset to train these models given an original training set of less than a thousand data points. We apply our approach in the low-resource, legal setting using the French Plum2Text dataset, as well as in English using WebNLG. We observe in our experiment that this combination improves the faithfulness of pre-trained neural text generators using both human and automatic evaluation. Moreover, our approach offers a level of interpretability by predicting the number of omissions and hallucinations present in a given generation with respect to the input data. Finally, we visualize our algorithm’s exploration of the hypothesis space at different steps during the decoding process.

In this paper, we introduce a new embedding-based metric relying on trainable ranking models to evaluate the semantic accuracy of neural data-to-text generators. This metric is especially well suited to semantically and factually assess the performance of a text generator when tables can be associated with multiple references and table values contain textual utterances. We first present how one can implement and further specialize the metric by training the underlying ranking models on a legal Data-to-Text dataset. We show how it may provide a more robust evaluation than other evaluation schemes in challenging settings using a dataset comprising paraphrases between the table values and their respective references. Finally, we evaluate its generalization capabilities on a well-known dataset, WebNLG, by comparing it with human evaluation and a recently introduced metric based on natural language inference. We then illustrate how it naturally characterizes, both quantitatively and qualitatively, omissions and hallucinations.

We hereby present our submission to the Shared Task in Evaluating Accuracy at the INLG 2021 Conference. Our evaluation protocol relies on three main components; rules and text classifiers that pre-annotate the dataset, a human annotator that validates the pre-annotations, and a web interface that facilitates this validation. Our submission consists in fact of two submissions; we first analyze solely the performance of the rules and classifiers (pre-annotations), and then the human evaluation aided by the former pre-annotations using the web interface (hybrid). The code for the web interface and the classifiers is publicly available.

In this paper, we reproduce the experiments of Artetxe et al. (2018b) regarding the robust self-learning method for fully unsupervised cross-lingual mappings of word embeddings. We show that the reproduction of their method is indeed feasible with some minor assumptions. We further investigate the robustness of their model by introducing four new languages that are less similar to English than the ones proposed by the original paper. In order to assess the stability of their model, we also conduct a grid search over sensible hyperparameters. We then propose key recommendations that apply to any research project in order to deliver fully reproducible research.

Plumitifs (dockets) were initially a tool for law clerks. Nowadays, they are used as summaries presenting all the steps of a judicial case. Information concerning parties’ identity, jurisdiction in charge of administering the case, and some information relating to the nature and the course of the preceding are available through plumitifs. They are publicly accessible but barely understandable; they are written using abbreviations and referring to provisions from the Criminal Code of Canada, which makes them hard to reason about. In this paper, we propose a simple yet efficient multi-source language generation architecture that leverages both the plumitif and the Criminal Code’s content to generate intelligible plumitifs descriptions. It goes without saying that ethical considerations rise with these sensitive documents made readable and available at scale, legitimate concerns that we address in this paper. This is, to the best of our knowledge, the first application of plumitifs descriptions generation made available for French speakers along with an ethical discussion about the topic.

We propose a novel way to handle out of vocabulary (OOV) words in downstream natural language processing (NLP) tasks. We implement a network that predicts useful embeddings for OOV words based on their morphology and on the context in which they appear. Our model also incorporates an attention mechanism indicating the focus allocated to the left context words, the right context words or the word’s characters, hence making the prediction more interpretable. The model is a “drop-in” module that is jointly trained with the downstream task’s neural network, thus producing embeddings specialized for the task at hand. When the task is mostly syntactical, we observe that our model aims most of its attention on surface form characters. On the other hand, for tasks more semantical, the network allocates more attention to the surrounding words. In all our tests, the module helps the network to achieve better performances in comparison to the use of simple random embeddings.