Natural Language Inference (NLI) has been a cornerstone task in evaluating language models’ inferential reasoning capabilities. However, the standard three-way classification scheme used in NLI has well-known shortcomings in evaluating models’ ability to capture the nuances of natural human reasoning. In this paper, we argue that the operationalization of the neutral label in current NLI datasets has low validity, is interpreted inconsistently, and that at least one important sense of neutrality is often ignored. We uncover the detrimental impact of these shortcomings, which in some cases leads to annotation datasets that actually decrease performance on downstream tasks. We compare approaches of handling annotator disagreement and identify flaws in a recent NLI dataset that designs an annotator study based on a problematic operationalization. Our findings highlight the need for a more refined evaluation framework for NLI, and we hope to spark further discussion and action in the NLP community.

Writing high-quality test questions (items) is critical to building educational measures but has traditionally also been a time-consuming process. One promising avenue for alleviating this is automated item generation, whereby methods from artificial intelligence (AI) are used to generate new items with minimal human intervention. Researchers have explored using large language models (LLMs) to generate new items with equivalent psychometric properties to human-written ones. But can LLMs generate items with improved psychometric properties, even when existing items have poor validity evidence? We investigate this using items from a natural language inference (NLI) dataset. We develop a novel prompting strategy based on selecting items with both the best and worst properties to use in the prompt and use GPT-3 to generate new NLI items. We find that the GPT-3 items show improved psychometric properties in many cases, whilst also possessing good content, convergent and discriminant validity evidence. Collectively, our results demonstrate the potential of employing LLMs to ease the item development process and suggest that the careful use of prompting may allow for iterative improvement of item quality.

Reasoning using negation is known to be difficult for transformer-based language models. While previous studies have used the tools of psycholinguistics to probe a transformer’s ability to reason over negation, none have focused on the types of negation studied in developmental psychology. We explore how well transformers can process such categories of negation, by framing the problem as a natural language inference (NLI) task. We curate a set of diagnostic questions for our target categories from popular NLI datasets and evaluate how well a suite of models reason over them. We find that models perform consistently better only on certain categories, suggesting clear distinctions in how they are processed.

Transformer-based language models (LMs) continue to advance state-of-the-art performance on NLP benchmark tasks, including tasks designed to mimic human-inspired “commonsense” competencies. To better understand the degree to which LMs can be said to have certain linguistic reasoning skills, researchers are beginning to adapt the tools and concepts of the field of psychometrics. But to what extent can the benefits flow in the other direction? I.e., can LMs be of use in predicting what the psychometric properties of test items will be when those items are given to human participants? We gather responses from numerous human participants and LMs (transformer- and non-transformer-based) on a broad diagnostic test of linguistic competencies. We then use the responses to calculate standard psychometric properties of the items in the diagnostic test, using the human responses and the LM responses separately. We then determine how well these two sets of predictions match. We find cases in which transformer-based LMs predict psychometric properties consistently well in certain categories but consistently poorly in others, thus providing new insights into fundamental similarities and differences between human and LM reasoning.

Despite the increasing number of large and comprehensive machine translation (MT) systems, evaluation of these methods in various languages has been restrained by the lack of high-quality parallel corpora as well as engagement with the people that speak these languages. In this study, we present an evaluation of state-of-the-art approaches to training and evaluating MT systems in 22 languages from the Turkic language family, most of which being extremely under-explored. First, we adopt the TIL Corpus with a few key improvements to the training and the evaluation sets. Then, we train 26 bilingual baselines as well as a multi-way neural MT (MNMT) model using the corpus and perform an extensive analysis using automatic metrics as well as human evaluations. We find that the MNMT model outperforms almost all bilingual baselines in the out-of-domain test sets and finetuning the model on a downstream task of a single pair also results in a huge performance boost in both low- and high-resource scenarios. Our attentive analysis of evaluation criteria for MT models in Turkic languages also points to the necessity for further research in this direction. We release the corpus splits, test sets as well as models to the public.

If two sentences have the same meaning, it should follow that they are equivalent in their inferential properties, i.e., each sentence should textually entail the other. However, many paraphrase datasets currently in widespread use rely on a sense of paraphrase based on word overlap and syntax. Can we teach them instead to identify paraphrases in a way that draws on the inferential properties of the sentences, and is not over-reliant on lexical and syntactic similarities of a sentence pair? We apply the adversarial paradigm to this question, and introduce a new adversarial method of dataset creation for paraphrase identification: the Adversarial Paraphrasing Task (APT), which asks participants to generate semantically equivalent (in the sense of mutually implicative) but lexically and syntactically disparate paraphrases. These sentence pairs can then be used both to test paraphrase identification models (which get barely random accuracy) and then improve their performance. To accelerate dataset generation, we explore automation of APT using T5, and show that the resulting dataset also improves accuracy. We discuss implications for paraphrase detection and release our dataset in the hope of making paraphrase detection models better able to detect sentence-level meaning equivalence.

Curriculum learning, a training strategy where training data are ordered based on their difficulty, has been shown to improve performance and reduce training time on various NLP tasks. While much work over the years has developed novel approaches for generating curricula, these strategies are typically only suited for the task they were designed for. This work explores developing a task-agnostic model for problem difficulty and applying it to the Stanford Natural Language Inference (SNLI) dataset. Using the human responses that come with the dev set of SNLI, we train both regression and classification models to predict how many annotators will answer a question correctly and then project the difficulty estimates onto the full SNLI train set to create the curriculum. We argue that our curriculum is effectively capturing difficulty for this task through various analyses of both the model and the predicted difficulty scores.