We introduce The Benchmark of Linguistic Minimal Pairs (BLiMP), 1 a challenge set for evaluating the linguistic knowledge of language models (LMs) on major grammatical phenomena in English. BLiMP consists of 67 individual datasets, each containing 1,000 minimal pairs—that is, pairs of minimally different sentences that contrast in grammatical acceptability and isolate specific phenomenon in syntax, morphology, or semantics. We generate the data according to linguist-crafted grammar templates, and human aggregate agreement with the labels is 96.4%. We evaluate n-gram, LSTM, and Transformer (GPT-2 and Transformer-XL) LMs by observing whether they assign a higher probability to the acceptable sentence in each minimal pair. We find that state-of-the-art models identify morphological contrasts related to agreement reliably, but they struggle with some subtle semantic and syntactic phenomena, such as negative polarity items and extraction islands.

Though state-of-the-art sentence representation models can perform tasks requiring significant knowledge of grammar, it is an open question how best to evaluate their grammatical knowledge. We explore five experimental methods inspired by prior work evaluating pretrained sentence representation models. We use a single linguistic phenomenon, negative polarity item (NPI) licensing in English, as a case study for our experiments. NPIs like any are grammatical only if they appear in a licensing environment like negation (Sue doesn't have any cats vs. *Sue has any cats). This phenomenon is challenging because of the variety of NPI licensing environments that exist. We introduce an artificially generated dataset that manipulates key features of NPI licensing for the experiments. We find that BERT has significant knowledge of these features, but its success varies widely across different experimental methods. We conclude that a variety of methods is necessary to reveal all relevant aspects of a model's grammatical knowledge in a given domain. 1 Other prominent theories of NPI licensing are based on notions of non-veridicality (

It is well documented that NLP models learn social biases, but little work has been done on how these biases manifest in model outputs for applied tasks like question answering (QA). We introduce the Bias Benchmark for QA (BBQ), a dataset of question sets constructed by the authors that highlight attested social biases against people belonging to protected classes along nine social dimensions relevant for U.S. English-speaking contexts. Our task evaluates model responses at two levels: (i) given an under-informative context, we test how strongly responses refect social biases, and (ii) given an adequately informative context, we test whether the model's biases override a correct answer choice. We fnd that models often rely on stereotypes when the context is under-informative, meaning the model's outputs consistently reproduce harmful biases in this setting. Though models are more accurate when the context provides an informative answer, they still rely on stereotypes and average up to 3.4 percentage points higher accuracy when the correct answer aligns with a social bias than when it conficts, with this difference widening to over 5 points on examples targeting gender for most models tested.

No abstract

The relationship between syntactic, semantic, and conceptual processes in language comprehension is a central question to the neurobiology of language. Several studies have suggested that conceptual combination in particular can be localized to the left anterior temporal lobe (LATL), while syntactic processes are more often associated with the posterior temporal lobe or inferior frontal gyrus. However, LATL activity can also correlate with syntactic computations, particularly in narrative comprehension. Here we investigated the degree to which LATL conceptual combination is dependent on syntax, specifically asking whether rapid (∼200 ms) magnetoencephalography effects of conceptual combination in the LATL can occur in the absence of licit syntactic phrase closure and in the absence of a semantically plausible output for the composition. We find that such effects do occur: LATL effects of conceptual combination were observed even when there was no syntactic phrase closure or plausible meaning. But syntactic closure did have an additive effect such that LATL signals were the highest for expressions that composed both conceptually and syntactically. Our findings conform to an account in which LATL conceptual composition is influenced by local syntactic composition but is also able to operate without it.