The perceived toxicity of language can vary based on someone's identity and beliefs, but this variation is often ignored when collecting toxic language datasets, resulting in dataset and model biases. We seek to understand the who, why, and what behind biases in toxicity annotations. In two online studies with demographically and politically diverse participants, we investigate the effect of annotator identities (who) and beliefs (why), drawing from social psychology research about hate speech, free speech, racist beliefs, political leaning, and more. We disentangle what is annotated as toxic by considering posts with three characteristics: anti-Black language, African American English (AAE) dialect, and vulgarity. Our results show strong associations between annotator identity and beliefs and their ratings of toxicity. Notably, more conservative annotators and those who scored highly on our scale for racist beliefs were less likely to rate anti-Black language as toxic, but more likely to rate AAE as toxic. We additionally present a case study illustrating how a popular toxicity detection system's ratings inherently reflect only specific beliefs and perspectives. Our findings call for contextualizing toxicity labels in social variables, which raises immense implications for toxic language annotation and detection.
When an NLP model is trained on text data from one time period and tested or deployed on data from another, the resulting temporal misalignment can degrade end-task performance. In this work, we establish a suite of eight diverse tasks across different domains (social media, science papers, news, and reviews) and periods of time (spanning five years or more) to quantify the effects of temporal misalignment. Our study is focused on the ubiquitous setting where a pretrained model is optionally adapted through continued domain-specific pretraining, followed by task-specific finetuning. We establish a suite of tasks across multiple domains to study temporal misalignment in modern NLP systems. We find stronger effects of temporal misalignment on task performance than have been previously reported. We also find that, while temporal adaptation through continued pretraining can help, these gains are small compared to task-specific finetuning on data from the target time period. Our findings motivate continued research to improve temporal robustness of NLP models.
Transformer architectures have achieved state-of-the-art results on a variety of sequence modeling tasks. However, their attention mechanism comes with a quadratic complexity in sequence lengths, making the computational overhead prohibitive, especially for long sequences. Attention context can be seen as a random-access memory with each token taking a slot. Under this perspective, the memory size grows linearly with the sequence length, and so does the overhead of reading from it. One way to improve the efficiency is to bound the memory size. We show that disparate approaches can be subsumed into one abstraction, attention with bounded-memory control (ABC), and they vary in their organization of the memory. ABC reveals new, unexplored possibilities. First, it connects several efficient attention variants that would otherwise seem apart. Second, this abstraction gives new insights--an established approach (Wang et al., 2020b) previously thought to be not applicable in causal attention, actually is. Last, we present a new instance of ABC, which draws inspiration from existing ABC approaches, but replaces their heuristic memory-organizing functions with a learned, contextualized one. Our experiments on language modeling, machine translation, and masked language model finetuning show that our approach outperforms previous efficient attention models; compared to the strong transformer baselines, it significantly improves the inference time and space efficiency with no or negligible accuracy loss.
Research in NLP is often supported by experimental results, and improved reporting of such results can lead to better understanding and more reproducible science. In this paper we analyze three statistical estimators for expected validation performance, a tool used for reporting performance (e.g., accuracy) as a function of computational budget (e.g., number of hyperparameter tuning experiments). Where previous work analyzing such estimators focused on the bias, we also examine the variance and mean squared error (MSE). In both synthetic and realistic scenarios, we evaluate three estimators and find the unbiased estimator has the highest variance, and the estimator with the smallest variance has the largest bias; the estimator with the smallest MSE strikes a balance between bias and variance, displaying a classic bias-variance tradeoff. We use expected validation performance to compare between different models, and analyze how frequently each estimator leads to drawing incorrect conclusions about which of two models performs best. We find that the two biased estimators lead to the fewest incorrect conclusions, which hints at the importance of minimizing variance and MSE.
Representation learning for text via pretraining a language model on a large corpus has become a standard starting point for building NLP systems. This approach stands in contrast to autoencoders, also trained on raw text, but with the objective of learning to encode each input as a vector that allows full reconstruction. Autoencoders are attractive because of their latent space structure and generative properties. We therefore explore the construction of a sentence-level autoencoder from a pretrained, frozen transformer language model. We adapt the masked language modeling objective as a generative, denoising one, while only training a sentence bottleneck and a single-layer modified transformer decoder. We demonstrate that the sentence representations discovered by our model achieve better quality than previous methods that extract representations from pretrained transformers on text similarity tasks, style transfer (an example of controlled generation), and single-sentence classification tasks in the GLUE benchmark, while using fewer parameters than large pretrained models.
Since the introduction of the transformer model by Vaswani et al. (2017), a fundamental question remains open: how to achieve extrapolation at inference time to longer sequences than seen during training? We first show that extrapolation can be improved by changing the position representation method, though we find that existing proposals do not allow efficient extrapolation. We introduce a simple and efficient method, Attention with Linear Biases (ALiBi), that allows for extrapolation. ALiBi does not add positional embeddings to the word embeddings; instead, it biases the query-key attention scores with a term that is proportional to their distance. We show that this method allows training a 1.3 billion parameter model on input sequences of length 1024 that extrapolates to input sequences of length 2048, achieving the same perplexity as a sinusoidal position embedding model trained on inputs of length 2048, 11% faster and using 11% less memory. ALiBi's inductive bias towards recency allows it to outperform multiple strong position methods on the WikiText-103 benchmark. Finally, we provide analysis of ALiBi to understand why it leads to better performance.
We introduce a new domain expert mixture (DEMix) layer that enables conditioning a language model (LM) on the domain of the input text. A DEMix layer is a collection of expert feedforward networks, each specialized to a domain, that makes the LM modular: experts can be mixed, added or removed after initial training. Extensive experiments with autoregressive transformer LMs (up to 1.3B parameters) show that DEMix layers reduce test-time perplexity, increase training efficiency, and enable rapid adaptation with little overhead. We show that mixing experts during inference, using a parameter-free weighted ensemble, allows the model to better generalize to heterogeneous or unseen domains. We also show that experts can be added to iteratively incorporate new domains without forgetting older ones, and that experts can be removed to restrict access to unwanted domains, without additional training. Overall, these results demonstrate benefits of explicitly conditioning on textual domains during language modeling.
Human evaluations are typically considered the gold standard in natural language generation, but as models' fluency improves, how well can evaluators detect and judge machine-generated text? We run a study assessing non-experts' ability to distinguish between human- and machine-authored text (GPT2 and GPT3) in three domains (stories, news articles, and recipes). We find that, without training, evaluators distinguished between GPT3- and human-authored text at random chance level. We explore three approaches for quickly training evaluators to better identify GPT3-authored text (detailed instructions, annotated examples, and paired examples) and find that while evaluators' accuracy improved up to 55%, it did not significantly improve across the three domains. Given the inconsistent results across text domains and the often contradictory reasons evaluators gave for their judgments, we examine the role untrained human evaluations play in NLG evaluation and provide recommendations to NLG researchers for improving human evaluations of text generated from state-of-the-art models.
Modern neural text generation systems can produce remarkably fluent and grammatical texts. While earlier language models suffered from repetition and syntactic errors, the errors made by contemporary models are often semantic, narrative, or discourse failures. To facilitate research of these complex error types, we introduce a new structured, crowdsourced error annotation schema called Scarecrow. The error categories used in Scarecrow -- such as redundancy, commonsense errors, and incoherence -- were identified by combining expert analysis with several pilot rounds of ontology-free crowd annotation to arrive at a schema which covers the error phenomena found in real machine generated text. We use Scarecrow to collect 13k annotations of 1.3k human and machine generate paragraphs of English language news text, amounting to over 41k spans each labeled with its error category, severity, a natural language explanation, and antecedent span (where relevant). We collect annotations for text generated by state-of-the-art systems with varying known performance levels, from GPT-2 Small through the largest GPT-3. We isolate several factors for detailed analysis, including parameter count, training data, and decoding technique. Our results show both expected and surprising differences across these settings. These findings demonstrate the value of Scarecrow annotations in the assessment of current and future text generation systems. We release our complete annotation toolkit and dataset at https://yao-dou.github.io/scarecrow/.