arXiv subjects

Janet B. Pierrehumbert

Publications and source records attributed to Janet B. Pierrehumbert.

27 records · Page 2Linked to original sources

arXiv2022

Two Contrasting Data Annotation Paradigms for Subjective NLP Tasks

Labelled data is the foundation of most natural language processing tasks. However, labelling data is difficult and there often are diverse valid beliefs about what the correct data labels should be. So far, dataset creators have acknowledged annotator subjectivity, but rarely actively managed it in the annotation process. This has led to partly-subjective datasets that fail to serve a clear downstream use. To address this issue, we propose two contrasting paradigms for data annotation. The descriptive paradigm encourages annotator subjectivity, whereas the prescriptive paradigm discourages it. Descriptive annotation allows for the surveying and modelling of different beliefs, whereas prescriptive annotation enables the training of models that consistently apply one belief. We discuss benefits and challenges in implementing both paradigms, and argue that dataset creators should explicitly aim for one or the other to facilitate the intended use of their dataset. Lastly, we conduct an annotation experiment using hate speech data that illustrates the contrast between the two paradigms.

cs.CL↗

arXiv2021

Temporal Adaptation of BERT and Performance on Downstream Document Classification: Insights from Social Media

Language use differs between domains and even within a domain, language use changes over time. For pre-trained language models like BERT, domain adaptation through continued pre-training has been shown to improve performance on in-domain downstream tasks. In this article, we investigate whether temporal adaptation can bring additional benefits. For this purpose, we introduce a corpus of social media comments sampled over three years. It contains unlabelled data for adaptation and evaluation on an upstream masked language modelling task as well as labelled data for fine-tuning and evaluation on a downstream document classification task. We find that temporality matters for both tasks: temporal adaptation improves upstream and temporal fine-tuning downstream task performance. Time-specific models generally perform better on past than on future test sets, which matches evidence on the bursty usage of topical words. However, adapting BERT to time and domain does not improve performance on the downstream task over only adapting to domain. Token-level analysis shows that temporal adaptation captures event-driven changes in language use in the downstream task, but not those changes that are actually relevant to task performance. Based on our findings, we discuss when temporal adaptation may be more effective.

cs.CL↗

arXiv2021

Dynamic Contextualized Word Embeddings

Static word embeddings that represent words by a single vector cannot capture the variability of word meaning in different linguistic and extralinguistic contexts. Building on prior work on contextualized and dynamic word embeddings, we introduce dynamic contextualized word embeddings that represent words as a function of both linguistic and extralinguistic context. Based on a pretrained language model (PLM), dynamic contextualized word embeddings model time and social space jointly, which makes them attractive for a range of NLP tasks involving semantic variability. We highlight potential application scenarios by means of qualitative and quantitative analyses on four English datasets.

cs.CL↗

arXiv2021

Superbizarre Is Not Superb: Derivational Morphology Improves BERT's Interpretation of Complex Words

How does the input segmentation of pretrained language models (PLMs) affect their interpretations of complex words? We present the first study investigating this question, taking BERT as the example PLM and focusing on its semantic representations of English derivatives. We show that PLMs can be interpreted as serial dual-route models, i.e., the meanings of complex words are either stored or else need to be computed from the subwords, which implies that maximally meaningful input tokens should allow for the best generalization on new words. This hypothesis is confirmed by a series of semantic probing tasks on which DelBERT (Derivation leveraging BERT), a model with derivational input segmentation, substantially outperforms BERT with WordPiece segmentation. Our results suggest that the generalization capabilities of PLMs could be further improved if a morphologically-informed vocabulary of input tokens were used.

cs.CL↗

arXiv2021

HateCheck: Functional Tests for Hate Speech Detection Models

Detecting online hate is a difficult task that even state-of-the-art models struggle with. Typically, hate speech detection models are evaluated by measuring their performance on held-out test data using metrics such as accuracy and F1 score. However, this approach makes it difficult to identify specific model weak points. It also risks overestimating generalisable model performance due to increasingly well-evidenced systematic gaps and biases in hate speech datasets. To enable more targeted diagnostic insights, we introduce HateCheck, a suite of functional tests for hate speech detection models. We specify 29 model functionalities motivated by a review of previous research and a series of interviews with civil society stakeholders. We craft test cases for each functionality and validate their quality through a structured annotation process. To illustrate HateCheck's utility, we test near-state-of-the-art transformer models as well as two popular commercial models, revealing critical model weaknesses.

cs.CL↗

arXiv2020

DagoBERT: Generating Derivational Morphology with a Pretrained Language Model

Can pretrained language models (PLMs) generate derivationally complex words? We present the first study investigating this question, taking BERT as the example PLM. We examine BERT's derivational capabilities in different settings, ranging from using the unmodified pretrained model to full finetuning. Our best model, DagoBERT (Derivationally and generatively optimized BERT), clearly outperforms the previous state of the art in derivation generation (DG). Furthermore, our experiments show that the input segmentation crucially impacts BERT's derivational knowledge, suggesting that the performance of PLMs could be further improved if a morphologically informed vocabulary of units were used.

cs.CL↗

arXiv2014

A model of grassroots changes in linguistic systems

Linguistic norms emerge in human communities because people imitate each other. A shared linguistic system provides people with the benefits of shared knowledge and coordinated planning. Once norms are in place, why would they ever change? This question, echoing broad questions in the theory of social dynamics, has particular force in relation to language. By definition, an innovator is in the minority when the innovation first occurs. In some areas of social dynamics, important minorities can strongly influence the majority through their power, fame, or use of broadcast media. But most linguistic changes are grassroots developments that originate with ordinary people. Here, we develop a novel model of communicative behavior in communities, and identify a mechanism for arbitrary innovations by ordinary people to have a good chance of being widely adopted. To imitate each other, people must form a mental representation of what other people do. Each time they speak, they must also decide which form to produce themselves. We introduce a new decision function that enables us to smoothly explore the space between two types of behavior: probability matching (matching the probabilities of incoming experience) and regularization (producing some forms disproportionately often). Using Monte Carlo methods, we explore the interactions amongst the degree of regularization, the distribution of biases in a network, and the network position of the innovator. We identify two regimes for the widespread adoption of arbritrary innovations, viewed as informational cascades in the network. With moderate regularization of experienced input, average people (not well-connected people) are the most likely source of successful innovations. Our results shed light on a major outstanding puzzle in the theory of language change. The framework also holds promise for understanding the dynamics of other social norms.

cs.CL↗

arXiv2011

Niche as a determinant of word fate in online groups

Patterns of word use both reflect and influence a myriad of human activities and interactions. Like other entities that are reproduced and evolve, words rise or decline depending upon a complex interplay between {their intrinsic properties and the environments in which they function}. Using Internet discussion communities as model systems, we define the concept of a word niche as the relationship between the word and the characteristic features of the environments in which it is used. We develop a method to quantify two important aspects of the size of the word niche: the range of individuals using the word and the range of topics it is used to discuss. Controlling for word frequency, we show that these aspects of the word niche are strong determinants of changes in word frequency. Previous studies have already indicated that word frequency itself is a correlate of word success at historical time scales. Our analysis of changes in word frequencies over time reveals that the relative sizes of word niches are far more important than word frequencies in the dynamics of the entire vocabulary at shorter time scales, as the language adapts to new concepts and social groupings. We also distinguish endogenous versus exogenous factors as additional contributors to the fates of words, and demonstrate the force of this distinction in the rise of novel words. Our results indicate that short-term nonstationarity in word statistics is strongly driven by individual proclivities, including inclinations to provide novel information and to project a distinctive social identity.

cs.CL↗

arXiv2009

Beyond word frequency: Bursts, lulls, and scaling in the temporal distributions of words

Background: Zipf's discovery that word frequency distributions obey a power law established parallels between biological and physical processes, and language, laying the groundwork for a complex systems perspective on human communication. More recent research has also identified scaling regularities in the dynamics underlying the successive occurrences of events, suggesting the possibility of similar findings for language as well. Methodology/Principal Findings: By considering frequent words in USENET discussion groups and in disparate databases where the language has different levels of formality, here we show that the distributions of distances between successive occurrences of the same word display bursty deviations from a Poisson process and are well characterized by a stretched exponential (Weibull) scaling. The extent of this deviation depends strongly on semantic type -- a measure of the logicality of each word -- and less strongly on frequency. We develop a generative model of this behavior that fully determines the dynamics of word usage. Conclusions/Significance: Recurrence patterns of words are well described by a stretched exponential distribution of recurrence times, an empirical scaling that cannot be anticipated from Zipf's law. Because the use of words provides a uniquely precise and powerful lens on human thought and activity, our findings also have implications for other overt manifestations of collective human dynamics.

cs.CL↗