skip to main content


Title: Natural Language Inference with Mixed Effects
There is growing evidence that the prevalence of disagreement in the raw annotations used to construct natural language inference datasets makes the common practice of aggregating those annotations to a single label problematic. We propose a generic method that allows one to skip the aggregation step and train on the raw annotations directly without subjecting the model to unwanted noise that can arise from annotator response biases. We demonstrate that this method, which generalizes the notion of a mixed effects model by incorporating annotator random effects into any existing neural model, improves performance over models that do not incorporate such effects.  more » « less
Award ID(s):
1940981
NSF-PAR ID:
10300467
Author(s) / Creator(s):
; ;
Date Published:
Journal Name:
The 9th Joint Conference on Lexical and Computational Semantics (*SEM 2020)
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. null (Ed.)
    There is growing evidence that the prevalence of disagreement in the raw annotations used to construct natural language inference datasets makes the common practice of aggregating those annotations to a single label problematic. We propose a generic method that allows one to skip the aggregation step and train on the raw annotations directly without subjecting the model to unwanted noise that can arise from annotator response biases. We demonstrate that this method, which generalizes the notion of a mixed effects model by incorporating annotator random effects into any existing neural model, improves performance over models that do not incorporate such effects. 
    more » « less
  2. Hand-annotated data can vary due to factors such as subjective differences, intra-rater variability, and differing annotator expertise. We study annotations from different ex- perts who labelled the same behavior classes on a set of an- imal behavior videos, and observe a variation in annotation styles. We propose a new method using program synthesis to help interpret annotation differences for behavior analysis. Our model selects relevant trajectory features and learns a temporal filter as part of a program, which corresponds to estimated importance an annotator places on that feature at each timestamp. Our experiments on a dataset from behav- ioral neuroscience demonstrate that compared to baseline approaches, our method is more accurate at capturing an- notator labels and learns interpretable temporal filters. We believe that our method can lead to greater reproducibility of behavior annotations used in scientific studies. We plan to release our code. 
    more » « less
  3. As machine learning methods become more powerful and capture more nuances of human behavior, biases in the dataset can shape what the model learns and is evaluated on. This paper explores and attempts to quantify the uncertainties and biases due to annotator demographics when creating sentiment analysis datasets. We ask >1000 crowdworkers to provide their demographic information and annotations for multimodal sentiment data and its component modalities. We show that demographic differences among annotators impute a significant effect on their ratings, and that these effects also occur in each component modality. We compare predictions of different state-of-the-art multimodal machine learning algorithms against annotations provided by different demographic groups, and find that changing annotator demographics can cause >4.5 in accuracy difference when determining positive versus negative sentiment. Our findings underscore the importance of accounting for crowdworker attributes, such as demographics, when building datasets, evaluating algorithms, and interpreting results for sentiment analysis.

     
    more » « less
  4. As text generated by large language models proliferates, it becomes vital to understand how humans engage with such text, and whether or not they are able to detect when the text they are reading did not originate with a human writer. Prior work on human detection of generated text focuses on the case where an entire passage is either human-written or machine-generated. In this paper, we study a more realistic setting where text begins as human-written and transitions to being generated by state-of-the-art neural language models. We show that, while annotators often struggle at this task, there is substantial variance in annotator skill and that given proper incentives, annotators can improve at this task over time. Furthermore, we conduct a detailed comparison study and analyze how a variety of variables (model size, decoding strategy, fine-tuning, prompt genre, etc.) affect human detection performance. Finally, we collect error annotations from our participants and use them to show that certain textual genres influence models to make different types of errors and that certain sentence-level features correlate highly with annotator selection. We release the RoFT dataset: a collection of over 21,000 human annotations paired with error classifications to encourage future work in human detection and evaluation of generated text. 
    more » « less
  5. As text generated by large language models proliferates, it becomes vital to understand how humans engage with such text, and whether or not they are able to detect when the text they are reading did not originate with a human writer. Prior work on human detection of generated text focuses on the case where an entire passage is either human-written or machine-generated. In this paper, we study a more realistic setting where text begins as human-written and transitions to being generated by state-of-the-art neural language models. We show that, while annotators often struggle at this task, there is substantial variance in annotator skill and that given proper incentives, annotators can improve at this task over time. Furthermore, we conduct a detailed comparison study and analyze how a variety of variables (model size, decoding strategy, fine-tuning, prompt genre, etc.) affect human detection performance. Finally, we collect error annotations from our participants and use them to show that certain textual genres influence models to make different types of errors and that certain sentence-level features correlate highly with annotator selection. We release the RoFT dataset: a collection of over 21,000 human annotations paired with error classifications to encourage future work in human detection and evaluation of generated text. 
    more » « less