Are We Modeling the Task or the Annotator? An Investigation of Annotator Bias in Natural Language Understanding Datasets

Crowdsourcing has been the prevalent paradigm for creating natural language understanding datasets in recent years. A common crowdsourcing practice is to recruit a small number of high-quality workers, and have them massively generate examples. Having only a few workers generate the majority of examples raises concerns about data diversity, especially when workers freely generate sentences. In this paper, we perform a series of experiments showing these concerns are evident in three recent NLP datasets. We show that model performance improves when training with annotator identifiers as features, and that models are able to recognize the most productive annotators. Moreover, we show that often models do not generalize well to examples from annotators that did not contribute to the training set. Our findings suggest that annotator bias should be monitored during dataset creation, and that test set annotators should be disjoint from training set annotators.


Introduction
Generating large datasets has become one of the main drivers of progress in natural language understanding (NLU). The prevalent method for creating new datasets is through crowdsourcing, where examples are generated by workers (Zaidan and Callison-Burch, 2011;Richardson et al., 2013;Bowman et al., 2015;Rajpurkar et al., 2016;Trischler et al., 2017). A common recent practice is to choose a small group of workers who produce high-quality annotations, and massively generate examples using these workers.
Having only a few workers annotate the majority of dataset examples raises concerns about data diversity and the ability of models to generalize, especially when the crowdsourcing task is to generate free text. If an annotator consistently uses language patterns that correlate with the labels, a neural model can pick up on those, which can lead to an over-estimation of model performance.
First, we establish that annotator information improves model performance by supplying annotator IDs as part of the input features. Second, we show that models are able to recognize annotators that generated many examples, illustrating that annotator information is captured by the model. Last, we test whether models generalize to annotators that were not seen at training time. We observe that often generalization to new annotators fails, and that augmenting the training set with a small number of examples from these annotators substantially increases performance.
Taken together, our experiments show that annotator bias exists in current NLU datasets, which can lead to problems in model generalization to new users. Hence, we propose that annotator bias should be monitored at data collection time and to tackle it by having the test set include examples from a disjoint set of annotators.
In a typical process, annotators are recruited and screened (Sabou et al., 2014), often resulting in a small group that creates most of the dataset examples. Mihaylov et al. (2018) recruited a few dozens of qualified workers that wrote 5,957 questions. Suhr et al. (2018) recruited 99 workers for a sentence writing task, who created more than 100,000 examples. Williams et al. (2018) recruited 387 workers for writing over 400,000 sentences, while Krishna et al. (2017) had 33,000 workers contributing 1.7 million examples.
These examples demonstrate that datasets are often constructed using a small number of annotators, approximately 1 annotator per 10 2 -10 3 examples. Furthermore, (see Section 3), the annotator distribution is skewed with a few annotators creating the vast majority of the dataset. In tasks that involve creative language writing, this may have implications on data diversity, and lead to an over-estimation of model performance.

Experimental Setup
We focus on crowdsourcing tasks where workers produce full-length sentences. We first describe the datasets we test our hypothesis on, and then provide details on the model and training.
Datasets We consider recent NLU datasets, for which the annotator IDs are available.
• MNLI (matched) (Williams et al., 2018): A NLI dataset. Each example was created by introducing an annotator with a premise sentence and asking her to write a hypothesis sentence that is either entailed, contradicted or is neutral with respect to the premise. • OPENBOOKQA (Mihaylov et al., 2018): A multiple-choice question answering dataset, focusing on multi-hop reasoning. Each question and its answer distractors were written by a worker, based on a given scientific fact. • COMMONSENSEQA (Talmor et al., 2019): A multiple-choice question answering dataset, focused on commonsense knowledge. Questions were written by crowdworkers, who try to bridge between two concepts extracted from CONCEPTNET (Speer et al., 2017) Table 1 summarizes the size and number of annotators who worked on each dataset. Figure  Model We use the pretrained BERT-base (Devlin et al., 2018), a strong model obtaining close to state-of-the-art performance on all three datasets. We add a single linear layer over BERT outputs and apply the same fine-tuning procedure in all experiments: fine-tuning for 3 epochs, using batch size 10, learning rate 2 × 10 −5 , and maximum sequence length of 128 word pieces.

Experiments and Results
We now describe a series of experiments for quantifying annotator bias, aiming to answer the following questions: 1) Do models perform better when exposed to the annotator ID? 2) Can models detect the annotators from their generated examples? 3) Do models generalize across annotators?
The utility of annotator information Our first experiment aims to determine whether models perform better given perfect information on the annotator ID. To this end, we follow the standard way of feeding input to BERT (Devlin et al., 2018) and concatenate the annotator ID as an additional feature to every example in every dataset. Formally, we replace every example (x = (w 1 , ..., w |x| ), y) created by annotator z, with the example ((z, w 1 , ..., w |x| ), y), where z is a textual unique annotator identifier, x is the input sequence and y is the gold label.
We compare performance on the original datasets and their new version. Adding the annotator ID improves model performance across all datasets (  Annotator recognition Perfect annotator information improves performance, but it is possible that a model can recognize the annotators from the input sequence only, even without being exposed to the annotator ID explicitly. In the next experiment, we investigate the ability of models to recognize the annotators from the input. To this end, we fine-tune BERT-base to predict annotator IDs from examples. We limit the task to 6 labels of the top-5 most productive annotators of each dataset and an OTHER label for all other annotators. Formally, we replace every example (x, y) created by annotator z, with the example (x,z), wherez = z if z is in the top-5 annotators andz = OTHER otherwise. Figure 2 shows the F1-score for the top-5 annotators of every dataset (y-axis), and the fraction of dataset examples created by each annotator (x-axis). Overall, annotators who write many examples are recognized better by the model: The model struggles to recognize MNLI annotators with F1 scores below 0.5, and excels at recognizing annotators from COMMONSENSEQA with scores between 0.76-0.91. For the top annotator of OPENBOOKQA, who created 24% of the dataset examples, the model obtains a high F1 score of 0.8. To conclude the first two experiments, annotator ID information is useful for the downstream task, and also can be predicted with high probability from the input for a large fraction  We fine-tune BERT-base and evaluate it on the development set, and compare the results to a random data split of identical size. We repeat every experiment 3 times, except for multi-annotator experiments on OPENBOOKQA and COMMON-SENSEQA which we repeat 9 times due to high variance. Table 3 shows the mean and standard deviation of performance difference (p.d.) between each annotator(s) split and its corresponding ran-dom split, where negative numbers indicate that performance on the annotator split was lower.
Our clearest finding is that in OPENBOOKQA performance on the the multi-annotator split is dramatically lower than on a random split in all 5 annotator sets, where performance drops by up to 23 accuracy points. This shows that the model does not generalize to examples generated by unseen annotators. In the other datasets, results on the multi-annotator split are more varied, where performance drops in roughly half the cases, sometimes substantially -up to 10 accuracy points in COMMONSENSEQA and 5 in MNLI.
In the single-annotator splits, in roughly half the cases performance on the annotator split was lower than the random split. However, measuring p.d. only for single annotators might be misleading, because specific annotators vary in the difficulty of examples they produce. Thus, running a model on a new annotator that produces easy examples will not result in decreased performance. Next, we propose a more fine-grained experiment that controls for these two confounding factors.
For every single-annotator data split S train , S dev , we perform a series of augmentation experiments, where we move a random fraction of k examples from S dev to S train , for k = 0.1, 0.2, 0.3. We keep the size of S train constant by randomly removing examples from it before augmentation. We repeat experiments multiple times as before, and report the p.d mean and standard deviation in Figure 3.
Results for both COMMONSENSEQA and OPENBOOKQA show a rapid increase of 10-20 accuracy point for all top-5 annotators, given only a small number of their generated examples. This shows that the examples generated by these annotators are not inherently difficult, and that the model can substantially improve performance by being exposed to the language that the annotators generate. Conversely, performance changes are marginal for MNLI, suggesting generalization patterns are mostly due to example difficulty. The different results for MNLI compared to those observed for OPENBOOKQA and COMMONSENSEQA may be attributed to the lessskewed annotator distribution and large number of examples in MNLI (see Figure 1 and Table 1), which make the model more robust to small perturbations in the data distribution.

Discussion and Conclusions
This study set out to investigate whether prevalent crowdsourcing practices for building NLU datasets introduce an annotator bias in the data that leads to an over-estimation of model performance.
We established that perfect annotator information can improve model performance, and that the language generated by annotators often reveals their identity. Moreover, we tested the ability of models to generalize to unseen annotators in three recent NLU datasets, and found that in two of these datasets annotator bias is evident. These findings may be explained by the annotator distributions and the size of these datasets. Skewed annotator distributions with only a few annotators creating the vast majority of examples are more prone to biases.
Our results suggest that annotator bias should be monitored in crowdsourcing tasks involving free text generation by annotators. This can be done by testing model performance on new annotators during data collection. Moreover, to tackle annotator bias, we propose that training set annotators should be separated from test-set annotators.