The effect of noisy labels on the performance of NLP systems has been studied extensively for system training. In this paper, we focus on the effect that noisy labels have on system evaluation. Using automated scoring as an example, we demonstrate that the quality of human ratings used for system evaluation have a substantial impact on traditional performance metrics, making it impossible to compare system evaluations on labels with different quality. We propose that a new metric, PRMSE, developed within the educational measurement community, can help address this issue, and provide practical guidelines on using PRMSE.

For the last 5 years, we have developed and maintained RSMTool – an open-source tool for evaluating NLP systems that automatically score written and spoken responses. RSMTool is designed to be cross-disciplinary, borrowing heavily from NLP, machine learning, and educational measurement. Its cross-disciplinary nature has required us to learn a user-centered development approach in terms of both design and implementation. We share some of these lessons in this paper.

The issues of algorithmic fairness and bias have recently featured prominently in many publications highlighting the fact that training the algorithms for maximum performance may often result in predictions that are biased against various groups. Educational applications based on NLP and speech processing technologies often combine multiple complex machine learning algorithms and are thus vulnerable to the same sources of bias as other machine learning systems. Yet such systems can have high impact on people’s lives especially when deployed as part of high-stakes tests. In this paper we discuss different definitions of fairness and possible ways to apply them to educational applications. We then use simulated and real data to consider how test-takers’ native language backgrounds can affect their automated scores on an English language proficiency assessment. We illustrate that total fairness may not be achievable and that different definitions of fairness may require different solutions.

Literacy is crucial for functioning in modern society. It underpins everything from educational attainment and employment opportunities to health outcomes. We describe My Turn To Read, an app that uses interleaved reading to help developing and struggling readers improve reading skills while reading for meaning and pleasure. We hypothesize that the longer-term impact of the app will be to help users become better, more confident readers with an increased stamina for extended reading. We describe the technology and present preliminary evidence in support of this hypothesis.

Automated scoring engines are usually trained and evaluated against human scores and compared to the benchmark of human-human agreement. In this paper we compare the performance of an automated speech scoring engine using two corpora: a corpus of almost 700,000 randomly sampled spoken responses with scores assigned by one or two raters during operational scoring, and a corpus of 16,500 exemplar responses with scores reviewed by multiple expert raters. We show that the choice of corpus used for model evaluation has a major effect on estimates of system performance with r varying between 0.64 and 0.80. Surprisingly, this is not the case for the choice of corpus for model training: when the training corpus is sufficiently large, the systems trained on different corpora showed almost identical performance when evaluated on the same corpus. We show that this effect is consistent across several learning algorithms. We conclude that evaluating the model on a corpus of exemplar responses if one is available provides additional evidence about system validity; at the same time, investing effort into creating a corpus of exemplar responses for model training is unlikely to lead to a substantial gain in model performance.

In this study, we develop content features for an automated scoring system of non-native English speakers’ spontaneous speech. The features calculate the lexical similarity between the question text and the ASR word hypothesis of the spoken response, based on traditional word vector models or word embeddings. The proposed features do not require any sample training responses for each question, and this is a strong advantage since collecting question-specific data is an expensive task, and sometimes even impossible due to concerns about question exposure. We explore the impact of these new features on the automated scoring of two different question types: (a) providing opinions on familiar topics and (b) answering a question about a stimulus material. The proposed features showed statistically significant correlations with the oral proficiency scores, and the combination of new features with the speech-driven features achieved a small but significant further improvement for the latter question type. Further analyses suggested that the new features were effective in assigning more accurate scores for responses with serious content issues.

Using a case study, we show that variation in oral reading rate across passages for professional narrators is consistent across readers and much of it can be explained using features of the texts being read. While text complexity is a poor predictor of the reading rate, a substantial share of variability can be explained by timing and story-based factors with performance reaching r=0.75 for unseen passages and narrator.

In large-scale educational assessments, the use of automated scoring has recently become quite common. While the majority of student responses can be processed and scored without difficulty, there are a small number of responses that have atypical characteristics that make it difficult for an automated scoring system to assign a correct score. We describe a pipeline that detects and processes these kinds of responses at run-time. We present the most frequent kinds of what are called non-scorable responses along with effective filtering models based on various NLP and speech processing technologies. We give an overview of two operational automated scoring systems —one for essay scoring and one for speech scoring— and describe the filtering models they use. Finally, we present an evaluation and analysis of filtering models used for spoken responses in an assessment of language proficiency.

Automated scoring of written and spoken responses is an NLP application that can significantly impact lives especially when deployed as part of high-stakes tests such as the GRE® and the TOEFL®. Ethical considerations require that automated scoring algorithms treat all test-takers fairly. The educational measurement community has done significant research on fairness in assessments and automated scoring systems must incorporate their recommendations. The best way to do that is by making available automated, non-proprietary tools to NLP researchers that directly incorporate these recommendations and generate the analyses needed to help identify and resolve biases in their scoring systems. In this paper, we attempt to provide such a solution.

We consider the automatic scoring of a task for which both the content of the response as well its spoken fluency are important. We combine features from a text-only content scoring system originally designed for written responses with several categories of acoustic features. Although adding any single category of acoustic features to the text-only system on its own does not significantly improve performance, adding all acoustic features together does yield a small but significant improvement. These results are consistent for responses to open-ended questions and to questions focused on some given source material.

This paper is a preliminary report on using text complexity measurement in the service of a new educational application. We describe a reading intervention where a child takes turns reading a book aloud with a virtual reading partner. Our ultimate goal is to provide meaningful feedback to the parent or the teacher by continuously tracking the child’s improvement in reading fluency. We show that this would not be a simple endeavor, due to an intricate relationship between text complexity from the point of view of comprehension and reading rate.

We explore various supervised learning strategies for automated scoring of content knowledge for a large corpus of 130 different content-based questions spanning four subject areas (Science, Math, English Language Arts, and Social Studies) and containing over 230,000 responses scored by human raters. Based on our analyses, we provide specific recommendations for content scoring. These are based on patterns observed across multiple questions and assessments and are, therefore, likely to generalize to other scenarios and prove useful to the community as automated content scoring becomes more popular in schools and classrooms.

In this paper we explore to what extent the difficulty of listening items in an English language proficiency test can be predicted by the textual properties of the prompt. We show that a system based on multiple text complexity features can predict item difficulty for several different item types and for some items achieves higher accuracy than human estimates of item difficulty.