Courtney Napoles

Ground Truth for Grammatical Error Correction Metrics

How do we know which grammatical error correction (GEC) system is best? A number of metrics have been proposed over the years, each motivated by weaknesses of previous metrics; however, the metrics themselves have not been compared to an empirical gold standard grounded in human judgments. We conducted the first human evaluation of GEC system outputs, and show that the rankings produced by metrics such as MaxMatch and I-measure do not correlate well with this ground truth. As a step towards better metrics, we also propose GLEU, a simple variant of BLEU, modified to account for both the source and the reference, and show that it hews much more closely to human judgments.

Finding Good Conversations Online: The Yahoo News Annotated Comments Corpus.

This work presents a dataset and annotation scheme for the new task of identifying “good” conversations that occur online, which we call ERICs: Engaging, Respectful, and/or Informative Conversations. We develop a taxonomy to reflect features of entire threads and individual comments which we believe contribute to identifying ERICs; code a novel dataset of Yahoo News comment threads (2.4k threads and 10k comments) and 1k threads from the Internet Argument Corpus; and analyze the features characteristic of ERICs. This is one of the largest annotated corpora of online human dialogues, with the most detailed set of annotations. It will be valuable for identifying ERICs and other aspects of argumentation, dialogue, and discourse.

There's No Comparison: Reference-less Evaluation Metrics in Grammatical Error Correction.

Current methods for automatically evaluating grammatical error correction (GEC) systems rely on gold-standard references. However, these methods suffer from penalizing grammatical edits that are correct but not in the gold standard. We show that reference-less grammaticality metrics correlate very strongly with human judgments and are competitive with the leading reference-based evaluation metrics. By interpolating both methods, we achieve state-of-the-art correlation with human judgments. Finally, we show that GEC metrics are much more reliable when they are calculated at the sentence level instead of the corpus level. We have set up a CodaLab site for benchmarking GEC output using a common dataset and different evaluation metrics.

A Report on the Automatic Evaluation of Scientific Writing Shared Task.

The Automated Evaluation of Scientific Writing, or AESW, is the task of identifying sentences in need of correction to ensure their appropriateness in a scientific prose. The data set comes from a professional editing company, VTeX, with two aligned versions of the same text – before and after editing – and covers a variety of textual infelicities that proofreaders have edited. While previous shared tasks focused solely on grammatical errors (Dale and Kilgarriff, 2011; Dale et al., 2012; Ng et al., 2013; Ng et al., 2014), this time edits cover other types of linguistic misfits as well, including those that almost certainly could be interpreted as style issues and similar “matters of opinion”. The latter arise because of different language editing traditions, experience, and the absence of uniform agreement on what “good” scientific language should look like. Initiating this task, we expected the participating teams to help identify the characteristics of “good” scientific language, and help create a consensus of which language improvements are acceptable (or necessary). Six participating teams took on the challenge.

Sentential Paraphrasing as Black-Box Machine Translation

We present a simple, prepackaged solution to generating paraphrases of English sentences. We use the Paraphrase Database (PPDB) for monolingual sentence rewriting and provide machine translation language packs: prepackaged, tuned models that can be downloaded and used to generate paraphrases on a standard Unix environment. The language packs can be treated as a black box or customized to specific tasks. In this demonstration, we will explain how to use the included interactive webbased tool to generate sentential paraphrases.

The Effect of Multiple Grammatical Errors on Processing Non-Native Writing

In this work, we estimate the deterioration of NLP processing given an estimate of the amount and nature of grammatical errors in a text. From a corpus of essays written by English-language learners, we extract ungrammatical sentences, controlling the number and types of errors in each sentence. We focus on six categories of errors that are commonly made by English-language learners, and consider sentences containing one or more of these errors. To evaluate the effect of grammatical errors, we measure the deterioration of ungrammatical dependency parses using the labeled F-score, an adaptation of the labeled attachment score. We find notable differences between the influence of individual error types on the dependency parse, as well as interactions between multiple errors.

Automatically Scoring Freshman Writing: A Preliminary Investigation

In this work, we explore applications of automatic essay scoring (AES) to a corpus of essays written by college freshmen and discuss the challenges we faced. While most AES systems evaluate highly constrained writing, we developed a system that handles open-ended, long-form writing. We present a novel corpus for this task, containing more than 3,000 essays and drafts written for a freshman writing course. We describe statistical analysis of the corpus and identify problems with automatically scoring this type of data. Finally, we demonstrate how to overcome grader bias by using a multi-task setup, and predict scores as well as human graders on a different dataset. Finally, we discuss how AES can help teachers assign more uniform grades.