Reference: Xiong, W., Litman, D. J., & Schunn, C. D. (2012). Natural language processing techniques for researching and improving peer feedback. Journal of Writing Research, 4(2), 155-176.
- Feedback on writing is seen to improve students’ writing, but the process is resource intensive.
- Possible options to reduce the workload in giving feedback:
- Direct feedback using technology assisted approaches (from grammar checks to complex computational linguistics).
- Peer Review [Considered in this paper].
- Peer review:
- Good feedback from a group of peers is found to be as useful as the instructor’s feedback and even weaker writers are seen to provide useful feedback to stronger writers (See references in original paper).
- When providing feedback on other students’ work, students become mindful of the mistakes and improve their own writing.
- Some web-based peer review systems: PeerMark in turnitin.com, SWoRD (used in this study) and Calibrated Peer Review.
- Challenge lies in the form of feedback provided by peers – peer feedback might not be in a form useful to make revisions. Key features identified to aid revisions:
- Localized information (Providing exact location details like paragraph, page numbers or quotations).
- Concrete solution (Suggesting possible solution rather than just pointing the problem).
- Research problem: Studying peer review is hard with a large amount of feedback data.
- Practical problem: Identifying useful feedback for students and possible interventions to help them provide good feedback.
- To automatically process peer feedback and identify the presence or absence of the two key features (Providing feedback on feedback for students and automatically coding feedback for researchers).
- Refer prototype shown in figure 1 of the original paper that suggests students to provide localized and explicit solutions.
Technical – How? (Details explained in study 1 and study 2)
- Building a domain lexicon from common unigrams and bigrams in student papers.
- Counting basic features like domain words, modals, negations, overlap between comment and paper etc. from each feedback.
- Creating a logic model to identify the type of feedback (Contains localization information or not/ contains explicit solution or not) – classification task in machine learning.
Method and Results:
Study 1 – Localization Detection:
- Each feedback comment represented as a vector of the four attributes below:
- regularExpressionTag: Regular expressions to match phrases that use location in a comment (E.g. “on page 5”).
- #domainWord: Counting the number of domain-related words in a comment (based on the domain lexicon gathered from frequent terms in student papers).
- sub-domain-obj, deDeterminer: Extracting syntactic attributes (sub-domain-obj) and count of words like “this, that, these, those” which are demonstrative determiners.
- windowSize, #overlaps: Extracting the length of matching words from the document to identify quotes (windowSize) and words overlapped.
- Weka models to automatically code localization information. The decision tree model had better accuracy (77%, recall 82%, precision 73%) in predicting if a feedback was localized or not. To refer the rules that made up the decision tree, take a look at Figure 2 of the original paper.
Study 2 – Solution Detection:
- Feedback comment represented as vectors using the three types of attributes(Refer table 2 in the original paper for details).
- Simple features like word count and the order of comment in overall feedback.
- Essay attributes to capture the relationship between the comment and the essay and domain topics.
- Keyword attributes semi-automatically learned based on semantic and syntactic functions.
- Logistic Regression model to detect the presence/ absence of explicit solutions (accuracy 83%, recall 91%, precision 83%). Domain-topic words followed by suggestions were highly associated with prediction. Detailed coefficients of attributes predicting presence of solution can be referred in Table 3 of the original paper.
Study 3: Can Research Rely on Automatic Coding?
- Comparing automatically coded data to hand coded data to see if the accuracy is sufficiently high for practical implementation.
- Helpfulness ratings by peers and 2 experts (content, writing experts) on peer comments at a review level.
- To account for expert ratings:
- Regression analysis using feedback type proportions (praise only comments, summary only comments, problem/solution containing comments), proportion localized critical comments, and proportion solution providing comments as predictors.
- 10 fold cross validation – SVM best fit.
- To check if same models are built using machine coded and hand coded data – 10 stepwise regressions. Refer Table 4 in the original paper to see the feedback features commonly included in the model by the different raters – Different features were helpful for different raters.
- Overall regression model is similar to hand coded localization data (Most of the positivity, solution and localization were similar between hand coding and automatic coding).
- Predictive models for detecting localization and solution information are statistical tools and do not provide deep content insights.
- To be integrated into SWoRD to provide real time feedback on comments.
- Technical note: Comments were already pre-processed – segmented into idea units by hand; data split by hand into comment type (summary, praise, criticism).
- Future work:
- Examine impact of feedback on feedback comments
- Obtaining generalization across courses
- Improving accuracy of prediction