Background: Employing natural language processing and latent semantic analysis, the current work was completed as a constituent part of a larger research project for designing and launching artificial intelligence in the form of deep artificial neural networks. The models were evaluated on a proprietary corpus retrieved from a data warehouse, where it was extracted from MyReviewers, a sophisticated web application purposed for peer review in written communication, which was actively used in several higher education institutions. The corpus of laboratory reports in STEM annotated by instructors and students was used to train the models. Under the Common Rule, research ethics were ensured by protecting the privacy of subjects and maintaining the confidentiality of data, which mandated corpus de-identification.Literature Review: De-identification and pseudonymization of textual data remains an actively studied research question for several decades. Its importance is stipulated by numerous laws and regulations in the United States and internationally with HIPAA Privacy Rule and FERPA.Research Question: Text de-identification requires a significant amount of manual post-processing for eliminating faculty and student names.  This work investigated automated and semi-automated methods for de-identifying student and faculty entities while preserving author names in cited sources and reference lists. It was hypothesized that a natural language processing toolkit and an artificial neural network model with named entity recognition capabilities would facilitate text processing and reduce the amount of manual labor required for post-processing after matching essays to a list of users’ names. The suggested techniques were applied with supplied pre-trained models without additional tagging and training. The goal of the study was to evaluate three approaches and find the most efficient one among those using a users’ list, a named entity recognition toolkit, and an artificial neural network.Research Methodology: The current work studied de-identification of STEM laboratory reports and evaluated the performance of the three techniques: brute forth search with a user lists, named entity recognition with the OpenNLP machine learning toolkit, and NeuroNER, an artificial neural network for named entity recognition built on the TensorFlow platform. The complexity of the given task was determined by the dilemma, where names belonging to students, instructors, or teaching assistants must be removed, while the rest of the names (e.g., authors of referenced papers) must be preserved.Results: The evaluation of the three selected methods demonstrated that automating de-identification of STEM lab reports is not possible in the setting, when named entity recognition methods are employed with pre-trained models. The highest results were achieved by the users’ list technique with 0.79 precision, 0.75 recall, and 0.77 F1 measure, which significantly outweighed OpenNLP with 0.06 precision, 0.14 recall, and 0.09 F1, and NeuroNER with 0.14 precision, 0.56 recall, and 0.23 F1.Discussion: Low performance of OpenNLP and NeuroNER toolkits was explained by the complexity of the task and unattainability of customized models due to imposed time constraints. An approach for masking possible de-identification errors is suggested.Conclusion: Unlike multiple cases described in the related work, de-identification of laboratory reports in STEM remained a non-trivial labor-intensive task. Applied out of the box, a machine learning toolkit and an artificial neural network technique did not enhance performance of the brute forth approach based on user list matching.Directions for Future Research: Customized tagging and training on the STEM corpus were presumed to advance outcomes of machine learning and predominantly artificial intelligence methods. Application of other natural language toolkits may lead to deducing a more effective solution.

Objective: We report on a comparative study of the language used by middle school students in their answers to a constructed response test of science inquiry knowledge. Background: Text analyses using statistical models have been conducted across a number of disciplines to identify topics in a journal, to extract topics in Twitter messages, and to investigate political preferences. In education, relatively few studies have analyzed the text of students’ written answers to investigate topics underlying the answers. Methodology: Two types of linguistic analysis were compared to investigate their utility in understanding students’ learning of scientific investigation practices. A statistical method, latent Dirichlet allocation (LDA), was used to extract topics from the texts of student responses. In the LDA model, topics are viewed as multinomial distributions over the vocabulary of documents. These topics were examined for content and used to characterize student responses on the constructed response items. The change from pre-test to post-test in proportions of use of each of the topics was related to students’ learning. Next, a qualitative method, systemic functional linguistic (SFL) analysis, was used to analyze the text of student responses on the same test of science inquiry knowledge. Student assessments were analyzed for two linguistic features that are important for convincing scientific communication: technical vocabulary usage and high lexical density. In this way, we investigated whether human judgement regarding the changes observed from texts based on the SFL framework agreed with the inference regarding the changes observed from the texts through LDA. Research questions: Two research questions were investigated in this study: (1) What do the LDA and SFL analyses tell us about students’ answers? (2) What are the similarities and differences of the two analyses? Data: The data for this study were taken from an NSF-funded host study on teaching science inquiry skills to middle school students who were a mix of both native English speakers and English-language learners. The primary objective was to enable participants to learn to take ownership of scientific language through the use of language-rich science investigation practices. The LDA analysis used a sample of 252 students’ pre-and post-assessments. The SFL analysis used a second sample of 90 students’ pre- and post-assessments. Results: In the LDA analysis, three topics were detected in student responses: “preponderance of everyday language (Topic 1),” “preponderance of general academic language (Topic 2),” and “preponderance of discipline-specific language (Topic 3).” Students’ use of topics changed from pre-test to post-test. Students on the post-test tended to have higher proportions of Topic 3 than students on the pre-test. In the SFL analysis, students tended to use more technical vocabulary and have higher lexical density in their written responses on the post-test than on the pre-test. Discussion: Results from the LDA and SFL analyses suggest that students responded using more discipline-specific language on the post-test than on the pre-test. In addition, the results of the two linguistic features from the SFL analysis, technical vocabulary usage and lexical density, were compared with the results from the LDA analysis. • Conclusion: Results of the LDA and SFL analyses were consistent with each other and clearly showed that students improved in their ability to use the discipline-specific and academic terminology of the language of scientific communication.