Handing in a paper or exercise and merely receiving "bad" or "incorrect" as feedback is not very helpful when the goal is to improve.Unfortunately, this is currently the kind of feedback given by many Automatic Short Answer Grading (ASAG) systems.One of the reasons for this is a lack of content-focused elaborated feedback datasets.To encourage research on explainable and understandable feedback systems, we present the Short Answer Feedback dataset (SAF).Similar to other ASAG datasets, SAF contains learner responses and reference answers to German and English questions.However, instead of only assigning a label or score to the learners' answers, SAF also contains elaborated feedback explaining the given score.Thus, SAF enables supervised training of models that grade answers and explain where and why mistakes were made.This paper discusses the need for enhanced feedback models in real-world pedagogical scenarios, describes the dataset annotation process, gives a comprehensive analysis of SAF, and provides T5-based baselines for future comparison. 1

We present Bloom Library, a linguistically diverse set of multimodal and multilingual datasets for language modeling, image captioning, visual storytelling, and speech synthesis/recognition.These datasets represent either the most, or among the most, multilingual datasets for each of the included downstream tasks.In total, the initial release of the Bloom Library datasets covers 363 languages across 32 language families.We train downstream task models for various languages represented in the data, showing the viability of the data for future work in low-resource, multimodal NLP and establishing the first known baselines for these downstream tasks in certain languages (e.g., Bisu [bzi], with an estimated population of 700 users).Some of these first-of-their-kind baselines are comparable to state-of-the-art performance for higher-resourced languages.The Bloom Library datasets are released under Creative Commons licenses on the Hugging Face datasets hub to catalyze more linguistically diverse research in the included downstream tasks.

Practice is essential for language learning. This is true for writing as well as speaking. However, in contrast to writing, it can be challenging to offer students sufficient time to practice speaking while receiving corrective feedback from a teacher. Considering the importance of corrective feedback for language mastery, automatic feedback systems could provide vital assistance through additional supervised speaking exercise. For this reason, this paper proposes an end-to-end feedback generation pipeline to correct grammar errors in unconstrained speech. The approach consists of four steps, converting raw speech files into transcripts with automatic speech recognition, removing disfluency, correcting grammatical errors, and preparing the feedback for presentation to the user. An explorative analysis of the pipeline with English language learners indicates that out-of-the-box automatic speech recognition models degrade in performance when used by language learners. However, training the model with only 15 minutes of learners’ speech decreases the word error rate almost by half.

Learners comprehend complex texts only if they understand the concepts involved, and assessing their comprehension works best if the assessment targets those concepts. Hence, assessment systems, such as question generators, often profit from learning-relevant concepts as input to solve their tasks. However, automatically detecting a given text’s relevant concepts is non-trivial. If a concept is deemed relevant heavily depends on the context and thus on automatic text understanding. Recent advancements in supervised machine learning improved state-of-the-art automatic text understanding. However, those methods need training corpora to learn the association between the relevant concepts and their text’s context. This work shows that textbooks comprise all the necessary information to construct such corpora. We introduce and evaluate an open-source automatic back-of-the-book index extractor and a corpus construction algorithm to derive training corpora from a set of high-quality PDF textbooks. We furthermore investigate to what extent state-of-the-art machine learning models can learn to extract concepts from constructed corpora. The results show that the models have decent performance and provide evidence generalization to unseen concepts.