ELLMath

Math assessments pose unique challenges for ELLs, especially in comprehending and responding to word problems due to language barriers, not mathematical ability. Our product will generate math assessments tailored to the language learning levels of individual students. It will feature capabilities to translate assessments into the student's native language and cultural settings when necessary, ensuring comprehension and fairness.

In the United States, a significant portion of educators experience burnout, with 44% of K-12 teachers and 35% of college or university faculty reporting that they very often or always feel overwhelmed by their workload. Our solution is specifically designed to alleviate this stress by enhancing the efficiency of assessment creation and delivery for educators working with diverse English language learners. For more details on teacher burnout, see Teacher Burnout: Challenges in K-12 and Higher Education.

Our product specifically targets English Language Learners (ELLs), a group that often encounters significant educational challenges. ELL students frequently face obstacles such as language barriers, cultural differences, and limited access to resources tailored to their unique needs, which can hinder their academic and social integration. This lack of opportunity can exacerbate disparities in educational outcomes, leaving ELL students at a disadvantage compared to their native-speaking peers.

In response, our solution aims to make learning more fair and engaging for these learners by personalizing and adapting assessments to meet their language and learning needs.

Additionally, ELL teachers are a scarce resource, often burdened with the complex task of designing assessments that accommodate diverse linguistic abilities. Our product alleviates this challenge by simplifying the creation of customized math assessments, enabling teachers to focus more on teaching and less on administrative tasks. This support is crucial for enhancing both the effectiveness of instruction and the overall educational experience for ELLs.

Simplification of English: Our product can automatically simplify the language used in both instructions and questions based on the student proficiency in English. This involves using shorter sentences, simpler vocabulary, and clear, concise language structures that make content more accessible to learners with varying levels of English proficiency.

Translation: For students who are new to English or who benefit from multilingual support, Our product can provide seamless translation of texts. This feature helps ensure that language barriers do not prevent students from demonstrating their true mathematical understanding or from engaging fully with the material.

Personalization: Beyond linguistic adjustments, personalization involves adapting the type of content, the mode of delivery (e.g., visual, auditory, textual), and the pacing of assignments and assessments to match each student’s unique learning style, cultural context and proficiency level. Our product can leverage data from ongoing interactions to continually refine and adjust the educational experience to optimize learning outcomes.

Culturally Relevant Content: Our product can customize assessments by incorporating culturally relevant examples and contexts that resonate with the diverse backgrounds of ELLs. This personalization can make assessments more engaging and relatable, which can improve comprehension and reduce anxiety associated with testing for ELLs from different countries.

In summary, Our solution elevates K-8 education by enabling fairer and more engaging assessments for students, while significantly simplifying assessment creation for teachers.

Our team, Balaji, Nikhil and Sowmya, brings a wealth of experience in education and AI from institutions like Harvard, Stanford, and IIT Madras. We've been teachers, and we've worked with teachers, giving us a unique insight into the challenges and needs of today’s diverse classrooms.

We have the necessary expertise in AI, assessments, technology etc.. to build the prototype. Having been a teacher and working with teachers in varied capacities, we understand the practical challenges deploying such a product in diverse settings. We will be doing school visits in SF and NYC to improve our understanding of the cultural context here.

As an immigrant in the US, we understand the practical challenges for ELLs to get integrated into the classroom and show progress.

We are in the process of building a prototype and hence we chose the concept stage.

Our solution introduces significant innovation to the math assessment of English Language Learners (ELLs) by integrating AI technologies that specifically address the unique challenges faced by this demographic. 

Adaptation for ELLs - While a plethora of tools exist for math adaptation for different levels, our solution focuses on ELLs and their diverse needs for adaptation on various axes - language learning, cultural context and personalization based on individual interests.

LLM Engineering with Domain Expertise - Large language models have a vast storehouse of general information that have to be mined using appropriate prompt engineering, in-context learning, retrieval augmentation or fine-tuning efforts. With our interdisciplinary team, we are poised to incorporate the domain knowledge from the field and experiment with a suite of techniques. Some examples of our domain knowledge include ensuring mathematical consistencies, being sensitive to the needs of the ELL learners and being cognizant of differences in linguistic and cultural situations and identifying the requirements of the instructors.

Instructor Focus: Our product has an instructor focus where the assessment data is expertly summarized and presented to the instructor that can help in lesson planning.

Privacy and Security - Recognizing the sensitivity of educational data, particularly for minors, our solution is designed with a strong commitment to privacy and security by incorporating the requisite anonymization modules before feeding data into LLMs. 

Natural Language Processing (NLP): NLP is critical for simplifying English in assessments, providing accurate translations, and enabling personalized feedback. We perform this task by experimenting with multiple large language models (LLMs) and various ways of eliciting information from them - fine-tuning, prompting and in-context learning to tailor the product for our needs. We also use LLMs for summarization of class assessments to assist the instructor. 

Retrieval-Augmented Generation (RAG): We use RAG to enhance the capabilities of our language models, enabling them to generate more accurate, contextually appropriate content for assessments. RAG involves retrieving relevant information from a large dataset (knowledge base) before generating language. Since we want to maintain mathematical correctness and respectfulness for cultural context, we use RAG to generate templates that would then be expanded by LLMs. This process is particularly useful in generating personalized and varied question sets and explanations that cater specifically to the language and educational needs of ELLs.

Dependency on Third-Party Models

Our system integrates both proprietary models developed in-house and third-party models where appropriate. For example, we use established third-party NLP models for basic language processing tasks and focus our development efforts on customizing these models with RAG and specific ML algorithms tailored to educational content. This hybrid approach allows us to leverage the strengths of widely validated models while ensuring that our specific application needs are optimally addressed.

Our prior research experiments have given the conviction that this kind of a solution would work in this context. Our resident AI expert has worked on NLP systems solving math problems for over a decade. She has documented the pitfalls in mathematical reasoning and understands the kind of prompting guard rails to be placed for sensible and mathematically sound adaptation in language, cultural contexts and other personalization. 

As we are developing the models for ELLs, who are a special interest group, utmost care will be taken to be respectful of cultural contexts and learning abilities. We combat bias by not allowing the LLMs to generate text at will, rather develop from a set of pre-apprved templates to ensure equity. Generating around the template - will make sure that additional info or bias should not come.

3 partime employees