1. Background

1.1 Modeling Language Performance

While exploring a governmental dataset, I noticed that personalized language learning is not universally accessible for everyone. In multi-grade classes with higher peer pressure and less individualized context, the positive correlation between SES and verbal IQ is even stronger than in single-grade class (p<.01). Curious and inspired, I trained a Linear Discriminant classifier to quantify the impact of this issue, predicting the probability of an intelligent student to obtain an above-median language scores with different SES.

1.2 Digital Sentiment Analysis

This discrepancy, partially resulting from the lack of individualized language learning environment, is stunning. I connected this statistical evidence with my research focus on foreign language effect at Boston University’s psycholinguistic lab. This theory proposed people tend to struggle in comprehending and communicating emotion-related content in second language. Lack of emotional context, to some extent, blocked their acquisition of second language.

What about a digital environment? Are people comfortable with expressing feelings online? Aiming to explore the viability of leveraging digital persona as a mean to emotionally self-express and thus to develop the proficiency in a second language, I gathered a rich text corpus of online chat & social media posts and ran a series of sentiment analyses.

Most online discourses are modestly and moderately emotional. Little of them are purely neutral. Absolute sentiment is strongly positively skewed, indicating the presence of extreme emotional catharsis online. Adopting the interactional format of digital social media can thus offer a viable way to integrate emotional self-expression meaningfully into contextual language learning.

2. Explore

After assessing technical viability, I step into narrow down actual user pain points using qualitative methods.

2.1 Diary Study

I conducted a contextual inquiry with frequent users of Duolingo, observing how they complete practice sessions over an entire week and asked their overall impression towards the production and self-expression aspects of the course.

2.2 Competitive User Testing

Right after diary study, 8 users were interviewed about the desirable features of digital self-expression from their perspectives. Referring to the social media products they liked, I centered the discussion around features they frequently used. Their quotes were categorized into meaningful codes.

3. Framework

Competitive benchmarking revealed a set of shared patterns for effective language learning, informing feature selection and AR spatial design.

3.1 Differentiation & Priorities

Based on key dimensions identified through thematic analysis, I constructed perceptual map of various benchmarks in e-learning industry, positioning my product at “high context” and “mid proficiency” combination. Using Kano Model, I evaluated their priorities and selected a subset of them into a coherent user journey and a primitive AR architecture that supports immersive production-focused practice.

3.2 System Architecture

4. Solution

4.1 Design System

Cosmospeaker’s pleasantly warm design system is established upon core concepts of “embrace, support, immerse”. All design elements are systematically and consistently organized into reusable variables and components variants that can directly implemented by software developers.

4.2 Interactive Prototype

Conversation-based language learning complemented with lightweight, situated AR.

Users customize learning companions to build expressive skills through task-oriented conversation. Sessions flexibly shift between conversational UI and optional AR, grounding practice in context. Learning is structured around CEFR-aligned vocabulary while adapting to individual preferences.