As the autumn rain lashed against the windows of the Turing Lab this week, the [CS-ADV] Computer Science cohort was engaged in a heated debate—not about code syntax, but about the definition of a joke.
For the annual “Code & Culture” hackathon, Year 12 students were tasked with building a Natural Language Processing (NLP) model capable of performing real-time sentiment analysis on the unique dialect of Virtanen International College: a chaotic blend of English, Finnish, and international slang often referred to as “Finglish.”
The objective, supervised by Ms. Sarah Jenkins, was to train a neural network to categorise student conversations in the common room as “Positive,” “Neutral,” or “Conflict.”
The ‘Silence’ Bug
The project, titled Babel-25, ran smoothly during the initial data ingestion phase. Students fed the model thousands of lines of anonymised transcripts from the school’s debate club and social discord servers. However, when the model was tested on live audio in the cafeteria, it began throwing critical error flags.
“We ran into a cultural wall,” explained Kaito Tanaka, the student lead for the Data Set team. “The model was trained on standard American English datasets. It flagged long pauses in conversation as ‘Social Discomfort’ or ‘Data Loss.’ It didn’t understand that in Finland, a ten-second silence between friends is comfortable, not awkward. The AI was diagnosing the entire Finnish student body with social anxiety.”
Sarcasm vs. Algorithm
The second, more complex failure occurred with sentiment tracking. The model consistently misidentified British dry humour and Finnish deadpan delivery as “Hostility.”
During a test run, a Year 13 student famously remarked to a friend about the torrential October weather: “Lovely day for a swim, isn’t it?”
The Babel-25 algorithm immediately flagged the statement as “Psychotic/Delusional” (Confidence: 94%) because the semantic meaning (swimming) violently clashed with the meteorological data (freezing rain). The AI lacked the contextual layer to detect the sarcasm.
The Manual Override
Realising that their “black box” solution was failing, Ms. Jenkins paused the coding phase and ordered a “Manual Annotation Sprint.” For three days, the students had to sit with printed transcripts and manually tag thousands of lines of dialogue to teach the machine the nuances of human tone.
“It was tedious, unglamorous work,” Ms. Jenkins noted. “The students wanted to write cool Python scripts. Instead, they had to sit there and explain to a computer why ‘Break a leg’ is a wish for good luck, not a threat of violence. It taught them the most important lesson in AI: an algorithm is only as smart as the human bias fed into it.”
A Nuanced Success
By Friday afternoon, the recalibrated model was running with a 78% accuracy rate—imperfect, but functional. It successfully identified a “Finglish” exchange where students switched between three languages in a single sentence, correctly tagging the sentiment as “Collaborative.”
The project has garnered interest from the University of Turku’s Department of Digital Humanities, which has requested access to the students’ annotated “Finglish” dataset for further research into code-switching behaviours.
Conclusion
The Babel-25 terminal has been left running in the lab corner, quietly listening and learning. While it still occasionally struggles to understand why students complain about “too much homework” while smiling, it serves as a reminder that while code is binary, human communication remains beautifully, stubbornly analogue.
Leave a Reply