Learning Chinese (well enough to order from the menu) using MTurk

TL;DR: I used MTurk and PostgreSQL to create vocab flash cards for Chinese restaurant menus, ranked from most common to least common terms.

It’s an embarrassing scenario: you’re dining with friends at a Chinese restaurant when the server drops off the menus, including the non-English “secret menu”. All eyes turn to you, as the only Chinese person in the group. “You’re Chinese right? Can you read this?”

I’ve created a tool to solve this problem: a deck of 1,214 flash cards featuring vocabulary found on Chinese menus. The cards are ordered from most common (鸡: chicken) to least common, so if you study them in order, you’ll quickly learn the characters and phrases you’re most likely to encounter.

To do this, I compiled a set of photos of Chinese restaurant menus from places like restaurant websites and Yelp. What I needed was to extract all of the Chinese text but the images weren’t good enough for OCR. Different fonts, complex layouts, and low-quality phone photos of menus (with glare, perspective skew, etc.) made it impossible for OCR to process the text reliably. And even if OCR could get it 90% right, the transcription errors would pollute the data, potentially resulting in some nonsensical datasets.

Thus, MTurk. Using human transcribers, I was able to get all images transcribed with minimal effort on my part. I had to:

• Upload the images somewhere
• Create a set of MTurk tasks, one for each image
• Write a clear description of what I wanted workers to do
• Pay money!
• Download and spot-check results

Ultimately I paid $26.40 to get the set of about 150 images transcribed. There was some overhead in approving or rejecting the work – despite making the instructions as clear as I could, some workers transcribed only the English text, or entered gibberish.

The next step was to dump everything into a PostgreSQL database to make manipulating the data a bit easier. Using psql’s \copy command also made it easier to ad-hoc reformat the data I needed into CSVs.

From there, I used tools I previously wrote to count characters and multi-character words, creating a list of vocab along with the number of occurrences. The code also grabs definitions from the free CC-CEDICT (available under CC-BY-SA), resulting in a set of “notes” that can be imported by Anki.

Admittedly the data set could be expanded. With the source data set that I collected, a lot of words only show up once. Words like 巴西 (Brazil), 生姜 (raw ginger), 七喜 (7-Up) are all mixed up in the bottom of the vocab barrel. Such is language learning…

Learning the top 131 cards is enough to recognize 50% of the words in the source menus, and 451 achieves 75% recognition. However, 925 cards are needed for 90% recognition. These numbers are actually worse than reality, as simplified and traditional variants are separate cards, and you will also have learned the constituent characters of many multi-character words. For example, 魚香 is card #39, but 魚 is #19 and 香 is #26.

If you’re interested, have a go at it! The deck of flash cards is available for Anki – all you need to do is import the deck and set aside a few minutes each day to study cards.

Download

To download the vocab list or Anki flash cards, see the Github repository: https://github.com/erjiang/chinesemenus (look under Releases for the APKG Anki deck).