A NLP pipeline that identifies Hong Kong Instagram users solely based on username patterns.
This binary classification project identifies Hong Kong Instagram users without relying on real-time scraping or API calls on the users' metadata. Instead, my model uses the underlying syllable patterns in usernames for the classification task.
An example use case for my identifier can be creating a lightweight IG ad bot that exclusively engages with Hong Kong IG users for social media marketing purposes.
Scikit-learn's Logistic Regression, Random Forest and SVM are chosen as the baseline models for evaluation.
The training data is collected from hypeauditor.com using a Selenium-based web scraper.
try HK-Insta-Identifier yourself on my streamlit app!
The core principle of this classification task revolves around Romanized Cantonese linguistic features and the behavior of NLTK syllable tokenizer.
Notably, the NLTK syllable tokenizer is not Romanized Cantonese-specific, but it could still provide a workaround by capturing some distinctive patterns.
The following are the visualizations of the distribution of repeated and unique syllables, and the top 10 most appeared syllables in Non-HK and HK usernames:
Terminologies:
- Vowel: a,e,i,o,u,y* and they can be a standalone syllable
- Consonant: a character that is not a vowel and cannot be a standalone syllable
- Consonant-vowel (CV) syllable: a syllable that contains both vowels and consonants, e.g. 'fi', 'ha', etc.
- Monosyllabic: a word with single/ one syllable
note: y sometimes can act as a vowel as well
-
Higher Appearance of Standalone Vowels in HK -> Unique Vowel Clusters
- Romanized Cantonese has a lot of unique adjacent vowels compared to English or other languages
- For example: {"張": "ch-eu-ng", "楊": " yeu-ng", "趙": "ch-iu", "游": " yau", ...}
- The tokenizer is not familiar with these clusters and might treat them as an individual syllable
-
Less CV syllables, More Unique Syllables in HK -> Complex Consonants Clusters
- Romanized Cantonese also has a lot of complex consonants combinations and some can even contain no vowels at all
- For example : {"翠": "ts-ui", "芷": " tsz", "吳": "ng", "郭": "kw-ok", ...}
- This confuses the tokenizer to group the consonants to other vowels, resulting in more unique syllables
-
Lower Overall Syllable Counts in HK -> Monosyllabic Chinese Characters
- Hong Kong People's name are mostly made up of 3 Chinese characters, and chinese characters are monosyllabic
- i.e. Hong Kong people's name at most have 3 syllables and leads to lower overall syllable counts in usernames
- So the maximum syllable count is around 75 ("ha") in Non-HK while it is only roughly 50 ("i") in HK
All these differences contributed as the patterns for the models to identify HK usernames from non-HK usernames.
In NLP, conventional tokens might be words, phrases, or subword units. On the contrary, syllabic tokenization is regarded as a rather "inconsistent" tokenization technique since the phonetics in the English language is also inconsistent in some extent, such as the "k" in "knife" or "olo" in "colonel".
However, I found that syllabic tokenization could still be the most suitable existing solution in tokenizing usernames with following reasons:
-
Usernames don't contain whitespaces
- Lack of whitespaces in the IG usernames makes the traditional tokenizers that heavily rely on whitespaces can't work properly
-
Usernames are not sentences
- In other words, usernames are too short to extract a "word" as a unit for the features
-
Usernames are not proper English
- Any conventional tokenizers won't have the word embeddings for usernames, so a subword tokenizer that tokenizes a word based on the prefixes and suffixes would also not work
-
No existing Romanized Cantonese-specific tokenizer
- As demonstrated in the last part, syllabic tokenizer can somehow still be able to extract some unique patterns, albeit the lack of Romanized Cantonese word embeddings
learn more in Forbidden Spellings & NLP pipeline deep dive: Why doesn't anyone tokenize by syllables?
After running a GridSearchCV, it was found that both Logistic Regression (LR) and Support Vector Machines (SVM) yielded the best testing results with 0.742. On the other hand, Random Forest (RF) with 0.691 showed the worst performance due to potential underfitting.
please consider checking out hk_ig_clf.ipynb or hk_ig_clf.pdf for full details.
Username analysis is a super super complicated topic because of the freedom and creativity users have when choosing their usernames, for example:
- Private account holders may choose not to include their government (Cantonese) names in their usernames
- English names are widely adopted by many Hong Kong users and greatly reduce the visibility of linguistic patterns tied to Cantonese
- There is often overlap among Romanized Chinese dialects, making it difficult to distinguish between users from different regions
So while my approach might seem effective enough(? ), you have to carefully consider these factors when interpreting the results of this project.
If you find this project interesting, feel free to star🌟or fork this repo and explore new ideas to build on this work - collaboration is always welcome!
Possible future directions could include developing a Romanized Cantonese-specific tokenizer and incorporating users' bios information.
Thank you again for taking the time to learn about this project!
back to top