0

I wrote a rule-based keyword detection and classification program specialized in my language (Vietnamese) and would like to know where this app is useful. Here how the program work:

  1. First you input the prompt (which is a bunch of keywords), e.g. fish 50k
  2. Then it will automatically label/classify the prompt like this:
Object: fish
Type of Object: food
Place of transaction: market
Type of place of transaction: offline
Consumer: myself
Type of consumer: myself
Price: 50000 VND

The program can make this classification based on a config you declare, e.g.:

- Dimension name: Object
  Classification:
    - Food: fish, meat
    - Appliance: computer, speaker
  Default value: meat
...

Which problems do you see this app will be useful? In general, where have you seen rule-based classification being applied? Especially in the context of ChatGPT and its GPT store? What domains, fields or industries have the need to use rule-based approaches? I think there should be a review on how this technique is applied in various field, but I can't find one.

In my understanding, there are two types of approaches in NLP: rule-based and statistic-based. Rule-based approach is simple, understandable and need not training, while statistic-based is better if the rules are complex and you have good training data. I think rule-based classification is much cheaper and more accurate than statistical-based classification. Is that correct?

Ooker
  • 123
  • 7

1 Answers1

2

Pretty much all of them. I can't see a field that couldn't benefit from some rule-based approach. The thing here is that we often talk about ML Models when, typically, they are ML Systems. What's the difference? A system combines multiple models, rules, sanity checks, and logical paths from input to output.

In a perfect world, we would only use rules. Because, after all, why not? They are explicit, explainable, don't change, and can be theoretically analyzed to confirm their completeness and accuracy. But here is why not:

  • Good rules are hard to define
  • Defining a complete ruleset is extremely challenging, though parts of the set could be defined.
  • Rulesets are typically large and require expertise to build and maintain
  • Rules don't adapt. So, if you deal with systems that are partially unknown or subject to change, rules may falter after a while.

Final thought: if you think about it, all statistics-based or ML classifiers can be viewed as parameterized rules.

Valentin Calomme
  • 6,256
  • 3
  • 23
  • 54