learning approach for an expert system 
*

*Step 1: I developed an expert system that classifies data into two classes, each rule in this expert system has a certainty factor (expert opinions).

*Step 2: The input of the system is a group of sentences and the output is classified sentences into two classes like c1, c2.

*Step 3: The user has to mark the sentences placed in the wrong.

*Step 4: the system shall edit rules certainty factors depending on the user's feedback, to make the system more accurate.

I want a machine learning algorithm that uses my rules to classify the data.
not: l have a labeled dataset
 A: For future readers, OP and I had a small back and forth in the comments under his question that clarified what he was working on.
Direct answer to the posed question:
I think doing exactly what you asked (having an ML tool adjust your existing rules) is effectively impossible (extremely complex task).
My recommendation:
Instead of having ML adjust your existing "expert" rules, you should just have an ML algorithm become your "expert". You mentioned you have a labelled dataset, so you should just treat this as a supervised learning problem, where it is up to the ML algorithm to become the expert and "decide" what the best rules are for classifying your text. Again, since your dataset is already labelled (since you already know the correct classification for each sentence), your ML algorithm will receive constant feedback and try to adjust to get the best result.
Now the harder question is: "What data do I give the ML model so that it can make these predictions?"
Well, this is where your existing expert rules come in. Don't throw them away! You can transform them into features for the learner, and the learner will decide how to weight/utilize that information.
There's two ways you can do this. Let's say, as an easy example, that one of your expert rules is that "If sentence is longer than 7 words, it's generally class 1 and if it's shorter, it's generally class 2". From this rule, you can extract two features:
One feature is to feed "number of words in sentence" to the algorithm, and it will try to find the best cut off point (maybe it's 9 and not 7). You feed it the same data your rules use, and let it make the determination.
The second thing you can do is along with that raw data "number of words in sentence", you can feed in some kind of binary 1/0 determination based on your existing rules, so that the algorithm can potentially weight that in as well.
So for the same example, you have two features. Based on that rule you already have, you can feed in the # words in sentence, plus your own expert determination of what the class "should" be based on that information.
I hope this makes sense.
For the ML model, I would heavily recommend RandomForest. It easily handles different kinds of inputs (categorical, numerical etc.), it has few hyperparameters you need to tune, is quick to train, generally has great performance, and it gives an easily interpretable probability output associated with each class prediction. It's just overall a very "plug and play", beginner-friendly algorithm compared to others, and it generally delivers comparable results to other algorithms.
Best of luck!
