Multiple Sentence Inputs about lime

[marcotcr]

Oh, this is interesting - you have both text AND continuous features.
If all you care about is explaining the text parts while keeping the continuous values fixed, it should be easy, as the only difference is that you have two sentences instead of one. What you would need to do in this case is to put both sentences in a string, separated by a character that is not caught by the split_expression parameter (see the initializer of LimeTextExplainer). You then could define a classifier_fn that takes whatever text LIME gives is, splits it into two sentences, computes the sentence vectors and etc. Your code would look something like this:

def get_classifier_fn(continuous_features):  
    def classifier_fn(text_list):
        ret = []
        for text in text_list: 
            sentence1, sentence2 = text.split(SPLIT_CHARACTER)
            # assuming this returns a number between 0 and 1
            ret.append(model.predict_similarity(get_embedding(sentence1), 
                                                get_embedding(sentence2),
                                                continuous_features))
        return ret
    return classifier_fn

Then, if you wanted to explain a particular instance, you would have to do something like this:

def explain_instance(instance):  
   text = '%s %s %s' % (instance.sentence1, SPLIT_CHARACTER, instance.sentence2)
   fn = get_classifier_fn(instance.continuous_features)
   explainer = LimeTextExplainer(split_expression=SPLIT_EXPRESSION, bow=BOW)
   return explainer.explain_instance(text, fn, labels=(0,))

You may have to be a bit creative if you want to use the visualizations we have, since the text displayed will have both sentences. Also, you have to think if it makes sense to use bow=True or False in this case. If you're doing anything with sequences, I would say False makes more sense.

If you want to explain the impact of the words and the numerical features, you'll have to write a function that perturbs both at the same time. If this is the case, you would want to do a mix of __data_labels_distances in lime_text.py and __data_inverse in lime_tabular.py.

Both are viable ideas, the second one is definitely a bit more work.
I would be very interested to see what you end up with, I've never thought about this particular use case. If you are comfortable sharing what the application is, please tell me over email : ).
Let me know if you have any more questions too.

Best,

[jaderabbit]

Thank you so so much! Your explanations are super helpful. For now, I'm particularly interested in the text components so will focus on that. I will definitely drop you a mail :)

[LeZhengThu]

Hi, in code "explainer.explain_instance(text, fn, labels=(0,))", the default value is "labels=(1,)" and explanation in your original code is "labels: iterable with labels to be explained." I wonder what this variable exactly stands for? Thanks.

[marcotcr]

This parameter is there for multi-class classification problems, and it defines for which classes (or labels) we want to generate explanations for. The default is 1 because usually what we have is two-class classification, where we only care about explaining label 1 (since the explanation for label 0 is the opposite).
The code assumes that the classifier function returns an array with a column per label (the probability for each), while in this example we are only outputting one number. That is why I set labels=(0,).

[LeZhengThu]

Thanks so much for your quick response. BTW, is there a chance to only explain several particular features? Which parameter should I set?

[marcotcr]

Do you mean restrict the explanation to a particular set of features? If so, there is no parameter for that.

[Aureole-1210]

Thank you so so much! Your explanations are super helpful. For now, I'm particularly interested in the text components so will focus on that. I will definitely drop you a mail :)

Thank you so so much! Your explanations are super helpful. For now, I'm particularly interested in the text components so will focus on that. I will definitely drop you a mail :)

Hi! I am conducting the same task as yours. But I find the final score of every token is similar. I think it's incorrect. Could you please tell me how you complete it?
More appreciate!!!!!!!!

[JINHXu]

LimeTextExplainer

can someone give an example SPLIT_EXPRESSION? 🙏

[JINHXu]

@marcotcr can you give an example for SPLIT_EXPRESSION? 🙏 I have a very similar use case meanwhile would not know what to pass to SPLIT_EXPRESSION. many thanks! 🙏 documentations here are not helping:

split_expression – Regex string or callable. If regex string, will be used with re.split. If callable, the function should return a list of tokens.

[JINHXu]

apparently without using SPLIT_EXPRESSION, the split_char is considered part of the input and get interpreted