Classifying short strings of text with additional context

Question

I have a list of short strings each identifying a city. Misspellings are very common. The example below shows some of these short strings, along with the correct city they're supposed to match.

string	city
amsterdam	amsterdam
asmterddam	amsterdam
amstterdm	amsterdam
new york	new york
new yrok	new york
nwe york	new york
neew york	new york
nw york	new york

I would like to train a classifier that takes the input string and then predict the most likely city to be identified. However there is a subtlety which is that the correct city will depend not only on the input string, but also on other factors such as the current location of the person, and of which person is writing.

For instance, the following strings could mean different cities based on these other features:

current_location	person	string	city
uk	John	dratford	dartford
uk	Jack	dartfford	dartford
uk	Jill	datrfrod	dartford
norway	Jill	dartfoord	dartford
norway	John	datrfrod	dartfjord
norway	James	datrfrod	dartfjord
sweden	Olaf	dratfjod	dartfjord
uk	Olaf	dratfoord	dartfjord

As we can see, the same input string can mean a different actual city depending on who writes it, and where the writer is currently located. This means that just a fuzzy search using tf-idf wouldn't be enough to make correct predictions.

I'm thinking of two different approaches to implement an appropriate classifier:

• boosted trees using character n-grams as features + the other categorical features
• neural network using convolutions on letters + the other categorical features

One advantage here is that the string will always be limited in size (max about 50/60 characters).

One thing I'm struggling with in particular, for the case of boosted trees, is that, how to make the model learn about the order of character n-grams? For instance, the sequence mst coming after the sequence ams is indicative of amsterdam, but I don't know how to give the model a sense of sequence, where this sequence could be anywhere in the string? (e.g. if the input string is aamsterdam then this sequence would be in position 2, instead of position 1).

So in short, I'd be looking for guidance in the best way to decompose and model the input string, in order to transform it into features usable by either boosted trees or a (convolutional?) neural network, along the other categorical input features.

I'm obviously not looking for a complete solution here, just general guidance would be hugely appreciated.

Answer 1

There are typically many discrete stages in a natural language processing (NLP) pipeline. Two possible stages are normalization and classification. Strings are normalized before being classified. Normalization (aka, standardization) makes string data more consistent. Mapping spelling variates to the same representation is an example. Mapping spelling variations to a consistent representation does not require machine learning. It can be with something like edit distance look-up in a dictionary.

After normalization, classification is more straightforward. Classification models depend on the quality and quantity of features, aka feature engineering. Predicting a city most likely won't happen at the character-level. Modeling at the token-level would probably be more useful.