Using word embedding to extend words for searching POI names

Question

I am developing my own mobile app related to digital map. One of the functions is searching POIs (points of interest) in the map according to relevance between user query and POI name.

Besides the POIs whose names contain exact words in the query, the app also needs to return those whose names are semantically related. For example, searching 'flower' should return POI names that contain 'flower' as well as those that contain 'florist'. Likewise, searching 'animal' should return 'animal' as well as 'veterinary'.

That said, I need to extend words in the query semantically. For example, 'flower' has to be extended to ['flower', 'florist']. I have tried to use word embeddings: using the words corresponding to most similar vectors as extensions. Due to the fact I don't have user review data right now and most of the POI names are very short, I used trained word2vec model published by Google. But the results turn out to be not what I expect: most similar words of 'flower' given by word2vec are words like 'roses'and 'orchid', and 'florist' is not even in the top 100 most similar list. Likewise, 'animal' gives 'dog', 'pets', 'cats' etc. Not very useful for my use case.

I think simply using word embedding similarity may not be enough. I may need to build some advanced model based on word embedding. Do you have any suggestions?

Answer 1

I think word embeddings are overkill in this particular case.

My suggestion would be to go for a simple dictionary based approach: compose sets of semantically related words, and then use those to expand your query terms. This might take a bit longer to set up, but has several advantages:

1. simplicity: you can't make many mistakes with this

2. transparency: you know exactly why a certain term matches, and another one doesn't

3. accuracy: you have tight control over the whole process; if some term is wrong, you remove it from the set. You cannot do that with embeddings

4. resources: a dictionary-based approach is far simpler and needs less storage

'Old tech' doesn't sound as sexy as the latest deep learning stuff, but unless you want to have this as a toy project to learn about how to do things with embeddings I would say the latter are the wrong tool for the job. At least you can be sure that it works, and if it doesn't you can easily fix it.