4 ms·

Very cool! I've been working for a while on extracting "semantic" from naked text (mostly news). One of the big limits of word2vec is that the semantic should

by ecesena 11y ago

Very cool! I've been working for a while on extracting "semantic" from naked text (mostly news). One of the big limits of word2vec is that the semantic should be related with words proximity in sentences, which works in some cases, but not always. I'll give an example for all: travel/geography. If you query something like Italy [1], the results are other European states. But if you're looking for news in Italy, or to plan a vacation to Italy, or for some Italian food... or anything related to Italy itself, you probably don't expect "Spain" to be the first result. It would be nice to have some sort of easy way in word2vec to define domains and their relationship with words proximity in sentences, to overcome situations like this one. [1] https://sense2vec.spacy.io/?Italy%7CGPE https://sense2vec.spacy.io/?Italy%7CGPE

syllogism 11y ago
Well, you can't do this in the web API obviously, but there are a few ways you could do this. This is a good idea really -- thanks for bringing this up. One way would be to predefine the entity types or tags that you want to get in your results. So you could ask for things like Italy that are nouns. The other way is to use the vector space. The classic demonstration of this is the arithmetic, doing like "Italy|GPE - *|GPE + food". My results for this have been very mixed. I wouldn't expect the query above to work. I would think you'd have more luck specifying the query as a combination of constraints: first query for foods in some way, and then sort them by distance from Italy.
nl 11y ago
Don't think of Word2Vec (etc) as a search - it's a different thing. It's more like a recommender: If I like Pizza in Napoli where should I go in Spain and what should I eat there? Italy:Pizza -> Spain:? Italy:Napoli -> Spain:?

4 ms·

Very cool! I've been working for a while on extracting "semantic" from naked text (mostly news). One of the big limits of word2vec is that the semantic should

by ecesena 11y ago

syllogism 11y ago
Well, you can't do this in the web API obviously, but there are a few ways you could do this. This is a good idea really -- thanks for bringing this up. One way would be to predefine the entity types or tags that you want to get in your results. So you could ask for things like Italy that are nouns. The other way is to use the vector space. The classic demonstration of this is the arithmetic, doing like "Italy|GPE - *|GPE + food". My results for this have been very mixed. I wouldn't expect the query above to work. I would think you'd have more luck specifying the query as a combination of constraints: first query for foods in some way, and then sort them by distance from Italy.
nl 11y ago
Don't think of Word2Vec (etc) as a search - it's a different thing. It's more like a recommender: If I like Pizza in Napoli where should I go in Spain and what should I eat there? Italy:Pizza -> Spain:? Italy:Napoli -> Spain:?