We recently open sourced a library that makes it easy to quickly get started with pre-trained word vector models like word2vec from Google, Facebook, and Stanford. It adds a ton of extra functionality like fast similiarity indexing with Annoy and fairly robust out-of-vocabulary word lookups (handling misspellings) out of the box. If you want to quickly get started with pre-built models, it might be worth checking out along with Gensim:
Are document vectors supported analogously to word vectors? It's a bit peculiar to me that gensim withholds useful word2vec methods from the Doc2Vec class, like most_similar_from. I'd really enjoy having a common interface to both vector spaces.
Currently, it doesn't support document vectors, but it should be possible to add it and maintain the same interface. We're open to accepting a pull request if your interested in building that feature.
Hey, I like the work that you guys are doing and have been working on similar things since last 2 years. Are you guys hiring? My personal email id is banga.kamal@gmail.com.
> One important thing to note is that Word2Vec does not consider the positional variable of the context words.
There is a very interesting new approach of learning Gaussian mixtures from words, which addresses this problem: https://github.com/benathi/word2gm
It also solves another interesting problem, which is that Word2Vec doesn't always put "categories" close to their actual contents.
Eg. I wrote a simple AI for playing Codenames with word vectors: https://github.com/thomasahle/codenames
But where a human use the clue "Countries 2" for "Germany" and "England", the Glove vectors (at least in my implementation) much prefers the clue "France 2", which is pretty confusing at first.
The Gaussian mixture representation is quite interesting, so thanks for the link. But it seems like they do not make use of the relative position of words in the text either. They simply maximize the margin of the expected likelihood for words that occur in the same context over those that do not.
However, in most cases it's probably actually better to ignore word order, since that captures more semantic relatedness rather than the syntactic features that influence word order. It also lets you translate between languages that do not necessarily have similar grammar: "Word Translation Without Parallel Data" https://arxiv.org/abs/1710.04087
Whether it's okay to ignore word order really depends on your goals. In my work, tracking word order to build a sentence representation is necessary to get composition rules correct. You can get away with ignoring order surprisingly often but when it's needed, it's vital. When aggregated across an entire corpus, order ends up mattering a lot, in an absolute sense (something looking only at accuracy numbers will hide).
Compare the following:
1) You must state the benefit.
2) You must benefit the state.
3) Feed fish vs. Fish feed.
Sometimes limits in inference:
1) A loves B.
2) Cats eat fish.
3) Cats are mammals.
Sometimes the sequence preceding or following will be important:
Right, what I meant was not that order doesn't matter at all, but that order mostly doesn't matter for the meaning of individual words (except for some homonyms which can be disambiguated using syntactical information).
When you want to handle whole sentences, obviously order becomes much more relevant, but for that you'd feed the word vectors into an LSTM or similar to handle the order-dependence.
This is much more nuanced criticism of language vectors than Chomsky's impatient hatred of them. There are interviews worth watching in which he loathes them for all the reasons you mention. Thankfully, I suppose, they rarely disappoint me. :)
Word2Vec actually employs a default strategy of pruning the most common 0.0001% of words in the corpus. My understanding is that this permits the RNN layer to train on the unadulterated text rather than an unnatural concentrate of tokens, which is supposedly beneficial.
Interestingly there's some controversy whether this benefit extends to punctuation. Most W2V tutorials instruct the programmer to apply tokenizing functions that leave only space-separated word characters, annihilating everything else, but I've actually observed higher accuracy scores from models that tokenize non-word characters individually, i.e. as strings of length 1. I'm not sure exactly why this is the case, but there is a common sense interpretation which might explain. Output from the precursory tokenizer output is often transduced by word2phrase, an optional bigram recognizer. Mikolov strongly recommends using it, but I'm sure it struggles to infer pointwise mutual information from a text that's missing ~70% of its contents.
Valid point. I should have made it clearer that NFL stands for National Football League. I was trying to demonstrate the difference in domain, one being sports and the other being computer science.
17 comments
[ 12.0 ms ] story [ 812 ms ] threadhttps://github.com/plasticityai/magnitude/
There is a very interesting new approach of learning Gaussian mixtures from words, which addresses this problem: https://github.com/benathi/word2gm
It also solves another interesting problem, which is that Word2Vec doesn't always put "categories" close to their actual contents.
Eg. I wrote a simple AI for playing Codenames with word vectors: https://github.com/thomasahle/codenames But where a human use the clue "Countries 2" for "Germany" and "England", the Glove vectors (at least in my implementation) much prefers the clue "France 2", which is pretty confusing at first.
However, in most cases it's probably actually better to ignore word order, since that captures more semantic relatedness rather than the syntactic features that influence word order. It also lets you translate between languages that do not necessarily have similar grammar: "Word Translation Without Parallel Data" https://arxiv.org/abs/1710.04087
Compare the following:
1) You must state the benefit.
2) You must benefit the state.
3) Feed fish vs. Fish feed.
Sometimes limits in inference:
1) A loves B.
2) Cats eat fish.
3) Cats are mammals.
Sometimes the sequence preceding or following will be important:
1) Problems with learning
2) Learning with problems
3) It has a learning problem
4) A learning problem for it.
Or consider an example from:
https://www.researchgate.net/publication/2335962_How_Well_Ca...
1) It was not the sales manager who hit the bottle that day, but the office worker with the serious drinking problem.
2) That day the office manager, who was drinking, hit the problem sales worker with the bottle, but it was not serious.
Order ends up mattering when you want to capture nuance or power higher order inference.
When you want to handle whole sentences, obviously order becomes much more relevant, but for that you'd feed the word vectors into an LSTM or similar to handle the order-dependence.
Interestingly there's some controversy whether this benefit extends to punctuation. Most W2V tutorials instruct the programmer to apply tokenizing functions that leave only space-separated word characters, annihilating everything else, but I've actually observed higher accuracy scores from models that tokenize non-word characters individually, i.e. as strings of length 1. I'm not sure exactly why this is the case, but there is a common sense interpretation which might explain. Output from the precursory tokenizer output is often transduced by word2phrase, an optional bigram recognizer. Mikolov strongly recommends using it, but I'm sure it struggles to infer pointwise mutual information from a text that's missing ~70% of its contents.
I will say that he’s wrong that NFL never co-occurs with ML. I’ve had discussions involving the No Free Lunch theorem by its initials.