In this tutorial, it teaches how to develop a simple encoder-decoder model with attention using pytorch.
However, in the encoder or decoder, self.embedding = nn.Embedding(input_size, hidden_size) (or similar) is defined. In pytorch documents, nn.Embedding is defined as "A simple lookup table that stores embeddings of a fixed dictionary and size."
So I am confused that, in the initialization, where does this lookup table has come from? Does it initialize some random embeddings for the indices and then they will be trained? Is it really necessary to be in the encoder/decoder part?
Thanks in advance.
Answering the last bit first: Yes, we do need Embedding or an equivalent. At least when dealing with discrete inputs (e.g. letters or words of a language), because these tokens come encoded as integers (e.g. 'a' -> 1, 'b' -> 2, etc.), but those numbers do not carry meaning: The letter 'b' is not "like 'a', but more", which its original encoding would suggest. So we provide the Embedding so that the network can learn how to represent these letters by something useful, e.g. making vowels similar to one another in some way.
During the initialization, the embedding vector are sampled randomly, in the same fashion as other weights in the model, and also get optimized with the rest of the model. It is also possible to initialize them from some pretrained embeddings (e.g. from word2vec, Glove, FastText), but caution must then be exercised not to destroy them by backprop through randomly initialized model.
Embeddings are not stricly necessary, but it would be very wasteful to force network to learn that 13314 ('items') is very similar to 89137 ('values'), but completely different to 13315 ('japan'). And it would probably not even remotely converge anyway.
Related
I've used gensim Word2Vec to learn the embedding of monetary amounts and other numeric data in bank transaction memos. The goal is to use this to be able to extract these amounts and currencies from future input strings.
Design
Our input strings are something like
"AMAZON.COM TXNw98e7r3347 USD 49.00 # 1.283"
During preprocessing, I tokenize and also replace all tokens that have the possibility of being a monetary amount (string consisting only of digits, commas, and <= 1 decimal point/period) with a special VALUE_TOKEN. And I also manually replace exchange rates with RATE_TOKEN. The result would be
["AMAZON", ".COM", "TXNw", "98", "e", "7", "r", "3347", "USD", "VALUE_TOKEN", "#", "RATE_TOKEN"]
With all my preprocessed lists of strings in list data, I generate model
model = Word2Vec(data, window=3, min_count=3)
The embeddings of model that I'm most interested in are that of VALUE_TOKEN, RATE_TOKEN, as well as any currencies (USD, EUR, CAD, etc.). Now that I generated the model, I'm not sure what to do with it.
Problem
Say I have a new string that the model has never seen before,
new_string = "EUR 299.99 RATE 1.3289 WITH FEE 5.00"
I would like to use model to identify which tokens of new_string is most contextually similar to VALUE_TOKEN (which should return ["299.99", "5.00"]), which is closest to RATE_TOKEN ("1.3289"). It should be able to classify these based on the learned embedding. I can preprocess new_string the way I do with the training data, but because I don't know the exchange rate before hand, all three tokens of ["299.99", "5.00", "1.3289"] will be tagged the same (either with VALUE_TOKEN or a new UNIDENTIFIED_TOKEN).
I've looked into methods like most_similar and similarity but don't think they work for tokens that are not necessarily in the vocabulary. What methods should I use to do this? Is this the right approach?
Word2vec's fuzzy, dense embedded token representations don't strike me as the right tool for what you're doing, though they might perhaps be an indirect contributor to a hybrid approach.
In particular:
The word2vec algorithm originated from, & has the most consistent public results, when applied to natural-language texts, with their particular patterns of relative token frequences, and varied co-occurrences. Certainly, many ahave applied it, with success, to other kinds of text/record data, but such uses may require a lot more preprocessing/parameter-tuning, and to the extent the underlying data has some fixed, highly-repetitive scheme, might be more amenable to other approaches.
If you replace all known values with 'VALUE_TOKEN', & all known rates with 'RATE_TOKEN', then the model is only going to learn token-vectors for 'VALUE_TOKEN' & 'RATE_TOKEN'. Such a model won't be able to supply any vector for non-replaced tokens it's never seen like '$1.2345' or '299.99'. Even collapsing all those to 'UNIDENTIFIED_TOKEN' just limits the model to whatever it learned earlier was the vector for 'UNIDENTIFIED_TOKEN' (if any, in the training data).
I've not noticed existing word2vec implementations offering an interface for inferring the word-vector for new unknown-vectors, from just one or several new examples of its appearance in-context. They could, in the same style of new-document-vector inference used by 'Paragraph Vectors'/Doc2Vec, but just don't.) The closest I've seen is Gensim's predict_output_word(), which does a CBOW-like forward-propagation on negative-sampling models, to every 'output node' (one per known word), to give a ranked list of the known-words most-likely to appear given some context words.
That predict_output_word() might, if fed surrounding known-tokens, contribute to your needs by whether it says your 'VALUE_TOKEN' or 'RATE_TOKEN' is a more-likely model-prediction. You could adapt its code to only evaluate those two candidates, if you're always sure the right answer is one or the other, for a speed-up. A simple comparison of the average-of-context-word-vectors, and the candidate-answer vectors, might be as effective as the full forward-propagation.
Alternatively, you might want use the word2vec model solely as a source of features (via context-words) for some other classifier, which is trained to answer VALUE or TOKEN. This other classifier's input might include things like:
some average of the vectors of all nearby tokens
the full vectors of closest neighbors
a one-hot encoding ('bag-of-words') of all nearby (or 'preceding') or 'following) known-tokens, assuming the vocabulary of non-numerical tokens is fairly short & highly indicative
?
If the data streams might include arbitrary new or corrupted tokens whose meaning might be inferrable from substrings, you could consider a FastText model as well.
Here I have a word2vec model, suppose I use the google-news-300 model
import gensim.downloader as api
word2vec_model300 = api.load('word2vec-google-news-300')
I want to find the similar words for "AI" or "artifical intelligence", so I want to write
word2vec_model300.most_similar("artifical intelligence")
and I got errors
KeyError: "word 'artifical intelligence' not in vocabulary"
So what is the right way to extract similar words for bigram words?
Thanks in advance!
At one level, when a word-token isn't in a fixed set of word-vectors, the creators of that set of word-vectors chose not to train/model that word. So, anything you do will only be a crude workaround for its absence.
Note, though, that when Google prepared those vectors – based on a dataset of news articles from before 2012 – they also ran some statistical multigram-combinations on it, creating multigrams with connecting _ characters. So, first check if a vector for 'artificial_intelligence' might be present.
If it isn't, you could try other rough workarounds like averaging together the vectors for 'artificial' and 'intelligence' – though of course that won't really be what people mean by the distinct combination of those words, just meanings suggested by the independent words.
The Gensim .most_similar() method can take either a raw vectors you've created by operations such as averaging, or even a list of multiple words which it will average for you, as arguments via its explicit keyword positive parameter. For example:
word2vec_model300.most_similar(positive=[average_vector])
...or...
word2vec_model300.most_similar(positive=['artificial', 'intelligence'])
Finally, though Google's old vectors are handy, they're a bit old now, & from a particular domain (popular news articles) where senses may not match tose used in other domains (or more recently). So you may want to seek alternate vectors, or train your own if you have sufficient data from your area of interest, to have apprpriate meanings – including vectors for any particular multigrams you choose to tokenize in your data.
I am coming from Google BERT context (Bidirectional Encoder representations from Transformers). I have gone through architecture and codes. People say this is bidirectional by nature. To make it unidirectional attention some mask is to be applied.
Basically a transformer takes key, values and queries as input; uses encoder decoder architecture; and applies attention to these keys, queries and values. What I understood is we need to pass tokens explicitly rather than transformer understanding this by nature.
Can someone please explain what makes transformer is bidirectional by nature
Bidirectional is actually a carry-over term from RNN/LSTM. Transformer is much more than that.
Transformer and BERT can directly access all positions in the sequence, equivalent to having full random access memory of the sequence during encoding/decoding.
Classic RNN has only access to the hidden state and last token, e.g. encoding of word3 = f(hidden_state, word2), so it has to compress all previous words into a hidden state vector, theoretically possible but a severe limitation in practice. Bidirectional RNN/LSTM is slightly better. Memory networks is another way to work around this. Attention is yet another way to improve LSTM seq2seq models. The insight for Transformer is that you want full memory access and don't need the RNN at all!
Another piece of history: an important ingredient that let us deal with sequence structure without using RNN is positional encoding, which comes from CNN seq2seq model. It would not have been possible without this. Turns out, you don't need the CNN either, as CNN doesn't have full random access, but each convolution filter can only look at a number of neighboring words at a time.
Hence, Transformer is more like FFN, where encoding of word1 = f1(word1, word2, word3), and encoding of word3 = f2(word1, word2, word3). All positions available all the time.
You might also appreciate the beauty which is that the authors made it possible to compute attention for all positions in parallel, through the use of Q, K, V matrices. It's quite magical!
But understand this, you'll also appreciate the limitations of Transformer, that it requires O(N^2 * d) computation where N is the sequence length, precisely because we're doing N*N attention of all words with all other words. RNN, on the other hand, is linear in the sequence length and requires O(N * d^2) computation. d is the dimension of model hidden state.
Transformer just won't write a novel anytime soon!
On the following picture you will see in a really clear way why BERT is Bidirectional.
This is crucial since this forces the model to use information from the entire sentence simultaneously – regardless of the position – to make a good predictions.
BERT has been a clear break through allowed by the use of the notorious "attention is all you need" paper and architecture.
This Bidirectional idea (masked) is different from classic LSTM cells which till now used the forward or the backward method or both but not at the same time.
Edit:
this is done by the transformer. The attention is all you need paper is presenting an encoder-decoder system implementing a sequence to sequence framework. BERT is using this Transformer (sequence to sequence Bidirectional network) to do other NLP task. And this has been done by using a masked approach.
The important thing is: BERT uses Attention but Attention has been done for a translation and as such do not care for Bidirectional. But remove a word and you have Bidirectional.
So why BERT now?
well the Transformer is the first transduction model relying
entirely on self-attention to compute representations of its input and output without using sequencealigned RNNs or convolution. Meaning that this model allows a sentence Embedding far more effective than before. In fact, RNN based architectures are hard to parallelize and can have difficulty learning long-range dependencies within the input and output sequences. SO break through in architecture AND the use of this idea to train a network by masking a word (or more) leads to BERT.
Edit of Edit:
forget about the scale product, it's the inside the Attention which is inside A multi head attention itself inside the Transformer: you are looking to deep. The transformer is using the entire sequence every time to find the other sequence (In case of BERT it's the missing 0.15 percentage of the sentence) and that's it. The use of BERT as a language model is realy a transfer learning (see this)
As stated in your post, unidirectional can be done with a certain type of mask, bidirec is better. And it is used because the go from a full sentence to a full sentence but not the way classic Seq2seq is made (with LSTM and RNN) and as such can be used for LM.
BERT is a bidirectional transformer whereas the original transformer (Vaswani et al., 2017) is unidirectional. This can be shown by comparing the masks in the code.
Original Transformer
Tensorflow's tutorial is a good reference. The look_ahead_mask is what makes the Transformer unidirectional.
def create_look_ahead_mask(size):
mask = 1 - tf.linalg.band_part(tf.ones((size, size)), -1, 0)
return mask # (seq_len, seq_len)
If you trace the code, you can find the look_ahead_mask is applied to the attention_weights in the decoder. Basically each row in the attention_weights represents a attention query for token at certain position (the first row -> first token position; the second row -> second token position etc.). And the look_ahead_mask blacks out the tokens appear after this position in the decoder so it does not see the "future". In that sense, the decoder is unidirectional analogous to unidirectional in an RNN.
BERT
On the other hand, if you check the original BERT implementation (also in Tensorflow). There's only an optional input_mask applied to the entire BertModel. And if you follow the README on pre-training the model and run create_pretraining_data.py, you will observe that the input_mask is only used for padding the input sequence so for short sentences the unused tokens are ignored. Thus, attention in BERT can be applied to both the "past" and the "future" of a given token position. In that sense, the encoder in BERT is bidirectional analogous to bidirectional in an RNN.
I know this is an old post but for anyone coming back to this:
Adding to what Hai-Anh Trinh said, Transformers aren't 'bi-directional', it would be better to call them "omni-directional". Because of their self-attention method, they are able to consider every single word at the same time, simultaneously.
BERT on the other hand is "deeply bidirectional". This is because of the masked language model(MLM) pre-training objective that is used in BERT. (there are a lot of resources online, I can link some if need be)
It's easy to get confused so don't worry about it.
(https://arxiv.org/pdf/1810.04805.pdf; link to the original BERT paper)
(https://arxiv.org/pdf/1706.03762.pdf; link to the original Transformer paper)
I've been reading some NLP with Deep Learning papers and found Fine-tuning seems to be a simple but yet confusing concept. There's been the same question asked here but still not quite clear.
Fine-tuning pre-trained word embeddings to task-specific word embeddings as mentioned in papers like Y. Kim, “Convolutional Neural Networks for Sentence Classification,” and K. S. Tai, R. Socher, and C. D. Manning, “Improved Semantic Representations From Tree-Structured Long Short-Term Memory Networks,” had only a brief mention without getting into any details.
My question is:
Word Embeddings generated using word2vec or Glove as pretrained word vectors are used as input features (X) for downstream tasks like parsing or sentiment analysis, meaning those input vectors are plugged into a new neural network model for some specific task, while training this new model, somehow we can get updated task-specific word embeddings.
But as far as I know, during the training, what back-propagation does is updating the weights (W) of the model, it does not change the input features (X), so how exactly does the original word embeddings get fine-tuned? and where do these fine-tuned vectors come from?
Yes, if you feed the embedding vector as your input, you can't fine-tune the embeddings (at least easily). However, all the frameworks provide some sort of an EmbeddingLayer that takes as input an integer that is the class ordinal of the word/character/other input token, and performs a embedding lookup. Such an embedding layer is very similar to a fully connected layer that is fed a one-hot encoded class, but is way more efficient, as it only needs to fetch/change one row from the matrix on both front and back passes. More importantly, it allows the weights of the embedding to be learned.
So the classic way would be to feed the actual classes to the network instead of embeddings, and prepend the entire network with a embedding layer, that is initialized with word2vec / glove, and which continues learning the weights. It might also be reasonable to freeze them for several iterations at the beginning until the rest of the network starts doing something reasonable with them before you start fine tuning them.
One hot encoding is the base for constructing initial layer for embeddings. Once you train the network one hot encoding essentially serves as a table lookup. In fine-tuning step you can select data for specific works and mention variables that need to be fine tune when you define the optimizer using something like this
embedding_variables = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope="embedding_variables/kernel")
ft_optimizer = tf.train.AdamOptimizer(learning_rate=0.001,name='FineTune')
ft_op = ft_optimizer.minimize(mean_loss,var_list=embedding_variables)
where "embedding_variables/kernel" is the name of the next layer after one-hot encoding.
Core question : Right way(s) of using word-embeddings to represent text ?
I am building sentiment classification application for tweets. Classify tweets as - negative, neutral and positive.
I am doing this using Keras on top of theano and using word-embeddings (google's word2vec or Stanfords GloVe).
To represent tweet text I have done as follows:
used a pre-trained model (such as word2vec-twitter model) [M] to map words to their embeddings.
Use the words in the text to query M to get corresponding vectors. So if the tweet (T) is "Hello world" and M gives vectors V1 and V2 for the words 'Hello' and 'World'.
The tweet T can then be represented (V) as either V1+V2 (add vectors) or V1V2 (concatinate vectors)[These are 2 different strategies] [Concatenation means juxtaposition, so if V1, V2 are d-dimension vectors, in my example T is 2d dimension vector]
Then, the tweet T is represented by vector V.
If I follow the above, then My Dataset is nothing but vectors (which are sum or concatenation of word vectors depending on which strategy I use).
I am training a deepnet such as FFN, LSTM on this dataset. But my results arent coming out to be great.
Is this the right way to use word-embeddings to represent text ? What are the other better ways ?
Your feedback/critique will be of immense help.
I think that, for your purpose, it is better to think about another way of composing those vectors. The literature on word embeddings contains examples of criticisms to these kinds of composition (I will edit the answer with the correct references as soon as I find them).
I would suggest you to consider also other possible approaches, for instance:
Using the single word vectors as input to your net (I do not know your architecture, but the LSTM is recurrent so it can deal with sequences of words).
Using a full paragraph embedding (i.e. https://cs.stanford.edu/~quocle/paragraph_vector.pdf)
Summing them doesn't make any sense to be honest, because on summing them you get another vector which i don't think represents the semantics of "Hello World" or may be it does but it won't surely hold true for longer sentences in general
Instead it would be better to feed them as sequence as in that way it at least preserves sequence in meaningful way which seems to fit more to your problem.
e.g A hates apple Vs Apple hates A this difference would be captured when you feed them as sequence into RNN but their summation will be same.
I hope you get my point!