Google BERT and antonym detection - machine-learning

I recently learned about the following phenomenon: Google BERT word embeddings of well-known state-of-the-art models seem to ignore the measure of semantical contrast between antonyms in terms of the natural distance(norm2 or cosine distance) between the corresponding embeddings. For example:
The measure is the "cosine distance" (as oppose to the "cosine similarity"), that means closer vectors are supposed to have smaller distance between them. As one can see, BERT states "weak" and "powerful" to be closer than "strong" and "powerfull". The same behavior reproduces with the norm2 distance:
Is anyone familiar with this behavior and how one can overcome it and detect a contrast between word senses?

Related

Understanding embedding vectors dimension

In deep learning, in particularly NLP, words are transformed into a vector representation to be fed into a neural network such as an RNN. By referring to the link:
http://colah.github.io/posts/2014-07-NLP-RNNs-Representations/#Word%20Embeddings
In the section of Word Embeddings, it is said that:
A word embedding W:words→Rn is a paramaterized function mapping words
in
some language to high-dimensional vectors (perhaps 200 to 500 dimensions)
I do not understand the purpose of the dimension of the vectors. What does it mean to have a vector of 200 dimensions compared to a vector of 20 dimensions?
Does it improve the overall accuracy of the model? Could anyone give me a simple example regarding the choice of dimension of the vectors.
These word embeddings also called Distributed Word Embedding is based on
you know a word by the company it keeps
as quoted by John Rupert Firth
So we know the meaning of a word by its context. You can think of each scalar in the vector (of a word) represents its strength for a concept. This slide from Prof. Pawan Goyal explains it all.
So you want good vector size to capture decent amount of concepts but you do not want a too huge vector because it will then become the bottleneck in training of models where these embeddings are used.
Also the vector size is mostly fixed as most do not train their own embedding but rather use openly available embeddings as they are trained for many hours on huge data. So using them will force us to use an embedding layers with dimensions as given by the openly available embedding you are using (word2vec, glove etc)
Distributed Word Embeddings is a major milestone in the area of deep learning in NLP. They give better accuracy as compared of tfidf based embeddings.

What's the major difference between glove and word2vec?

What is the difference between word2vec and glove?
Are both the ways to train a word embedding? if yes then how can we use both?
Yes, they're both ways to train a word embedding. They both provide the same core output: one vector per word, with the vectors in a useful arrangement. That is, the vectors' relative distances/directions roughly correspond with human ideas of overall word relatedness, and even relatedness along certain salient semantic dimensions.
Word2Vec does incremental, 'sparse' training of a neural network, by repeatedly iterating over a training corpus.
GloVe works to fit vectors to model a giant word co-occurrence matrix built from the corpus.
Working from the same corpus, creating word-vectors of the same dimensionality, and devoting the same attention to meta-optimizations, the quality of their resulting word-vectors will be roughly similar. (When I've seen someone confidently claim one or the other is definitely better, they've often compared some tweaked/best-case use of one algorithm against some rough/arbitrary defaults of the other.)
I'm more familiar with Word2Vec, and my impression is that Word2Vec's training better scales to larger vocabularies, and has more tweakable settings that, if you have the time, might allow tuning your own trained word-vectors more to your specific application. (For example, using a small-versus-large window parameter can have a strong effect on whether a word's nearest-neighbors are 'drop-in replacement words' or more generally words-used-in-the-same-topics. Different downstream applications may prefer word-vectors that skew one way or the other.)
Conversely, some proponents of GLoVe tout that it does fairly well without needing metaparameter optimization.
You probably wouldn't use both, unless comparing them against each other, because they play the same role for any downstream applications of word-vectors.
Word2vec is a predictive model: trains by trying to predict a target word given a context (CBOW method) or the context words from the target (skip-gram method). It uses trainable embedding weights to map words to their corresponding embeddings, which are used to help the model make predictions. The loss function for training the model is related to how good the model’s predictions are, so as the model trains to make better predictions it will result in better embeddings.
The Glove is based on matrix factorization techniques on the word-context matrix. It first constructs a large matrix of (words x context) co-occurrence information, i.e. for each “word” (the rows), you count how frequently (matrix values) we see this word in some “context” (the columns) in a large corpus. The number of “contexts” would be very large, since it is essentially combinatorial in size. So we factorize this matrix to yield a lower-dimensional (word x features) matrix, where each row now yields a vector representation for each word. In general, this is done by minimizing a “reconstruction loss”. This loss tries to find the lower-dimensional representations which can explain most of the variance in the high-dimensional data.
Before GloVe, the algorithms of word representations can be divided into two main streams, the statistic-based (LDA) and learning-based (Word2Vec). LDA produces the low dimensional word vectors by singular value decomposition (SVD) on the co-occurrence matrix, while Word2Vec employs a three-layer neural network to do the center-context word pair classification task where word vectors are just the by-product.
The most amazing point from Word2Vec is that similar words are located together in the vector space and arithmetic operations on word vectors can pose semantic or syntactic relationships, e.g., “king” - “man” + “woman” -> “queen” or “better” - “good” + “bad” -> “worse”. However, LDA cannot maintain such linear relationship in vector space.
The motivation of GloVe is to force the model to learn such linear relationship based on the co-occurreence matrix explicitly. Essentially, GloVe is a log-bilinear model with a weighted least-squares objective. Obviously, it is a hybrid method that uses machine learning based on the statistic matrix, and this is the general difference between GloVe and Word2Vec.
If we dive into the deduction procedure of the equations in GloVe, we will find the difference inherent in the intuition. GloVe observes that ratios of word-word co-occurrence probabilities have the potential for encoding some form of meaning. Take the example from StanfordNLP (Global Vectors for Word Representation), to consider the co-occurrence probabilities for target words ice and steam with various probe words from the vocabulary:
As one might expect, ice co-occurs more frequently with solid than it
does with gas, whereas steam co-occurs more frequently with gas than
it does with solid.
Both words co-occur with their shared property water frequently, and both co-occur with the unrelated word fashion infrequently.
Only in the ratio of probabilities does noise from non-discriminative words like water and fashion cancel out, so that large values (much greater than 1) correlate well with properties specific to ice, and small values (much less than 1) correlate well with properties specific of steam.
However, Word2Vec works on the pure co-occurrence probabilities so that the probability that the words surrounding the target word to be the context is maximized.
In the practice, to speed up the training process, Word2Vec employs negative sampling to substitute the softmax fucntion by the sigmoid function operating on the real data and noise data. This emplicitly results in the clustering of words into a cone in the vector space while GloVe’s word vectors are located more discretely.

Document classification using word vectors

While I was classifying and clustering the documents written in natural language, I came up with a question ...
As word2vec and glove, and or etc, vectorize the word in distributed spaces, I wonder if there are any method recommended or commonly used for document vectorization USING word vectors.
For example,
Document1: "If you chase two rabbits, you will lose them both."
can be vectorized as,
[0.1425, 0.2718, 0.8187, .... , 0.1011]
I know about the one also known as doc2vec, that this document has n dimensions just like word2vec. But this is 1 x n dimensions and I have been testing around to find out the limits of using doc2vec.
So, I want to know how other people apply the word vectors for applications with steady size.
Just stacking vectors with m words will be formed m x n dimensional vectors. In this case, the vector dimension will not be uniformed since dimension m will depends on the number of words in document.
If: [0.1018, ... , 0.8717]
you: [0.5182, ... , 0.8981]
..: [...]
m th word: [...]
And this form is not favorable form to run some machine learning algorithms such as CNN. What are the suggested methods to produce the document vectors in steady form using word vectors?
It would be great if it is provided with papers as well.
Thanks!
The most simple approach to get a fixed-size vector from a text, when all you have is word-vectors, to average all the word-vectors together. (The vectors could be weighted, but if they haven't been unit-length-normalized, their raw magnitudes from training are somewhat of an indicator of their strength-of-single-meaning – polysemous/ambiguous words tend to have vectors with smaller magnitudes.) It works OK for many purposes.
Word vectors can be specifically trained to be better at composing like this, if the training texts are already associated with known classes. Facebook's FastText in its 'classification' mode does this; the word-vectors are optimized as much or more for predicting output classes of the texts they appear in, as they are for predicting their context-window neighbors (classic word2vec).
The 'Paragraph Vector' technique, often called 'doc2vec', gives every training text a sort-of floating pseudoword, that contributes to every prediction, and thus winds up with a word-vector-like position that may represent that full text, rather than the individual words/contexts.
There are many further variants, including some based on deeper predictive networks (eg 'Skip-thought Vectors'), or slightly different prediction targets (eg neighboring sentences in 'fastSent'), or other genericizations that can even include a mixture of symbolic and numeric inputs/targets during training (an option in Facebook's StarSpace, which explores other entity-vectorization possibilities related to word-vectors and FastText-like classification needs).
If you don't need to collapse a text to fixed-size vectors, but just compare texts, there are also techniques like "Word Mover's Distance" which take the "bag of word-vectors" for one text, and another, and give a similarity score.

What “information” in document vectors makes sentiment prediction work?

Sentiment prediction based on document vectors works pretty well, as examples show:
https://github.com/RaRe-Technologies/gensim/blob/develop/docs/notebooks/doc2vec-IMDB.ipynb
http://linanqiu.github.io/2015/10/07/word2vec-sentiment/
I wonder what pattern is in the vectors making that possible. I thought it should be similarity of vectors making that somehow possible. Gensim similarity measures rely on cosine similarity. Therefore, I tried the following:
Randomly initialised a fix “compare” vector, get cosine similarity of the “compare” vector with all other vectors in training and test set, use the similarities and the labels of the train set to estimate a logistic regression model, evaluate the model with the test set.
Looks like this, where train/test_arrays contain document vectors and train/test_labels labels either 0 or 1. (Notice, document vectors are obtained from genism doc2vec and are well trained, predicting the test set 80% right if directly used as input for the logistic regression):
fix_vec = numpy.random.rand(100,1)
def cos_distance_to_fix(x):
return scipy.spatial.distance.cosine(fix_vec, x)
train_arrays_cos = numpy.reshape(numpy.apply_along_axis(cos_distance_to_fix, axis=1, arr=train_arrays), newshape=(-1,1))
test_arrays_cos = numpy.reshape(numpy.apply_along_axis(cos_distance_to_fix, axis=1, arr=test_arrays), newshape=(-1,1))
classifier = LogisticRegression()
classifier.fit(train_arrays_cos, train_labels)
classifier.score(test_arrays_cos, test_labels)
It turns out, that this approach does not work, predicting the test set only to 50%....
So, my question is, what “information” is in the vectors, making the prediction based on vectors work, if it is not the similarity of vectors? Or is my approach simply not possible to capture similarity of vectors correct?
This is less a question about Doc2Vec than about machine-learning principles with high-dimensional data.
Your approach is collapsing 100-dimensions to a single dimension – the distance to your random point. Then, you're hoping that single dimension can still be predictive.
And roughly all LogisticRegression can do with that single-valued input is try to pick a threshold-number that, when your distance is on one side of that threshold, predicts a class – and on the other side, predicts not-that-class.
Recasting that single-threshold-distance back to the original 100-dimensional space, it's essentially trying to find a hypersphere, around your random point, that does a good job collecting all of a single class either inside or outside its volume.
What are the odds your randomly-placed center-point, plus one adjustable radius, can do that well, in a complex high-dimensional space? My hunch is: not a lot. And your results, no better than random guessing, seems to suggest the same.
The LogisticRegression with access to the full 100-dimensions finds a discriminating-frontier for assigning the class that's described by 100 coefficients and one intercept-value – and all of those 101 values (free parameters) can be adjusted to improve its classification performance.
In comparison, your alternative LogisticRegression with access to only the one 'distance-from-a-random-point' dimension can pick just one coefficient (for the distance) and an intercept/bias. It's got 1/100th as much information to work with, and only 2 free parameters to adjust.
As an analogy, consider a much simpler space: the surface of the Earth. Pick a 'random' point, like say the South Pole. If I then tell you that you are in an unknown place 8900 miles from the South Pole, can you answer whether you are more likely in the USA or China? Hardly – both of those 'classes' of location have lots of instances 8900 miles from the South Pole.
Only in the extremes will the distance tell you for sure which class (country) you're in – because there are parts of the USA's Alaska and Hawaii further north and south than parts of China. But even there, you can't manage well with just a single threshold: you'd need a rule which says, "less than X or greater than Y, in USA; otherwise unknown".
The 100-dimensional space of Doc2Vec vectors (or other rich data sources) will often only be sensibly divided by far more complicated rules. And, our intuitions about distances and volumes based on 2- or 3-dimensional spaces will often lead us astray, in high dimensions.
Still, the Earth analogy does suggest a way forward: there are some reference points on the globe that will work way better, when you know the distance to them, at deciding if you're in the USA or China. In particular, a point at the center of the US, or at the center of China, would work really well.
Similarly, you may get somewhat better classification accuracy if rather than a random fix_vec, you pick either (a) any point for which a class is already known; or (b) some average of all known points of one class. In either case, your fix_vec is then likely to be "in a neighborhood" of similar examples, rather than some random spot (that has no more essential relationship to your classes than the South Pole has to northern-Hemisphere temperate-zone countries).
(Also: alternatively picking N multiple random points, and then feeding the N distances to your regression, will preserve more of the information/shape of the original Doc2Vec data, and thus give the classifier a better chance of finding a useful separating-threshold. Two would likely do better than your one distance, and 100 might approach or surpass the 100 original dimensions.)
Finally, some comment about the Doc2Vec aspect:
Doc2Vec optimizes vectors that are somewhat-good, within their constrained model, at predicting the words of a text. Positive-sentiment words tend to occur together, as do negative-sentiment words, and so the trained doc-vectors tend to arrange themselves in similar positions when they need to predict similar-meaning-words. So there are likely to be 'neighborhoods' of the doc-vector space that correlate well with predominantly positive-sentiment or negative-sentiment words, and thus positive or negative sentiments.
These won't necessarily be two giant neighborhoods, 'positive' and 'negative', separated by a simple boundary –or even a small number of neighborhoods matching our ideas of 3-D solid volumes. And many subtleties of communication – such as sarcasm, referencing a not-held opinion to critique it, spending more time on negative aspects but ultimately concluding positive, etc – mean incursions of alternate-sentiment words into texts. A fully-language-comprehending human agent could understand these to conclude the 'true' sentiment, while these word-occurrence based methods will still be confused.
But with an adequate model, and the right number of free parameters, a classifier might capture some generalizable insight about the high-dimensional space. In that case, you can achieve reasonably-good predictions, using the Doc2Vec dimensions – as you've seen with the ~80%+ results on the full 100-dimensional vectors.

Document similarity: Vector embedding versus Tf-Idf performance?

I have a collection of documents, where each document is rapidly growing with time. The task is to find similar documents at any fixed time. I have two potential approaches:
A vector embedding (word2vec, GloVe or fasttext), averaging over word vectors in a document, and using cosine similarity.
Bag-of-Words: tf-idf or its variations such as BM25.
Will one of these yield a significantly better result? Has someone done a quantitative comparison of tf-idf versus averaging word2vec for document similarity?
Is there another approach, that allows to dynamically refine the document's vectors as more text is added?
doc2vec or word2vec ?
According to article, the performance of doc2vec or paragraph2vec is poor for short-length documents. [Learning Semantic Similarity for Very Short Texts, 2015, IEEE]
Short-length documents ...?
If you want to compare the similarity between short documents, you might want to vectorize the document via word2vec.
how construct ?
For example, you can construct a document vector with a weighted average vector using tf-idf.
similarity measure
In addition, I recommend using ts-ss rather than cosine or euclidean for similarity.
Please refer to the following article or the summary in github below.
"A Hybrid Geometric Approach for Measuring Similarity Level Among Documents and Document Clustering"
https://github.com/taki0112/Vector_Similarity
thank you
You have to try it: the answer may vary based on your corpus and application-specific perception of 'similarity'. Effectiveness may especially vary based on typical document lengths, so if "rapidly growing with time" also means "growing arbitrarily long", that could greatly affect what works over time (requiring adaptations for longer docs).
Also note that 'Paragraph Vectors' – where a vector is co-trained like a word vector to represent a range-of-text – may outperform a simple average-of-word-vectors, as an input to similarity/classification tasks. (Many references to 'Doc2Vec' specifically mean 'Paragraph Vectors', though the term 'Doc2Vec' is sometimes also used for any other way of turning a document into a single vector, like a simple average of word-vectors.)
You may also want to look at "Word Mover's Distance" (WMD), a measure of similarity between two texts that uses word-vectors, though not via any simple average. (However, it can be expensive to calculate, especially for longer documents.) For classification, there's a recent refinement called "Supervised Word Mover's Distance" which reweights/transforms word vectors to make them more sensitive to known categories. With enough evaluation/tuning data about which of your documents should be closer than others, an analogous technique could probably be applied to generic similarity tasks.
You also might consider trying Jaccard similarity, which uses basic set algebra to determine the verbal overlap in two documents (although it is somewhat similar to a BOW approach). A nice intro on it can be found here.

Resources