Data modeling, restructuring, and massaging: Difference between revisions
mNo edit summary |
|||
(17 intermediate revisions by the same user not shown) | |||
Line 1: | Line 1: | ||
{{#addbodyclass:tag_math}} | |||
{{Math notes}} | {{Math notes}} | ||
Line 267: | Line 268: | ||
==== | ====Computers and people and numbers==== | ||
<!-- | <!-- | ||
Where people are good at words and bad at numbers, computers are good at numbers and bad at words. | Where people are good at words and bad at numbers, computers are good at numbers and bad at words. | ||
Line 274: | Line 275: | ||
So it makes sense to express words ''as'' numbers? | So it makes sense to express words ''as'' numbers? | ||
That does go a moderate way, but it really matters ''how''. | |||
Line 353: | Line 354: | ||
--> | --> | ||
====vector space representations, word embeddings, and more==== | |||
<!-- | |||
==== | Around text, '''vector space representations''' are the the general idea that for each word (or similar units) you | ||
can calculate something that you an meaningfully compare to other. | |||
--> | |||
=====Just count in a big table===== | |||
<!-- | <!-- | ||
'''You could just count.'''[https://scikit-learn.org/stable/modules/generated/sklearn.feature_extraction.text.CountVectorizer.html] Given a collection of documents | |||
* decide what words to include | |||
* make a big documents-by-words table {{comment|(we like to say 'matrix' instead of 'table' when we get mathy, but it's the same concept really)}} | |||
* count | |||
One limitation is that including all words makes a humongous table {{comment|(and most cells will contain zero)}}. | |||
Yet not including them means we say they do not exist ''at all''. | |||
Another limitation is that counts ''as such'' are not directly usable, for dumb reasons like that | |||
if more count means more important, consider longer documents will have higher counts just because they are longer, | |||
and that 'the' and 'a' will be most important. | |||
'''You could use something like [[tf-idf]]'''[https://scikit-learn.org/stable/modules/generated/sklearn.feature_extraction.text.TfidfVectorizer.html#sklearn.feature_extraction.text.TfidfVectorizer], an extra step on top of the previous, e.g. | |||
* will downweigh 'the' by merit about being absolutely everywhere | |||
* reduces the effect of document length | |||
Another limitation is that that table's other axis being '' 'all' documents' seems huge for no good reason. | |||
are | There are plenty of tasks where those original training documents (that happened to be a training task long ago) are just... not useful. | ||
Sure there is information in those documents, but can't we learn from it, then put it in a more compact form? | |||
Another limitation is that these still focus on unique words. Even words that are inflections of others will be treated as entirely independent, | |||
so if one document uses only 'fish' and 'spoon', it might even be treated as entirely distint from one that says only 'fishing' and 'spooning'. | |||
Never that bad, but you can intuit why this isn't useful. | |||
One thing many tried is to use something like matrix methods - factorization, dimensionality reduction, and the likes. | |||
You don't need to know how the math works, but the point is that similarly expressed terms can be recognized as such, | |||
and e.g. 'fish' and 'fishing' will smush into each other, automatically because otherwise-similar documents will mention both | |||
{{comment|(looking from the perspective of the other axis, mayne we can also compare documents better, but unless comparing documents was your goal, that was actually something we were trying to get rid of)}} | |||
--> | |||
=====Word embeddings===== | |||
<!-- | |||
The above arrived in an area where those vectors no longer represent just the words, | |||
but contain some comparability to similar words. | |||
There are some older distributional similarity approaches - these were a little clunky in that | |||
they made for high dimensional, sparse vectors, where each one represented a specific context. | |||
They were sometimes more explainable, but somewhat unwieldy. | |||
'''"Word embeddings"''' often refer the word vector thing ''but more compactly'' somehow: | |||
we try to some somehow figure out a dense, semantically useful vector. | |||
: ''Dense'' compared to the input: when you can see text as 'one of ~100K possible words', then a vector of maybe two hundred dimensions that seems to do a good job packin enough meaning for good (dis)similarity comparisons is pretty good | |||
: ''Semantically useful'' in that those properties tend to be useful | |||
:: often focusing on such (dis)similarity comparisons - e.g. 'bat' may get a sense of animal, tool, and verb, and noun -- or rather, in this space appear close to animals and tools and verbs than e.g. 'banana' does. | |||
This amounts to even more up-front work, so why do this dense semantic thing? | |||
One is practical - classical methods run into some age-old machine learning problems like high dimensionality, sparsity, | |||
and there happen to be some word embedding methods that sidestep that (by cheating, but by cheating ''pretty well'' ). | |||
Also, putting all words in a single space lets us compare terms, sentences, and documents. | |||
If the vectors are good, this is a good approximation of ''semantic'' similarity. | |||
If we can agree on a basis between uses (e.g. build a reasonable set of vectors per natural language) | |||
we might even be able to give a basic idea of e.g. what a document is about. | |||
...that one turns out to be optimistic, for a slew of reasons. | |||
You can often get better results out of becoming a little domain-specific. | |||
Which them makes your vectors specific to just your system again. | |||
'''So how do you get these dense semantic vectors?''' | |||
There are varied ways to do this. | |||
You ''could'' start with well-annotated data, and that might be of higher quality in the end, | |||
but it is hard to come by annotation for many aspects over an entire language, | |||
It's a lot of work to even try - and that's still ignoring details like contextual ambiguity, | |||
analyses even people wouldn't quite agree on, the fact you have to impose a particular system | |||
so if it doesn't encode something you wanted, you have to do it ''again'' later. | |||
A recent trend is to put a little more trust in the assumptions of the [[distributional hypothesis]], | |||
e.g. that words in similar context will be comparable, | |||
and focus on words in context. | |||
For which we can use non-annotated data. We need a ''lot'' more of it for comparable quality, | |||
but people have collective produced a lot of text, internet and other. | |||
This was done in varied ways over the years (e.g. [[Latent Semantic Analysis]] applies somewhat), | |||
later with more complex math, and/or neural nets. | |||
Which may just train better - the way you handle its output is much the same. | |||
One of the techniques that kicked this off in more recent years is [[word2vec]], | |||
which doesn't do a lot more than looking what appears in similar contexts. | |||
(Its view is surprisingly narrow - what happens in a small window in a ''lot'' of data will tend to be more consistent. | |||
A larger window is not only more work but often too fuzzy) | |||
Its math is apparently {{search|Levy "Neural Word Embedding as Implicit Matrix Factorization"|fairly like the classical matrix factorization}}. | |||
'''Word embeddings''' often refers to learning vectors from context, | |||
though there are more varied meanings (some conflicting), | |||
so you may wish to read 'embeddings' as 'text vectors' and figure out yourself what the implementation actually is. | |||
Note that | |||
* you won't really know what these vectors mean. | |||
: You can fish this out, to a degree, because you can compare vectors. | |||
: say, if any given vector compares much better to 'hammer' or to 'see' (e.g. from examples sentences), you can start to figure out what it meant. | |||
* the thing you train does not necessarily | |||
--> | --> | ||
==== | =====Static embeddings===== | ||
<!-- | <!-- | ||
In the context of some of the later developments, the simpler variant implmentations are considered static vectors. | |||
'''Static vectors''' refer to systems where each word alwys gets the same vector. | |||
That usually means: | |||
* make vocabulary: each word gets an entry | |||
* learn vector for each item in the vocabulary | |||
'''Limits of static embeddings''' | |||
We previously mentioned that putting a single number on a word has issues. | |||
" | |||
We now point out that putting a single vector for a word have some issues. | |||
Yes, a lot of words will get a completely sensible sense. | |||
Yet if we made an enumerated vocabulary, and each item gets a vector, then the saw in "I sharpened the saw" and "we saw a bat" ''will'' be assigned the same vector; same for the bat in "I saw a bat" and "I'll bat an eyelash". | |||
The problem is that ''if'' that vector is treated as the semantic sense, both saws have exactly the same sense. | |||
...probably both the tool quality and the seeing quality, | |||
and any use will tell you it's slightly about tools | and any use will tell you it's slightly about tools | ||
{{comment|(by count, most 'saw's are the seeing, something that even unsupervised learning should tell you)}}. | {{comment|(by count, most 'saw's are the seeing, something that even unsupervised learning should tell you)}}. | ||
You can imagine it's not really an issue for words like antidisestablishmentarianism. | You can imagine it's not really an issue for words like antidisestablishmentarianism. | ||
Line 410: | Line 533: | ||
But saw, hm. | But saw, hm. | ||
In theory, it may assign | |||
* a vector to the verb saw that is more related to seeing than to cutting | |||
* a vector to the noun bat that is more related to other animals than it is to sports equipment | |||
That said, that is not a given. | |||
: There's at least four options and it might land on any of them, particularly for a sentence in isolation | |||
: the meanings we propose to be isolated here may get weirdly blended in training | |||
But for similar reasons, if you {{example|table a motion}}, it ''will'' associate in gestures and woodworking, because those are the more common things. | |||
'''"Can't you just have multiple saws, encode 'saw as a verb' differently from 'saw as an implement'?"''' | |||
Sure, now you can ''store'' that, but how do you decide? | Sure, now you can ''store'' that, but how do you decide? | ||
Line 514: | Line 649: | ||
In fact, some patterns are strong enough that even unseen words will get a decent estimation. | In fact, some patterns are strong enough that even unseen words will get a decent estimation. | ||
Say, give we spoiaued to an embedding-style parser, and it's going to guess it's a verb ''while also'' pointing out it's out of its vocabulary. | Say, give we spoiaued to an embedding-style parser, and it's going to guess it's a verb ''while also'' pointing out it's out of its vocabulary. | ||
Line 563: | Line 658: | ||
=====Contextual word embeddings===== | =====Contextual word embeddings===== | ||
<!-- | <!-- | ||
The first attempts at word embeddings, and many since, | The first attempts at word embeddings, and many since, | ||
Line 577: | Line 673: | ||
Line 622: | Line 709: | ||
Say, the probably-out-of-vocabulary apploid may get a decent guess | Say, the probably-out-of-vocabulary apploid may get a decent guess | ||
if we learned a vector for appl from e.g. apple. | if we learned a vector for appl from e.g. apple. | ||
Also, it starts dealing with misspellings a lot better. | |||
Understanding the language's morphology would probably do a little better, | Understanding the language's morphology would probably do a little better, | ||
Line 627: | Line 716: | ||
in part because inflection, compositional agglutination (e.g. turkish) | in part because inflection, compositional agglutination (e.g. turkish) | ||
and such are often ''largely'' regular. | and such are often ''largely'' regular. | ||
Line 634: | Line 724: | ||
For this reason it's often combined with [[bloom embeddings]]. | For this reason it's often combined with [[bloom embeddings]]. | ||
Examples: | Examples: | ||
fastText | fastText, floret, | ||
[https://d2l.ai/chapter_natural-language-processing-pretraining/subword-embedding.html] | [https://d2l.ai/chapter_natural-language-processing-pretraining/subword-embedding.html] | ||
Line 643: | Line 735: | ||
--> | --> | ||
=====Bloom embeddings===== | =====The hashing trick (also, Bloom embeddings)===== | ||
<!-- | <!-- | ||
The '''hashing trick''' works for everything from basic counting to contextual and sub-word embeddings -- just anywhere where you need to put a fixed bound on, and are willing to accept degrading performance beyond that. | |||
'''You can use the [[hashing trick]]'''[https://scikit-learn.org/stable/modules/generated/sklearn.feature_extraction.text.HashingVectorizer.html#sklearn.feature_extraction.text.HashingVectorizer] if you want to put an upper bound on memory -- but this comes at an immediate cost that may be avoided with a more clever plan. | |||
* squeezes all words into a fixed amount of entries | |||
* ...almost indiscriminantly, so the more words you push into fewer entries, the less accurate it is. | |||
* it's better than nothing given low amounts of memory, but there are usually better options | |||
The hashing trick can be applied to word embeddings, | |||
is sometimes called Bloom embeddings, because it's an idea ''akin'' to [[bloom filter]]s. | |||
But probably more because it's shorter than "embeddings with the hashing trick". | |||
Consider a language model that should be assigning vectors. | |||
When it sees [[out-of-vocabulary]] words, what do you do? | |||
Do you treat them as not existing at all? | |||
: ideally, we could do something quick and dirty that is better than nothing. | |||
Do you add just as many entries to the vocabulary? | Do you add just as many entries to the vocabulary? | ||
Line 657: | Line 767: | ||
Do you map all to a single unknown vector? | Do you map all to a single unknown vector? | ||
That's small, but makes them ''by definition'' indistinguishable. | That's small, but makes them ''by definition'' indistinguishable. | ||
If you wanted to do even a ''little'' extra contextual learning for them, | If you wanted to do even a ''little'' extra contextual learning for them, | ||
Line 669: | Line 780: | ||
* hope that the words that get assigned together aren't ''quite'' as conflicting as ''all at once''. | * hope that the words that get assigned together aren't ''quite'' as conflicting as ''all at once''. | ||
It's a ''very'' rough, | It's a ''very'' rough, | ||
* it is definitely better than nothing. | |||
* with a little forethought you could sort of share these vectors between documents | |||
:: in that the same words will map to the same entry every time | |||
Limitations: | |||
* when you smush things together and use what you previously learned | |||
* when you smush things together and ''learn'', you relate unrelated things | |||
This sort of bloom-like intermediate is also applied to subword embeddings, | |||
because it gives a ''sliding scale'' between | |||
'so large that it probably won't fit in RAM' and 'so smushed together it has become too fuzzy' | |||
some | |||
* [https://github.com/explosion/floret floret] (bloom embeddings for fastText) | |||
* thinc's HashEmbed [https://thinc.ai/docs/api-layers#hashembed] | |||
* spaCy’s MultiHashEmbed and HashEmbedCNN (which uses thinc's HashEmbed) | |||
Line 685: | Line 812: | ||
https://spacy.io/usage/v3-2#vectors | https://spacy.io/usage/v3-2#vectors | ||
--> | |||
=====Now we have nicer numbers, but how how I ''use'' them?===== | |||
<!-- | |||
* use the vectors as-is | |||
* adapt the embeddings with your own training | |||
:: starts with a good basis, refines for your use | |||
:: but: only deals with tokens already in there | |||
* there are also some ways to selectively alter vectors | |||
:: can be useful if you want to keep sharing the underling embeddings | |||
--> | |||
=====vectors - unsorted===== | |||
<!-- | |||
These vectors come from machine learning (of varying type and complexity). | |||
* LSA | |||
* [[word2vec]] - | |||
: technically patented? | |||
: T Mikolov et al. (2013) "Efficient Estimation of Word Representations in Vector Space" | |||
* tok2vec | |||
* FastText | |||
: https://fasttext.cc/ | |||
* lda2vec | |||
: https://multithreaded.stitchfix.com/blog/2016/05/27/lda2vec/#topic=38&lambda=1&term= | |||
All of this may still apply a single vector to the same word always (sometimes called static word embeddings). | |||
This is great for unambiguous content words, but less so for polysemy and | |||
--> | --> | ||
Line 761: | Line 930: | ||
word2vec is one of many ways to put semantic vectors to words (in the [[distributional hypothesis]] approach), | word2vec is one of many ways to put semantic vectors to words (in the [[distributional hypothesis]] approach), | ||
and refers to two techniques, using either [[bag-of-words]] and [[skip-gram]] as processing for a specific learner, | and refers to two techniques, using either [[bag-of-words]] and [[skip-gram]] as processing for a specific learner, | ||
as described in T Mikolov et al. (2013), "{{search|Efficient Estimation of Word Representations in Vector Space}}" | as described in T Mikolov et al. (2013), "{{search|Efficient Estimation of Word Representations in Vector Space}}", | ||
probably the one that kicked off this dense-vector idea into the interest. | |||
Word2vec amounts could be seen as building a classifier that predicts what word apear in a context, and/or what context appears around a word, | |||
which happens to do a decent task of classifying that word. | |||
That paper mentions | That paper mentions | ||
Line 770: | Line 945: | ||
but this may come from not really reading the paper you're copy-pasting the image from? | but this may come from not really reading the paper you're copy-pasting the image from? | ||
:: seems to be better at less-common words, but slower | :: seems to be better at less-common words, but slower | ||
NN implies [[one-hot]] coding, so not small, but it turns out to be moderately efficient{{verify}} | (NN implies [[one-hot]] coding, so not small, but it turns out to be moderately efficient{{verify}}) | ||
Latest revision as of 23:14, 21 April 2024
This is more for overview of my own than for teaching or exercise.
|
Intro
NLP data massage / putting meanings or numbers to words
Bag of words / bag of features
The bag-of-words model (more broadly bag-of-features model) use the collection of words in a context, unordered, in a multiset, a.k.a. bag.
In other words, we summarize a document (or part of it) it by appearance or count of words, and ignore things like adjacency and order - so any grammar.
In text processing
In introductions to Naive Bayes as used for spam filtering, its naivety essentially is this assumption that feature order does not matter.
Though real-world naive bayes spam filtering would take more complex features than single words (and may re-introduce adjacenct via n-grams or such), examples often use 1-grams for simplicity - which basically is bag of words, exc.
Other types of classifiers also make this assumption, or make it easy to do so.
Bag of features
While the idea is best known from text, hence bag-of-words, you can argue for bag of features, applying it to anything you can count, and may be useful even when considered independently.
For example, you may follow up object detection in an image with logic like "if this photo contains a person, and a dog, and grass" because each task may be easy enough individually, and the combination tends to narrow down what kind of photo it is.
In practice, the bag-of-features often refers to models that recognize parts of a whole object (e.g. "we detected a bunch of edges of road signs" might be easier and more robust than detecting it fully), and used in a number image tasks, such as feature extraction, object/image recognition, image search, (more flexible) near-duplicate detection, and such.
The idea that you can describe an image by the collection of small things we recognize in it, and that combined presence is typically already a strong indicator (particularly when you add some hypothesis testing). Exact placement can be useful, but often easily secondary.
See also:
N-gram notes
N-grams are contiguous sequence of length n.
They are most often seen in computational linguistics.
Applied to sequences of characters it can be useful e.g. in language identification,
but the more common application is to words.
As n-grams models only include dependency information when those relations are expressed through direct proximity, they are poor language models, but useful to things working off probabilities of combinations of words, for example for statistical parsing, collocation analysis, text classification, sliding window methods (e.g. sliding window POS tagger), (statistical) machine translation, and more
For example, for the already-tokenized input This is a sentence . the 2-grams would be:
- This is
- is a
- a sentence
- sentence .
...though depending on how special you do or do not want to treat the edges, people might fake some empty tokens at the edge, or some special start/end tokens.
Skip-grams
Note: Skip-grams seem to refer to now two different things.
An extension of n-grams where components need not be consecutive (though typically stay ordered).
A k-skip-n-gram is a length-n sequence where the components occur at distance at most k from each other.
They may be used to ignore stopwords, but perhaps more often they are intended to help reduce data sparsity, under a few assumptions.
They can help discover patterns on a larger scale, simply because skipping makes you look further for the same n. (also useful for things like fuzzy hashing).
Skip-grams apparently come from speech analysis, processing phonemes.
In word-level analysis their purpose is a little different. You could say that we acknowledge the sparsity problem, and decide to get more out of the data we have (focusing on context) rather than trying to smooth.
Actually, if you go looking, skip-grams are now often equated with a fairly specific analysis.
Syntactic n-grams
Flexgrams
Words as features - one-hot coding and such
Putting numbers to words
Computers and people and numbers
vector space representations, word embeddings, and more
Just count in a big table
Word embeddings
Static embeddings
Contextual word embeddings
Subword embeddings
The hashing trick (also, Bloom embeddings)
Now we have nicer numbers, but how how I use them?
vectors - unsorted
Moderately specific ideas and calculations
Collocations
Collocations are statistically idiosyncratic sequences - the math that is often used asks "do these adjacent words occur together more often than the occurrence of each individually would suggest?".
This doesn't ascribe any meaning, it just tends to signal anything from empty habitual etiquette, jargon, various substituted phrases, and many other things that go beyond purely compositional construction, because why other than common sentence structures would they co-occur so often?
...actually, there are varied takes on how useful collocations are, and why.
latent semantic analysis
Latent Semantic Analysis (LSA) is the application of Singular Value Decomposition on text analysis and search.
random indexing
https://en.wikipedia.org/wiki/Random_indexing
Topic modeling
Roughly the idea given documents that are about a particular topic, one would expect particular words to appear in the each more or less frequently.
Assuming such documents sharing topics, you can probably find groups of words that belong to those topics.
Assuming each document is primarily about one topic, you can expect a larger set of documents to yield multiple topics, and an assignment of one or more of these topics, so act like a soft/fuzzy clustering.
This is a relatively weak proposal in that it relies on a number of assumptions, but given that it requires zero training, it works better than you might expect when those assumptions are met. (the largest probably being your documents having singular topics).
https://en.wikipedia.org/wiki/Topic_model
word2vec
word2vec is one of many ways to put semantic vectors to words (in the distributional hypothesis approach), and refers to two techniques, using either bag-of-words and skip-gram as processing for a specific learner, as described in T Mikolov et al. (2013), "Efficient Estimation of Word Representations in Vector Space", probably the one that kicked off this dense-vector idea into the interest.
Word2vec amounts could be seen as building a classifier that predicts what word apear in a context, and/or what context appears around a word,
which happens to do a decent task of classifying that word.
That paper mentions
- its continuous bag of words (cbow) variant predicts the current word based on the words directly around it (ignoring order, hence bow(verify))
- its continuous skip-gram variant predicts surrounding words given the current word.
- Uses skip-grams as a concept/building block. Some people refer to this technique as just 'skip-gram' without the 'continuous',
but this may come from not really reading the paper you're copy-pasting the image from?
- seems to be better at less-common words, but slower
(NN implies one-hot coding, so not small, but it turns out to be moderately efficient(verify))