Unknown Words in N-Gram Modelling - nlp

What is the logic for grouping unknown words under the same token i.e., <UNK> and also include words with small probabilities?
Won't some rare words get assigned high probabilities if the <UNK> set grows in size?
This might work if all the <UNK> words belong to the same class in some sense, for example, proper nouns such as John, Tim, Sam can all use each other's probability as bi-grams "Hello John, Hello Tim, Hello Sam" are equally likely. But if this is not the case, won't this method run into problems?

Mapping rare words to <UNK> simply means that we delete those words and replace them with the token <UNK> in the training data. Thus our model does not know of any rare words. It is a crude form of smoothing because the model assumes that the token <UNK> will never actually occur in real data or better yet it ignores these n-grams altogether.
The problem that smoothing is trying to solve is data sparsity. This technique is probably the simplest way to deal with it. However, we can do better as #alvas shows in the comments.

Related

Idenfying bigrams using Gensim Phraser that contain the word "not," for sentiment analysis

I am working on a sentiment analysis project where I am analyzing a corpus of documents, and I am specifically not removing the word "not" as a stopword, so that I can use it to determine if a text agrees or disagrees with something. For instance, there is a difference between "not effective" and "effective" when discussing the COVID vaccine.
However, my phraser is not identifying any bigrams with the word "not." I presume this is because that token exists in such large numbers (particularly because I expanded contractions, so "isn't" -> "is not"), that the scoring function simply scores all bigrams with "not" too low. This would be because the standard phrase scoring function is:
(where min_count is a hyper parameter)
So, since "not" exists many thousands of times in the database, worda_count will be very large, leading to a large denominator and dropping the score considerably.
Is there a way to get around this, so "not" bigrams are scored effectively?
I can think of a few options off the top of my head:
Write my own scoring function that effectively has two scoring formula: the standard scoring formula, and a different scoring formula if the first word is "not".
I could include "not" in a list of connector_words, but gensim.models.phrases.Phraser specifically indicates that these connector words cannot be at the beginning or end of a phrase.
As you've discovered, the Phrases functionality in Gensim is pretty crude: it only combines words based on a meaning-oblivious statistical analysis. It's more likely to be helpful in promoting certain noun-phrases ('new_york') or idioms than generic syntactical reversals-of-meaning (as with an added 'not'). So whether you'll want to use it at all, I'm not sure.
You could try the most simpleminded thing possible: preprocess to always attach 'not' to the following word. Maybe it'll help!
You could also try some expensive grammar-aware preprocessing - the sort that labels words with parts-of-speech, & further identifies which other words/word-ranges a particular 'not' modifies. That might allow you to condiionally connect the 'not' to later words – maybe even non-contiguous words – & perhaps that will provide a lift to downstream sentiment-analysis.

Why word embedding technique works

I have look into some word embedding techniques, such as
CBOW: from context to single word. Weight matrix produced used as embedding vector
Skip gram: from word to context (from what I see, its acutally word to word, assingle prediction is enough). Again Weight matrix produced used as embedding
Introduction to these tools would always quote "cosine similarity", which says words of similar meanning would convert to similar vector.
But these methods all based on the 'context', account only for words around a target word. I should say they are 'syntagmatic' rather than 'paradigmatic'. So why the close in distance in a sentence indicate close in meaning? I can think of many counter example that frequently occurs
"Have a good day". (good and day are vastly different, though close in distance).
"toilet" "washroom" (two words of similar meaning, but a sentence contains one would unlikely to contain another)
Any possible explanation?
This sort of "why" isn't a great fit for StackOverflow, but some thoughts:
The essence of word2vec & similar embedding models may be compression: the model is forced to predict neighbors using far less internal state than would be required to remember the entire training set. So it has to force similar words together, in similar areas of the parameter space, and force groups of words into various useful relative-relationships.
So, in your second example of 'toilet' and 'washroom', even though they rarely appear together, they do tend to appear around the same neighboring words. (They're synonyms in many usages.) The model tries to predict them both, to similar levels, when typical words surround them. And vice-versa: when they appear, the model should generally predict the same sorts of words nearby.
To achieve that, their vectors must be nudged quite close by the iterative training. The only way to get 'toilet' and 'washroom' to predict the same neighbors, through the shallow feed-forward network, is to corral their word-vectors to nearby places. (And further, to the extent they have slightly different shades of meaning – with 'toilet' more the device & 'washroom' more the room – they'll still skew slightly apart from each other towards neighbors that are more 'objects' vs 'places'.)
Similarly, words that are formally antonyms, but easily stand-in for each-other in similar contexts, like 'hot' and 'cold', will be somewhat close to each other at the end of training. (And, their various nearer-synonyms will be clustered around them, as they tend to be used to describe similar nearby paradigmatically-warmer or -colder words.)
On the other hand, your example "have a good day" probably doesn't have a giant influence on either 'good' or 'day'. Both words' more unique (and thus predictively-useful) senses are more associated with other words. The word 'good' alone can appear everywhere, so has weak relationships everywhere, but still a strong relationship to other synonyms/antonyms on an evaluative ("good or bad", "likable or unlikable", "preferred or disliked", etc) scale.
All those random/non-predictive instances tend to cancel-out as noise; the relationships that have some ability to predict nearby words, even slightly, eventually find some relative/nearby arrangement in the high-dimensional space, so as to help the model for some training examples.
Note that a word2vec model isn't necessarily an effective way to predict nearby words. It might never be good at that task. But the attempt to become good at neighboring-word prediction, with fewer free parameters than would allow a perfect-lookup against training data, forces the model to reflect underlying semantic or syntactic patterns in the data.
(Note also that some research shows that a larger window influences word-vectors to reflect more topical/domain similarity – "these words are used about the same things, in the broad discourse about X" – while a tiny window makes the word-vectors reflect a more syntactic/typical similarity - "these words are drop-in replacements for each other, fitting the same role in a sentence". See for example Levy/Goldberg "Dependency-Based Word Embeddings", around its Table 1.)
‘Embedding’ mean a semantic vector representation. e.g. how to represent words such that synonyms are nearer than antonyms or other unrelated words.
Embeddings algorithms like Word2vec maps entities be it e-commerce
items or words (say in English language), to N-dimensional vectors.
Now since you have a mathematical representation of the entities in
a Euclidean space, you can use associated semantics such as distance
between vectors. e.g:
For a given item say ‘Levis Jeans’ recommend the most related items
which are often co-purchased with it.
This can be easily done: search the nearest vectors to the vector of
‘Levis Jeans’, and recommend them. You will find that the nearest
vectors correspond to items such as T-shirts etc., which are
relevant to the Levis Jeans. Similarly it preserves
distance/similarity between words e.g.: King - Queen = Man - Woman !
Yes, Word2vec captures such co-occurrance relationships, when
mapping the items/words to vectors also called as ‘item/word
embeddings’.
This is not specifically targeted to sentence embeddings but nevertheless here you get some crucial insights extremely relevant to the core logic behind embedding generation. Read till the end.

doc2vec: Pull documents from inferred document

i am new in word/paragraph embedding and trying to understand via doc2vec in GENSIM. I would like to seek advice on whether my understanding is incorrect. My understanding is that doc2vec is potentially able to return documents that may have semantically similar content. As a test, i tried the following and have the following questions.
Question 1: I noted that every run of training with the exact same parameters and examples will result in a model that produces very different results from previous trains (E.g. Different vectors and different ranking of similar documents eveytime).. Why is this so indeterministic? As such, can this be reliably used for any practical work?
Question 2: Why am i not getting the tag ids of the top similar documents instead?
Results: [('day',0.477),('2016',0.386)....
Question 2 answer: The problem was due to model.most_similar, should use model.docvecs.most_similar instead
Please advise if i misunderstood anything?
Data prep
I had created multiple documents with a sentence each. I had deliberately made it such that they are distinctly different semantically.
A: It is a fine summer weather, with the birds singing and sun shining bright.
B: It is a lovely day indeed, if only i had a degree in appreciating.
C: 2016-2017 Degree in Earth Science Earthly University
D: 2009-2010 Dip in Life and Nature Life College
Query: Degree in Philosophy from Thinking University from 2009 to 2010
Training
I trained the documents (tokens as words, running index as tag)
tdlist=[]
docstring=['It is a fine summer weather, with the birds singing and sun shining bright.',
'It is a lovely day indeed, if only i had a degree in appreciating.',
'2016-2017 Degree in Earth Science Earthly University',
'2009-2010 Dip in Life and Nature Life College']
counter=1
for para in docstring:
tokens=tokenize(para) #This will also strip punctuation
td=TaggedDocument(gensim.utils.to_unicode(str.encode(' '.join(tokens))).split(), str(counter))
tdlist.append(td)
counter=counter+1
model=gensim.models.Doc2Vec(tdlist,dm=0,alpha=0.025, size=20, min_alpha=0.025, min_count=0)
for epoch in range(200):
model.train(tdlist, total_examples=model.corpus_count, epochs=model.iter)
Inference
I then attempted to infer the query. Although they are many missing words in the vocab for the query, i would expect closest document similarity results for C and D. But the results only gave me a list of 'words' followed by a similarity score. I am unsure if my understanding is wrong. Below is my code extract.
mydocvector=model.infer_vector(['Degree' ,'in' ,'Philosophy' ,'from' ,'Thinking' ,'University', 'from', '2009', 'to', '2010'])
print(model.docvecs.most_similar(positive=[mydocvector])
Doc2Vec doesn't work well on toy-sized datasets - few documents, few total words, few words per document. You'll absolutely want more documents than vector dimensions (size), and ideally tens-of-thousands of documents or more.
The second argument to TaggedDocument should be a list of tags. By supplying a single string-of-an-int, each of its elements (characters) will be seen as tags. (With just documents 1 to 4 this won't yet hurt, but as soon as you have document 10, Doc2Vec will see it as tags 1 and 0, unless you supply it as ['10'] (a single-element list).
Yes, to find most-similar documents you use model.docvecs.most_similar() rather than model.most_similar() (which only operates on learned words, if any).
You are using dm=0 mode, which is a pretty good starting idea – it's fast and often a top-performer. But note that this mode doesn't train word-vectors too. So anything you ask for from the top model, like model['summer'] or model.most_similar('sun'), will be nonsense results based on randomly-initialized but never-trained words. (If you need words trained too, either add dbow_words=1 to the dm=0 mode, or use a dm=1 mode. But for pure doc-vectors, dm=0 is a pretty good choice.)
There's no need to call train() in a loop - or indeed at all, given the line above it. The form you've used to instantiate Doc2Vec, with an actual corpus tdlist as the first argument, already triggers model-setup and training, using the default number of iter passes (5) and the supplied alpha and min_alpha. Now, for Doc2Vec training you often want more passes (10 to 20 are common, though smaller datasets might benefit from even more). And for any training, for properly gradient-descent, you want the effective learning-rate alpha to gradually decline to a negligible value, such as the default 0.0001 (rather than a forced same-as-starting value).
The only situation where you'd usually call train() explicitly is if you instantiate the model without a corpus. In that case, you'd need to both call model.build_vocab(tdlist) (to let the model initialize with a discovered vocabulary), and then some form of train() - but you'd still need only one call to train, supplying the desired number of passes. (Allowing the default model.iter 5 passes, inside an outer loop of 200 iterations, means a total of 1000 passes over the data... and all at the same fixed alpha, which is not proper gradient-descent.)
When you have a beefier dataset, you may find results improve with a higher min_count. Usually words that appear only a few times can't contribute much meaning, and thus only serve as noise slowing training and interfering with other vectors becoming more expressive. (Don't assume "more words must equal better results".) Throwing out the singletons, or more, usually helps.
Regarding inference, almost none of the words in your inference text are in the training set. (I only see 'Degree', 'in', and 'University' repeated.) So in addition to all the issues above, inferring a good vector for the example text would be hard. With a richer training set, you'd likely get better results. It also often helps to increase the steps optional parameter to infer_vector() far above its default of 5.

Part of speech tagging : tagging unknown words

In the part of speech tagger, the best probable tags for the given sentence is determined using HMM by
P(T*) = argmax P(Word/Tag)*P(Tag/TagPrev)
T
But when 'Word' did not appear in the training corpus, P(Word/Tag) produces ZERO for given all possible tags, this leaves no room for choosing the best.
I have tried few ways,
1) Assigning small amount of probability for all unknown words, P(UnknownWord/AnyTag)~Epsilon... means this completely ignores the P(Word/Tag) for unknowns word by assigning the constant probability.. So decision making on unknown word is by prior probabilities.. As expected it is not producing good result.
2) Laplace Smoothing
I confused with this. I don't know what is difference between (1) and this. My way of understanding Laplace Smoothing adds the constant probability(lambda) to all unknown & Known words.. So the All Unknown words will get constant probability(fraction of lambda) and Known words probabilities will be the same relatively since all word's prob increased by Lambda.
Is the Laplace Smoothing same as the previous one ?
*)Is there any better way of dealing with unknown words ?
Your two approaches are similar, but, if I understand correctly, they differ in one key way. In (1) you are assigning extra mass to counts of unknown words and in (2) you are assigning extra mass to all counts. You definitely want to do (2) and not (1).
One of the problems with Laplace smoothing is that it give too much of a boost to unknown words and drags down the probabilities of high-probability words too much (relatively speaking). Your version (1) would actually worsen this problem. Basically, it would over-smooth.
Laplace smoothing words ok for an HMM, but it's not great. Most people do add-one smoothing but you could experiment with things like add-one-half or whatever.
If you want to move beyond this naive approach to smoothing, check out "one-count smoothing", as described in the Appendix of Jason Eisner's HMM tutorial. The basic idea here is that for unknown words more probability mass should be given to tags that appear with a wider variety of low frequency words. For example, since the tag NOUN appears on a large number of different words and DETERMINER appears on a small number of different words, it is more likely that an unseen word will be a NOUN.
If you want to get even fancier, you could use a Chinese Restaurant Process model taken from non-parametric Bayesian statistics to put a prior distribution on unseen word/tag combinations. Kevin Knight's Bayesian inference tutorial has details.
I think the HMM-based TnT tagger provides a better approach to handle unknown words (see the approach in TnT tagger's paper).
The accuracy results (for known words and unknown words) of TnT and other two POS and morphological taggers on 13 languages including Bulgarian, Czech, Dutch, English, French, German, Hindi, Italian, Portuguese, Spanish, Swedish, Thai and Vietnamese, can be found in this article.

Supervised Learning for User Behavior over Time

I want to use machine learning to identify the signature of a user who converts to a subscriber of a website given their behavior over time.
Let's say my website has 6 different features which can be used before subscribing and users can convert to a subscriber at any time.
For a given user I have stats which represent the intensity on a continuous range of that user's interaction with features 1-6 on a daily basis so:
D1: f1,f2,f3,f4,f5,f6
D2: f1,f2,f3,f4,f5,f6
D3: f1,f2,f3,f4,f5,f6
D4: f1,f2,f3,f4,f5,f6
Let's say on day 5, the user converts.
What machine using algorithms would help me identify which are the most common patterns in feature usage which lead to a conversion?
(I know this is a super basic classification question, but I couldn't find a good example using longitudinal data, where input vectors are ordered by time like I have)
To develop the problem further, let's assume that each feature has 3 intensities at which the user can interact (H, M, L).
We can then represent each user as a string of states of interaction intensity. So, for a user:
LLLLMM LLMMHH LLHHHH
Would mean on day one they only interacted significantly with features 5 and 6, but by the third day they were interacting highly with features 3 through 6.
N-gram Style
I could make these states words and the lifetime of a user a sentence. (Would probably need to add a "conversion" word to the vocabulary as well)
If I ran these "sentences" through an n-gram model, I could get the likely future state of a user given his/her past few state which is somewhat interesting. But, what I really want to know the most common sets of n-grams that lead to the conversion word. Rather than feeding in an n-gram and getting the next predicted word, I want to give the predicted word and get back the 10 most common n-grams (from my data) which would be likely to lead to the word.
Amaç Herdağdelen suggests identifying n-grams to practical n and then counting how many n-gram states each user has. Then correlating with conversion data (I guess no conversion word in this example). My concern is that there would be too many n-grams to make this method practical. (if each state has 729 possibilities, and we're using trigrams, thats a lot of possible trigrams!)
Alternatively, could I just go thru the data logging the n-grams which led to the conversion word and then run some type of clustering on them to see what the common paths are to a conversion?
Survival Style
Suggested by Iterator, I understand the analogy to a survival problem, but the literature here seems to focus on predicting time to death as opposed to the common sequence of events which leads to death. Further, when looking up the Cox Proportional Hazard model, I found that it does not event accommodate variables which change over time (its good for differentiating between static attributes like gender and ethnicity)- so it seems very much geared toward a different question than mine.
Decision Tree Style
This seems promising though I can't completely wrap my mind around how to structure the data. Since the data is not flat, is the tree modeling the chance of moving from one state to another down the line and when it leads to conversion or not? This is very different than the decision tree data literature I've been able to find.
Also, need clarity on how to identify patterns which lead to conversion instead a models predicts likely hood of conversion after a given sequence.
Theoretically, hidden markov models may be a suitable solution to your problem. The features on your site would constitute the alphabet, and you can use the sequence of interactions as positive or negative instances depending on whether a user finally subscribed or not. I don't have a guess about what the number of hidden states should be, but finding a suitable value for that parameter is part of the problem, after all.
As a side note, positive instances are trivial to identify, but the fact that a user has not subscribed so far doesn't necessarily mean s/he won't. You might consider to limit your data to sufficiently old users.
I would also consider converting the data to fixed-length vectors and apply conceptually simpler models that could give you some intuition about what's going on. You could use n-grams (consecutive interaction sequences of length n).
As an example, assuming that the interaction sequence of a given user ise "f1,f3,f5", "f1,f3,f5" would constitute a 3-gram (trigram). Similarly, for the same user and the same interaction sequence you would have "f1,f3" and "f3,f5" as the 2-grams (bigrams). In order to represent each user as a vector, you would identify all n-grams up to a practical n, and count how many times the user employed a given n-gram. Each column in the vector would represent the number of times a given n-gram is observed for a given user.
Then -- probably with the help of some suitable normalization techniques such as pointwise mutual information or tf-idf -- you could look at the correlation between the n-grams and the final outcome to get a sense of what's going on, carry out feature selection to find the most prominent sequences that users are involved in, or apply classification methods such as nearest neighbor, support machine or naive Bayes to build a predictive model.
This is rather like a survival analysis problem: over time the user will convert or will may drop out of the population, or will continue to appear in the data and not (yet) fall into neither camp. For that, you may find the Cox proportional hazards model useful.
If you wish to pursue things from a different angle, namely one more from the graphical models perspective, then a Kalman Filter may be more appealing. It is a generalization of HMMs, suggested by #AmaçHerdağdelen, which work for continuous spaces.
For ease of implementation, I'd recommend the survival approach. It is the easiest to analyze, describe, and improve. After you have a firm handle on the data, feel free to drop in other methods.
Other than Markov chains, I would suggest decision trees or Bayesian networks. Both of these would give you a likely hood of a user converting after a sequence.
I forgot to mention this earlier. You may also want to take a look at the Google PageRank algorithm. It would help you account for the user completely disappearing [not subscribing]. The results of that would help you to encourage certain features to be used. [Because they're more likely to give you a sale]
I think Ngramm is most promising approach, because all sequnce in data mining are treated as elements depndent on few basic steps(HMM, CRF, ACRF, Markov Fields) So I will try to use classifier based on 1-grams and 2 -grams.

Resources