How to interpret doc2vec classifier in terms of words? - gensim

I have trained a doc2vec (PV-DM) model in gensim on documents which fall into a few classes. I am working in a non-linguistic setting where both the number of documents and the number of unique words are small (~100 documents, ~100 words) for practical reasons. Each document has perhaps 10k tokens. My goal is to show that the doc2vec embeddings are more predictive of document class than simpler statistics and to explain which words (or perhaps word sequences, etc.) in each document are indicative of class.
I have good performance of a (cross-validated) classifier trained on the embeddings compared to one compared on the other statistic, but I am still unsure of how to connect the results of the classifier to any features of a given document. Is there a standard way to do this? My first inclination was to simply pass the co-learned word embeddings through the document classifier in order to see which words inhabited which classifier-partitioned regions of the embedding space. The document classes output on word embeddings are very consistent across cross validation splits, which is encouraging, although I don't know how to turn these effective labels into a statement to the effect of "Document X got label Y because of such and such properties of words A, B and C in the document".
Another idea is to look at similarities between word vectors and document vectors. The ordering of similar word vectors is pretty stable across random seeds and hyperparameters, but the output of this sort of labeling does not correspond at all to the output from the previous method.
Thanks for help in advance.
Edit: Here are some clarifying points. The tokens in the "documents" are ordered, and they are measured from a discrete-valued process whose states, I suspect, get their "meaning" from context in the sequence, much like words. There are only a handful of classes, usually between 3 and 5. The documents are given unique tags and the classes are not used for learning the embedding. The embeddings have rather dimension, always < 100, which are learned over many epochs, since I am only worried about overfitting when the classifier is learned, not the embeddings. For now, I'm using a multinomial logistic regressor for classification, but I'm not married to it. On that note, I've also tried using the normalized regressor coefficients as vector in the embedding space to which I can compare words, documents, etc.

That's a very small dataset (100 docs) and vocabulary (100 words) compared to much published work of Doc2Vec, which has usually used tens-of-thousands or millions of distinct documents.
That each doc is thousands of words and you're using PV-DM mode that mixes both doc-to-word and word-to-word contexts for training helps a bit. I'd still expect you might need to use a smaller-than-defualt dimensionaity (vector_size<<100), & more training epochs - but if it does seem to be working for you, great.
You don't mention how many classes you have, nor what classifier algorithm you're using, nor whether known classes are being mixed into the (often unsupervised) Doc2Vec training mode.
If you're only using known classes as the doc-tags, and your "a few" classes is, say, only 3, then to some extent you only have 3 unique "documents", which you're training on in fragments. Using only "a few" unique doctags might be prematurely hiding variety on the data that could be useful to a downstream classifier.
On the other hand, if you're giving each doc a unique ID - the original 'Paragraph Vectors' paper approach, and then you're feeding those to a downstream classifier, that can be OK alone, but may also benefit from adding the known-classes as extra tags, in addition to the per-doc IDs. (And perhaps if you have many classes, those may be OK as the only doc-tags. It can be worth comparing each approach.)
I haven't seen specific work on making Doc2Vec models explainable, other than the observation that when you are using a mode which co-trains both doc- and word- vectors, the doc-vectors & word-vectors have the same sort of useful similarities/neighborhoods/orientations as word-vectors alone tend to have.
You could simply try creating synthetic documents, or tampering with real documents' words via targeted removal/addition of candidate words, or blended mixes of documents with strong/correct classifier predictions, to see how much that changes either (a) their doc-vector, & the nearest other doc-vectors or class-vectors; or (b) the predictions/relative-confidences of any downstream classifier.
(A wishlist feature for Doc2Vec for a while has been to synthesize a pseudo-document from a doc-vector. See this issue for details, including a link to one partial implementation. While the mere ranked list of such words would be nonsense in natural language, it might give doc-vectors a certain "vividness".)
Whn you're not using real natural language, some useful things to keep in mind:
if your 'texts' are really unordered bags-of-tokens, then window may not really be an interesting parameter. Setting it to a very-large number can make sense (to essentially put all words in each others' windows), but may not be practical/appropriate given your large docs. Or, trying PV-DBOW instead - potentially even mixing known-classes & word-tokens in either tags or words.
the default ns_exponent=0.75 is inherited from word2vec & natural-language corpora, & at least one research paper (linked from the class documentation) suggests that for other applications, especially recommender systems, very different values may help.

Related

Determining the "goodness" of a phrase based on "grammatical" or "contextual" relevancy

Given a random string of words, I would like to assign a "goodness" score to the phrase, where "goodness" is some indication of grammatical and contextual relevancy.
For example:
"the green tree was tall" [Good score]
"delicious tires swim open" [Medium score]
"jump an con porch calmly" [Poor score]
I've been experimenting with the Natural Language Toolkit. I'd considered using a trained tagger to assign parts-of-speech to each word in a phrase, and then parse a corpus for occurrences of that POS pattern. This may give me an indication of grammatical "goodness". However, as the tagger itself is trained on the same corpus that I'm using for validation, I can't imagine the results would be reliable. This approach also does not take into consideration the contextual relevancy of the words.
Is anyone aware of existing projects or research into this sort of thing? How would you approach this?
You could employ two different approaches - supervised and semi-supervised.
Supervised
Assuming you have a labeled dataset of tuples of the form <sentence> <goodness label> (like the one in your examples), you could first split your dataset up in a train:test fold (e.g. 4:1).
Then you could simply use BERT feature vectors (these are pre-trained on large volumes of natural language text). The following piece of code gives you the vector for the sentence the green tree was tall (read more here).
nlp_features = pipeline('feature-extraction')
output = nlp_features('the green tree was tall')
np.array(output).shape # (Samples, Tokens, Vector Size)
Assuming you vectorize every sentence, you could then train a simple logistc regression model (sklearn) that learns a set of parameters to minimize the errors in these predictions on the training set and eventually you throw the test set sentences at this model to see how it behaves.
Instead of BERT, you could also use embedded vectors as inputs to an LSTM network for training the classifier (like the one here).
Semi-supervised
This is applicable when you don't have sufficient labeled data (although you need a few to get you started with).
In this case, I think what you could do is to map the words of a sentence into POS tag sequences, e.g.,
the green tree was tall --> ARTICLE ADJ NOUN VERB ADJ (see here for more details).
This step would make your method depend less on the words themselves. A model trained on these sequences would try to discover some latent distinguishing characteristics of good sentences from the bad ones.
In particular, you could run a standard text classification approach with Bidirectional LSTMs for training your classifier (this time not with words but with a much smaller vocabulary of POS tags).
You can use a transformer model from HuggingFace that is fine tuned for sentence correctness. Specifically, the model has to be fine tuned on the Corpus of Linguistic Acceptability (CoLA). Here's a medium article on HuggingFace, transformers, and the fine tuning process.
You can also get a model that's already fine-tuned and you can put in the text classification pipeline for HuggingFace's transformers library here. That site hosts fine-tuned models and you can search for a few others that are fine tuned for the CoLA task there.

How to measure similarity between words or very short text

I work on the problem of finding the nearest document in a list of documents. Each document is a word or a very short sentence (e.g. "jeans" or "machine tool" or "biological tomatoes"). By closest I mean close in a semantical way.
I have tried to use word2vec embeddings (from Mikolov article) but the closest words or more contextually linked than semanticaly linked ("jeans" is linked to "shoes" and not "trousers" as expected).
I have tried to use Bert encoding (https://mccormickml.com/2019/05/14/BERT-word-embeddings-tutorial/#32-understanding-the-output) using last layers but it faces the same issues.
I have tried elastic search, but it doesn't find semantical similarities.
(The task needs to be solved in French but maybe solving it in English is a good first step)
Note different sets of word-vectors may vary in how well they capture your desired 'semantic' similarities. (In particular, training with a shorter window may emphasize similarity among words that are drop-in replacements for each other, as opposed to just used-in-similar domains, as larger window values may emphasize. See this answer for more details.)
You may also want to take a look at "Word Mover's Distance" as a way to compare short texts that contain various mixes of somewhat-similar words. (It's fairly expensive, but should be practical on your short texts. It's available in the Python gensim library as wmdistance() on KeyedVectors instances.)
If you have training data where your specific multi-word phrases are used, in many natural-language-like subtly-varied contexts, you could consider combining all such phrases-of-interest into single tokens (like machine_tool or biological_tomatoes), and training your own domain-specific word-vectors.
For computing similarity between short texts which contains 2 or 3 words, you can use word2vec with getting the average vector of the sentence.
for example, if you have a text (machine tool) and want to represent it in one vector using word2vec so you have to get the vector of "machine" and the vector if "tool" then combine them in one vector by getting the average vector which is to add the two vectors and divide by 2 (the number of words). this will give you a vector representation for a sentence which is more than one word.
You can use also something like doc2vec which is designed on the top of word2vec and its purpose to get a vector for a sentence or paragraph.
You might try document embedding that is built on top of word2vec
However, notice that word and document embedding do not always capture "desired similarity", they just learn a language model on your corpus, they are heavy influenced by text size and word frequency.
How big is your corpus? If you need it just to perform some classification it might be better to train your vectors on a large dataset such as Google News corpus.

Is there a way to load pre-trained word vectors before training the doc2vec model?

I am trying to build a doc2vec model with more or less 10k sentences, after that I will use the model to find the most similar sentence in the model of some new sentences.
I have trained a gensim doc2vec model using the corpus(10k sentences) I have. This model can to some extend tell me if a new sentence is similar to some of the sentences in the corpus.
But, there is a problem: it may happen that there are words in new sentences which don't exist in the corpus, which means that they don't have a word embedding. If this happens, the prediction result will not be good.
As far as I know, the trained doc2vec model does have a matrix of doc vectors as well as a matrix of word vectors. So what I were thinking is to load a set of pre-trained word vectors, which contains a large number of words, and then train the model to get the doc vectors. Does it make sense? Is it possible with gensim? Or is there another way to do it?
Unlike what you might guess, typical Doc2Vec training does not train up word-vectors first, then compose doc-vectors using those word-vectors. Rather, in the modes that use word-vectors, the word-vectors trained in a simultaneous, interleaved fashion alongside the doc-vectors, both changing together. And in one fast and well-performing mode, PV-DBOW (dm=0 in gensim), word-vectors aren't trained or used at all.
So, gensim Doc2Vec doesn't support pre-loading state from elsewhere, and even if it did, it probably wouldn't provide the benefit you expect. (You could dig through the source code & perhaps force it by doing a bunch of initialization steps yourself. But then, if words were in the pre-loaded set, but not in your training data, training the rest of the active words would adjust the entire model in direction incompatible with the imported-but-untrained 'foreign' words. It's only the interleaved, tug-of-war co-training of the model's state which makes the various vectors meaningful in relation to each other.)
The most straightforward and reliable strategy would be to try to expand your training corpus, by finding more documents from a similar/compatible domain, to include multiple varied examples of any words you might encounter later. (If you thought some other word-vectors were apt enough for your domain, perhaps the texts that were used to train those word-vectors can be mixed-into your training corpus. That's a reasonable way to put the word/document data from that other source on equal footing in your model.)
And, as new documents arrive, you can also occasionally re-train the model from scratch, with the now-expanded corpus, letting newer documents contribute equally to the model's vocabulary and modeling strength.

N-gram text categorization category size difference compensation

Lately I've been mucking about with text categorization and language classification based on Cavnar and Trenkle's article "N-Gram-Based Text Categorization" as well as other related sources.
For doing language classification I've found this method to be very reliable and useful. The size of the documents used to generate the N-gram frequency profiles is fairly unimportant as long as they are "long enough" since I'm just using the most common n N-grams from the documents.
On the other hand well-functioning text categorization eludes me. I've tried with both my own implementations of various variations of the algorithms at hand, with and without various tweaks such as idf weighting and other peoples' implementations. It works quite well as long as I can generate somewhat similarly-sized frequency profiles for the category reference documents but the moment they start to differ just a bit too much the whole thing falls apart and the category with the shortest profile ends up getting a disproportionate number of documents assigned to it.
Now, my question is. What is the preferred method of compensating for this effect? It's obviously happening because the algorithm assumes a maximum distance for any given N-gram that equals the length of the category frequency profile but for some reason I just can't wrap my head around how to fix it. One reason I'm interested in this fix is actually because I'm trying to automate the generation of category profiles based on documents with a known category which can vary in length (and even if they are the same length the profiles may end up being different lengths). Is there a "best practice" solution to this?
If you are still interested, and assuming I understand your question correctly, the answer to your problem would be to normalise your n-gram frequencies.
The simplest way to do this, on a per document basis, is to count the total frequency of all n-grams in your document and divide each individual n-gram frequency by that number. The result is that every n-gram frequency weighting now relates to a percentage of the total document content, regardless of the overall length.
Using these percentages in your distance metrics will discount the size of the documents and instead focus on the actual make up of their content.
It might also be worth noting that the n-gram representation only makes up a very small part of an entire categorisation solution. You might also consider using dimensional reduction, different index weighting metrics and obviously different classification algorithms.
See here for an example of n-gram use in text classification
As I know the task is to count probability of generation some text by language model M.
Recently i was working on measuring the readaiblity of texts using semantic, synctatic and lexical properties. It can be also measured by language model approach.
To answer properly you should consider these questions:
Are you using log-likelihood approach?
What levels of N-Grams are you using? unigrams digrams or higher level?
How big are language corpuses that you use?
Using only digrams and unigrams i managed to classify some documents with nice results. If your classification is weak consider creating bigger language corpuse or using n-grams of lower levels.
Also remember that classifying some text to invalid category may be an error depending on length of text (randomly there are few words appearing in another language models).
Just consider making your language corpuses bigger and know that analysing short texts have higher probability of missclasification

how to get the similar texts from a lot of pages?

get the x most similar texts from a lot of texts to one text.
maybe change the page to text is better.
You should not compare the text to every text, because its too slow.
The ability of identifying similar documents/pages, whether web pages or more general forms of text or even of codes, has many practical applications. This topics is well represented in scholarly papers and also in less specialized forums. In spite of this relative wealth of documentation, it can be difficult to find the information and techniques relevant to a particular case.
By describing the specific problem at hand and associated requirements, it may be possible to provide you more guidance. In the meantime the following provides a few general ideas.
Many different functions may be used to measure, in some fashion, the similarity of pages. Selecting one (or possibly several) of these functions depends on various factors, including the amount of time and/or space one can allot the problem and also to the level of tolerance desired for noise.
Some of the simpler metrics are:
length of the longest common sequence of words
number of common words
number of common sequences of words of more than n words
number of common words for the top n most frequent words within each document.
length of the document
Some of the metrics above work better when normalized (for example to avoid favoring long pages which, through their sheer size have more chances of having similar words with other pages)
More complicated and/or computationally expensive measurements are:
Edit distance (which is in fact a generic term as there are many ways to measure the Edit distance. In general, the idea is to measure how many [editing] operations it would take to convert one text to the other.)
Algorithms derived from the Ratcliff/Obershelp algorithm (but counting words rather than letters)
Linear algebra-based measurements
Statistical methods such as Bayesian fitlers
In general, we can distinguish measurements/algorithms where most of the calculation can be done once for each document, followed by a extra pass aimed at comparing or combining these measurements (with relatively little extra computation), as opposed to the algorithms that require to deal with the documents to be compared in pairs.
Before choosing one (or indeed several such measures, along with some weighing coefficients), it is important to consider additional factors, beyond the similarity measurement per-se. for example, it may be beneficial to...
normalize the text in some fashion (in the case of web pages, in particular, similar page contents, or similar paragraphs are made to look less similar because of all the "decorum" associated with the page: headers, footers, advertisement panels, different markup etc.)
exploit markup (ex: giving more weight to similarities found in the title or in tables, than similarities found in plain text.
identify and eliminate domain-related (or even generally known) expressions. For example two completely different documents may appear similar is they have in common two "boiler plate" paragraphs pertaining to some legal disclaimer or some general purpose description, not truly associated with the essence of each cocument's content.
Tokenize texts, remove stop words and arrange in a term vector. Calculate tf-idf. Arrange all vectors in a matrix and calculate distances between them to find similar docs, using for example Jaccard index.
All depends on what you mean by "similar". If you mean "about the same subject", looking for matching N-grams usually works pretty well. For example, just make a map from trigrams to the text that contains them, and put all trigrams from all of your texts into that map. Then when you get your text to be matched, look up all its trigrams in your map and pick the most frequent texts that come back (perhaps with some normalization by length).
I don't know what you mean by similar, but perhaps you ought to load your texts into a search system like Lucene and pose your 'one text' to it as a query. Lucene does pre-index the texts so it can quickly find the most similar ones (by its lights) at query-time, as you asked.
You will have to define a function to measure the "difference" between two pages. I can imagine a variety of such functions, one of which you have to choose for your domain:
Difference of Keyword Sets - You can prune the document of the most common words in the dictionary, and then end up with a list of unique keywords per document. The difference funciton would then calculate the difference as the difference of the sets of keywords per document.
Difference of Text - Calculate each distance based upon the number of edits it takes to turn one doc into another using a text diffing algorithm (see Text Difference Algorithm.
Once you have a difference function, simply calculate the difference of your current doc with every other doc, then return the other doc that is closest.
If you need to do this a lot and you have a lot of documents, then the problem becomes a bit more difficult.

Resources