I have developed a pipeline to extract text from documents, preprocess the text, and train a gensim Doc2vec model on given documents. Given a document in my corpus, I would like to recommend other documents in the corpus.
I want to know how I can evaluate my model without having a pre-defined list of "good" recommendations. Any ideas?
One simple self-check that can be used to catch some big problems with a Doc2Vec model training pipeline – like gross misparameterizations, or insufficient data/epochs – is to re-infer vectors for the training texts (using .infer_vector()), and check that generally:
the bulk-trained vector for the same text is "close to" the re-inferred vector - such as its nearest-neighbor, or one of the top neighbors, in a .most_similar() operation on the re-inferred text
the overall list of nearest-neighbors (from .most_similar()) for the bulk-trained vector, & the re-inferred vector, are very similar.
They won't necessarily be identical, for reasons explained in Q11 & Q12 of the Gensim Project FAQ, but if they're wildly-different, then something foundational has gone wrong, like:
insufficient (in quantity or quality/form) training data
misparameterizations, like too few epochs or too-large (overfitting-prone) vectors for the quantity of data
Ultimately, though, the variety of data sources & intended uses & possible dimensions of "recommendation-worthy" mean that you need cusomt inputs, based on your project's needs, usually from the intended audience (or your own ability to simulate/represent it).
In the original paper introducing the "Paragraph Vector" algorithm (what's inside the Doc2Vec class), and a followup evaluating it on Wikipedia & arXiv articles, several of the evaluations used triplets of documents, where 2 of the triplet were conjectured to be "necessarily similar" based on some preexisting system's groupings, and the 3rd randomly-chosen.
The algorithm's performance, and relative performance under different parameter choices, was scored based on how often it placed the 2 presumptively-related documents closer-together than the 3rd randomly-chosen document.
For example, one of the original paper's evaluations use brief search-engine-result snippets as documents, and considered any 2 documents that appeared as sibling top-10 results for the same query as presumptively-related. Two of the followup paper's evaluation used the human-curated categories of Wikipedia or arXiv as signalling that articles of the same category should be presumptively-related.
It's imperfect, but allowed the creation of large evaluation sets from already-existing systems/data, which generally pointed results in the same direction as human senses-of-relatedness.
Perhaps you can find a similar preexisting guide for your data. Or, as you perform ad-hoc checking, be sure to capture every judgement you make, so that it becomes, over time, a growing dataset of desirable pairings that are either (a) better than some other result that was co-presented; or (b) just "presumably good enough" that they usually should rank higher than other random 3rd documents. A large amount of imprecision in such desirability-data is tolerable, as it can even out as the set of probe-pairings grows, and the power of being able to automate bulk quantitative evaluations (reusing old assessments against new parameters/models) drives far more overall improvement than any small glitches in the evaluations cost.
Related
I am comparing techniques and want to find out what is the best method to vector and reduce dimensions of a large number of text documents. I have already tested Bag of Words and TF-IDF and reduced dimensions with PCA, SVD, and NMF. Using these approaches I can reduce my data and know the best number of dimensions based on the variance explained.
However, I want to do the same with doc2vec, considering that doc2vec itself is a dimensional reducer, what is the best approach to find out the number of dimensions for my model? Is there any statistical measure that helps me find the best number of vector_size?
Thanks in advance!
There's no magic indicator for what's best; you should try a range of dimensionalities to see what scores well on your specific downstream evaluations, given your data & goals.
If using a doc2vec implementation that offers inference of out-of-training set documents (such as via the .infer_vector() method in Python gensim library), then a plausible sanity check for eliminating very-bad choices of vector_size (or other parameters) is to re-infer vectors for training-set documents.
If repeated re-inferences of the same text are are generally "close to" each other, and to the vector for that same document created by the full model training, that's a weak indicator that the model is at least behaving in a self-consistent way. (If the spread of results is large, that might indicate potential problems with insufficient data, too few training epochs, a too-large/overfit model, or other foundational issues.)
So I am doing a project on document similarity and right now my features are only the embeddings from Doc2Vec. Since that is not showing any good results, after hyperparameter optimization and word embedding before the doc embedding... What other features can I add, so as to get better results?
My dataset is 150 documents, 500-700 words each, with 10 topics(labels), each document having one topic. Documents are labeled on a document level, and that labeling is currently used only for evaluation purposes.
Edit: The following is answer to gojomo's questions and elaborating on my comment on his answer:
The evaluation of the model is done on the training set. I am comparing if the label is the same as the most similar document from the model. For this I am first getting the document vector using the model's method 'infer_vector' and then 'most_similar' to get the most similar document. The current results I am getting are 40-50% of accuracy. A satisfactory score would be of at least 65% and upwards.
Due to the purpose of this research and it's further use case, I am unable to get a larger dataset, that is why I was recommended by a professor, as this is a university project, to add some additional features to the document embeddings of Doc2Vec. As I had no idea what he ment, I am asking the community of stackoverflow.
The end goal of the model is to do clusterization of the documents, again the labels for now being used only for evaluation purposes.
If I don't get good results with this model, I will try out the simpler ones mentioned by #Adnan S #gojomo such as TF-IDF, Word Mover's Distance, Bag of words, just presumed I would get better results using Doc2Vec.
You should try creating TD-IDF with 2 and 3 grams to generate a vector representation for each document. You will have to train the vocabulary on all the 150 documents. Once you have the TF-IDF vector for each document, you can use cosine similarity between any two of them.
Here is a blog article with more details and doc page for sklearn.
How are you evaluating the results as not good, and how will you know when your results are adequate/good?
Note that just 150 docs of 400-700 words each is a tiny, tiny dataset: typical datasets used published Doc2Vec results include tens-of-thousands to millions of documents, of hundreds to thousands of words each.
It will be hard for any of the Word2Vec/Doc2Vec/etc-style algorithms to do much with so little data. (The gensim Doc2Vec implementation includes a similar toy dataset, of 300 docs of 200-300 words each, as part of its unit-testing framework, and to eke out even vaguely useful results, it must up the number of training epochs, and shrink the vector size, significantly.)
So if intending to use Doc2Vec-like algorithms, your top priority should be finding more training data. Even if, in the end, only ~150 docs are significant, collecting more documents that use similar domain language can help improve the model.
It's unclear what you mean when you say there are 10 topics, and 1 topic per document. Are those human-assigned categories, and are those included as part of the training texts or tags passed to the Doc2Vec algorithm? (It might be reasonable to include it, depending on what your end-goals and document-similarity evaluations consist of.)
Are these topics the same as the labelling you also mention, and are you ultimately trying to predict the topics, or just using the topics as a check of the similarity-results?
As #adnan-s suggests in the other answer, it may also be worth trying more-simple count-based 'bag of words' document representations, including potentially on word n-grams or even character n-grams, or TF-IDF weighted.
If you have adequate word-vectors, as trained from your data or from other compatible sources, the "Word Mover's Distance" measure can be another interesting way to compute pairwise similarities. (However, it may be too expensive to calculate between many-hundred-word texts - working much faster on shorter texts.)
As others have already suggested your training set of 150 documents probably isn't big enough to create good representations. You could, however, try to use a pre-trained model and infer the vectors of your documents.
Here is a link where you can download a (1.4GB) DBOW model trained on English Wikipedia pages, working with 300-dimensional document vectors. I obtained the link from jhlau/doc2vec GitHub repository. After downloading the model you can use it as follows:
from gensim.models import Doc2Vec
# load the downloaded model
model_path = "enwiki_dbow/doc2vec.bin"
model = Doc2Vec.load(model_path)
# infer vector for your document
doc_vector = model.infer_vector(doc_words)
Where doc_words is a list of words in your document.
This, however, may not work for you in case your documents are very specific. But you can still give it a try.
I'm trying to find out the similarity between 2 documents. I'm using Doc2vec Gensim to train around 10k documents. There are around 10 string type of tags. Each tag consists of a unique word and contains some sort of documents. Model is trained using distributed memory method.
Doc2Vec(alpha=0.025, min_alpha=0.0001, min_count=2, window=10, dm=1, dm_mean=1, epochs=50, seed=25, vector_size=100, workers=1)
I've tried both dm and dbow as well. dm gives better result(similarity score) as compared to dbow. I understood the concepts of dm vs dbow. But don't know which method is good for similarity measures between two documents.
First question: Which method is the best to perform well on similarities?
model.wv.n_similarity(<words_1>, <words_2>) gives similarity score using word vectors.
model.docvecs.similarity_unseen_docs(model, doc1, doc2) gives similarity score using doc vectors where doc1 and doc2 are not tags/ or indexes of doctags. Each doc1 and doc2 contains 10-20 words kind of sentences.
Both wv.n_similarity and docvecs.similarity_unseen_docs provide different similarity scores on same types of documents.
docvecs.similarity_unseen_docs gives little bit good results as compared to wv.n_similarity but wv.n_similarity sometimes also gives good results.
Question: What is the difference between docvecs.similarity_unseen_docs and wv.n_similarity? Can I use docvecs.similarity_unseen_docs to find the similarity score between unseen data (It might be a silly question)?
Why I asked because docvecs.similarity_unseen_docs provides similarity score on tags, not on actual words belonging to their tags. I'm not sure, please correct me here, if I'm wrong.
How can I convert cosine similarity score to probability?
Thanks.
model = Doc2Vec(alpha=0.025, min_alpha=0.0001, min_count=2, window=10, dm=1, dm_mean=1, epochs=50, seed=25, vector_size=100, workers=4)
# Training of the model
tagged_data = [TaggedDocument(words=_d, tags=[str(i)]) for i, _d in enumerate(<list_of_list_of_tokens>)]
model.build_vocab(tagged_data)
model.train(tagged_data, total_examples=model.corpus_count, epochs=model.epochs)
# Finding similarity score
model.wv.n_similarity(<doc_words1>, <doc_words2>)
model.random.seed(25)
model.docvecs.similarity_unseen_docs(model, <doc_words1>, <doc_words2>)
Both PV-DM mode (dm=1, the default) and PV-DBOW mode (dm=0) can work well. Which is better will depend on your data and goals. Once you have a robust way to quantitatively score the quality of a model's results, for your project goals – which you'll want to be able to tune all of the model's meta-parameters, including DM/DBOW mode – you can and should try both.
PV-DBOW trains fast, and often works very well on shortish-docs (a few dozens of words). Note, though, that this mode doesn't train usable word-vectors unless you also add the dbow_words=1 option, which will slow training.
Using model.wv.n_similarity() relies on word-vectors only. It averages each set f word-vectors, then reports the cosine-similarity between those two averages. (So, it will only be sensible in PV-DM mode, or PV-DBOW with dbow_words=1 activated.
Using model. docvecs.similarity_unseen_docs() uses infer_vector() to treat each of the supplied docs as new texts, for which a true Doc2Vec doc-vector (not a mere average-of-word-vectors) is calculated. (This method operates on lists-of-words, not lists-of-tags.)
Which is better is something you should test for your goals. The average-of-word-vectors is a simpler, faster technique for making a text-vector – but still works ok for a lot of purposes. The inferred doc-vectors take longer to calculate, but with a good model, may be better for some tasks.
Other notes on your setup:
often, setting min_count as low as 2 is a bad idea: those rare words don't have enough examples to mean much, and actually interfere with the quality of surrounding words
10k documents is on the smallish side for a training corpus, compared to published Doc2Vec results (which usually use tens-of-thousands to millions of documents).
published results often use 10-20 training epochs (though more, like your choice of 50, might be helpful especially for smaller corpuses)
on typical multi-core machines workers=1 will be much slower than the default (workers=3); on a machine with 8 or more cores, up to workers=8 is often a good idea. (Though, unless using the newer corpus_file input option, more workers up to the full count of 16, 32, etc cores doesn't help.)
classic Doc2Vec usage doesn't assign docs just known labels (as in your "10 string type of tags"), but unique IDs for each document. In some cases using, or adding, known labels as tags may help, but beware that if you're only supplying 10 tags, you've essentially turned your 10,000 documents into 10 documents (from the perspective of the model's view, which sees all texts with the same tag as if they were segments of one larger document with that tag). In plain PV-DBOW, training only 10 doc-vectors, of 100-dimensions each, from just 10 distinct examples wouldn't make much sense: it'd be prone to severe overfitting. (In PV-DM or PV-DBOW with dbow_words, the fact that the model is training both 10 doc-vectors and many hundreds/thousands of other vocabulary-word word-vectors would help offset the risk of overfitting.)
We've build a system that analyzes some data and outputs some results in plain English (i.e. no charts etc.). The current implementation relies on lots of templates and some randomization in order to give as much diversity to the text as possible.
We'd like to switch to something more advanced with the hope that the produced text is less repetitive and sounds less robotic. I've searched a lot on google but I cannot find something concrete to start from. Any ideas?
EDIT: The data fed to the NLG mechanism are in JSON format. Here is an example about web analytics data. The json file may contain for example a metric (e.g. visits), it's value in the last X days, whether the last value is expected or not and which dimensions (e.g. countries or marketing channels) affected its change.
The current implementation could give something like this:
Overall visits in the UK mainly from ABC email campaign reached 10K (+20% DoD) and were above the expected value by 10%. Users were mainly landing on XXX page while the increase was consistent across devices.
We're looking to finding a way to depend less on templates, sound even more natural and increase the vocabulary.
What you are looking for is a hot research area and a pretty tough task. Currently there is no way to generate 100% meaningful diverse and natural sentences. one approach to generate sentences is using n-grams. using these method you can generate sentences that look more natural and diverse that may look good but probably meaningless and grammatically incorrect.
A more up to date approach is using Deep learning. anyway if you want to generate meaningful sentences, maybe your best way is using your current template based method.
You can find an introduction to basics of n-gram based NLG here:
Generating Random Text with Bigrams
this tool sounds to implement some of the most famous techniques for natural language generation: simplenlg
Have you tried Neural Networks especially LSTM and GRU architectures? These models are the most recent developments in predicting sequences of words. Generating natural language means to generate a sequence of words such that it makes sense with respect to the input and earlier words in the sequence. This is equivalent to predicting time series. LSTM is designed for predicting time series. Hence, it is commonly used to predict a sequence of words, given an input sequence, an input word, or any other input that can be embedded in a vector.
Deep learning libraries such as Tensorflow, Keras, and Torch all have sequence to sequence implementations that can be used for generating natural language by predicting a sequence of words given an input.
Note that usually these models need a huge amount of training data.
You need to meet two criteria in order to benefit from such models:
You should be able to represent your input as a vector.
You need a relatively large amount of input/target pairs.
Background
For years I've been using my own Bayesian-like methods to categorize new items from external sources based on a large and continually updated training dataset.
There are three types of categorization done for each item:
30 categories, where each item must belong to one category, and at most two categories.
10 other categories, where each item is only associated with a category if there is a strong match, and each item can belong to as many categories as match.
4 other categories, where each item must belong to only one category, and if there isn't a strong match the item is assigned to a default category.
Each item consists of English text of around 2,000 characters. In my training dataset there are about 265,000 items, which contain a rough estimate of 10,000,000 features (unique three word phrases).
My homebrew methods have been fairly successful, but definitely have room for improvement. I've read the NLTK book's chapter "Learning to Classify Text", which was great and gave me a good overview of NLP classification techniques. I'd like to be able to experiment with different methods and parameters until I get the best classification results possible for my data.
The Question
What off-the-shelf NLP tools are available that can efficiently classify such a large dataset?
Those I've tried so far:
NLTK
TIMBL
I tried to train them with a dataset that consisted of less than 1% of the available training data: 1,700 items, 375,000 features. For NLTK I used a sparse binary format, and a similarly compact format for TIMBL.
Both seemed to rely on doing everything in memory, and quickly consumed all system memory. I can get them to work with tiny datasets, but nothing large. I suspect that if I tried incrementally adding the training data the same problem would occur either then or when doing the actual classification.
I've looked at Google's Prediction API, which seem to do much of what I'm looking for but not everything. I'd also like to avoid relying on an external service if possible.
About the choice of features: in testing with my homebrew methods over the years, three word phrases produced by far the best results. Although I could reduce the number of features by using words or two word phrases, that would most likely produce inferior results and would still be a large number of features.
After this post and based on the personal experience, I would recommend Vowpal Wabbit. It is said to have one of the fastest text classification algorithms.
MALLET has a number of classifiers (NB, MaxEnt, CRF, etc). It's written Andrew McCallum's group. SVMLib is another good option, but SVM models typically require a bit more tuning than MaxEnt. Alternatively some sort of online clustering like K-means might not be bad in this case.
SVMLib and MALLET are quite fast (C and Java) once you have your model trained. Model training can take a while though! Unfortunately it's not always easy to find example code. I have some examples of how to use MALLET programmatically (along with the Stanford Parser, which is slow and probably overkill for your purposes). NLTK is a great learning tool and is simple enough that is you can prototype what you are doing there, that's ideal.
NLP is more about features and data quality than which machine learning method you use. 3-grams might be good, but how about character n-grams across those? Ie, all the character ngrams in a 3-gram to account for spelling variations/stemming/etc? Named entities might also be useful, or some sort of lexicon.
I would recommend Mahout as it is intended for handling very large scale data sets.
The ML algorithms are built over Apache Hadoop(map/reduce), so scaling is inherent.
Take a look at classification section below and see if it helps.
https://cwiki.apache.org/confluence/display/MAHOUT/Algorithms
Have you tried MALLET?
I can't be sure that it will handle your particular dataset but I've found it to be quite robust in previous tests of mine.
However, I my focus was on topic modeling rather than classification per se.
Also, beware that with many NLP solutions you needn't input the "features" yourself (as the N-grams, i.e. the three-words-phrases and two-word-phrases mentioned in the question) but instead rely on the various NLP functions to produce their own statistical model.