All,
I have been running Y!LDA (https://github.com/shravanmn/Yahoo_LDA) on a set of documents and the results look great (or at least what I would expect). Now I want to use the resulting topics to perform a reverse query against the corpus. Does anyone know if the 3 human readable text files that are generated after the learntopics executable is run is the final output for this library? If so, is that what I need to parse to perform my queries? I am stuck with a little shoulder shrugging at this point...
Thanks,
Adam
If LDA is working the way I think it is (I use a java implementation, so explanations may vary) then what you get out are the three following things:
P(word,concept) -- The probability of getting a word given a concept. So, when LDA finishes figuring out what concepts exist within the corpus, this P(w,c) will tell you (in theory) which words map to which concepts.
A very naive method of determining concepts would be to load this file into a matrix and combine all these probabilities for all possible concepts for a test document in some method (add, multiply, Root-mean-squared) and rank order the concepts.
Do note that the above method does not recognize the various biases introduced by weakly represented topics or dominating topics in LDA. To accommodate that, you need more complicated algorithms (Gibbs sampling, for instance), but this will get you some results.
P(concept,document) -- If you are attempting to find the intrinsic concepts in the documents in the corpus, you would look here. You can use the documents as examples of documents that have a particular concept distribution, and compare your documents to the LDA corpus documents... There are uses for this, but it may not be as useful as the P(w,c).
Something else probably relating to the weights of words, documents, or concepts. This could be as simple as a set of concept examples with beta weights (for the concepts), or some other variables that are output from LDA. These may or may not be important depending on what you are doing. (If you are attempting to add a document to the LDA space, having the alpha or beta values -- very important.)
To answer your 'reverse lookup' question, to determine the concepts of the test document, use P(w,c) for each word w in the test document.
To determine which document is the most like the test document, determine the above concepts, then compare them to the concepts for each document found in P(c,d) (using each concept as a dimension in vector-space and then determining a cosine between the two documents tends to work alright).
To determine the similarity between two documents, same thing as above, just determine the cosine between the two concept-vectors.
Hope that helps.
Related
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.
How are keyword clouds constructed?
I know there are a lot of nlp methods, but I'm not sure how they solve the following problem:
You can have several items that each have a list of keywords relating to them.
(In my own program, these items are articles where I can use nlp methods to detect proper nouns, people, places, and (?) possibly subjects. This will be a very large list given a sufficiently sized article, but I will assume that I can winnow the list down using some method by comparing articles. How to do this properly is what I am confused about).
Each item can have a list of keywords, but how do they pick keywords such that the keywords aren't overly specific or overly general between each item?
For example, trivially "the" can be a keyword that is a lot of items.
While "supercalifragilistic" could only be in one.
I suppose that I could create a heuristic where if a word exists in n% of the items where n is sufficiently small, but will return a nice sublist (say 5% of 1000 articles is 50, which seems reasonable) then I could just use that. However, the issue that I take with this approach is that given two different sets of entirely different items, there is most likely some difference in interrelatedness between the items, and I'm throwing away that information.
This is very unsatisfying.
I feel that given the popularity of keyword clouds there must have been a solution created already. I don't want to use a library however as I want to understand and manipulate the assumptions in the math.
If anyone has any ideas please let me know.
Thanks!
EDIT:
freenode/programming/guardianx has suggested https://en.wikipedia.org/wiki/Tf%E2%80%93idf
tf-idf is ok btw, but the issue is that the weighting needs to be determined apriori. Given that two distinct collections of documents will have a different inherent similarity between documents, assuming an apriori weighting does not feel correct
freenode/programming/anon suggested https://en.wikipedia.org/wiki/Word2vec
I'm not sure I want something that uses a neural net (a little complicated for this problem?), but still considering.
Tf-idf is still a pretty standard method for extracting keywords. You can try a demo of a tf-idf-based keyword extractor (which has the idf vector, as you say apriori determined, estimated from Wikipedia). A popular alternative is the TextRank algorithm based on PageRank that has an off-the-shelf implementation in Gensim.
If you decide for your own implementation, note that all algorithms typically need plenty of tuning and text preprocessing to work correctly.
The minimum you need to do is removing stopwords that you know that they never can be a keyword (prepositions, articles, pronouns, etc.). If you want something fancier, you can use for instance Spacy to keep only desired parts of speech (nouns, verbs, adjectives). You can also include frequent multiword expressions (gensim has good function for automatic collocation detection), named entities (spacy can do it). You can get better results if you run coreference resolution and substitute pronouns with what they refer to... There are endless options for improvements.
I have to manually go through a long list of terms (~3500) which have been entered by users through the years. Beside other things, I want to reduce the list by looking for synonyms, typos and alternate spellings.
My work will be much easier if I can group the list into clusters of possible typos before starting. I was imagining to use some metric which can calculate the similarity to a term, e.g. in percent, and then cluster everything which has a similarity higher than some threshold. As I am going through it manually anyway, I don't mind a high failure rate, if it can keep the whole thing simple.
Ideally, there exists some easily available library to do this for me, implemented by people who know what they are doing. If there is no such, then at least one calculating a similarity metric for a pair of strings would be great, I can manage the clustering myself.
If this is not available either, do you know of a good algorithm which is simple to implement? I was first thinking a Hamming distance divided by word length will be a good metric, but noticed that while it will catch swapped letters, it won't handle deletions and insertions well (ptgs-1 will be caught as very similar to ptgs/1, but hematopoiesis won't be caught as very similar to haematopoiesis).
As for the requirements on the library/algorithm: it has to rely completely on spelling. I know that the usual NLP libraries don't work this way, but
there is no full text available for it to consider context.
it can't use a dictionary corpus of words, because the terms are far outside of any everyday language, frequently abbreviations of highly specialized terms.
Finally, I am most familiar with C# as a programming language, and I already have a C# pseudoscript which does some preliminary cleanup. If there is no one-step solution (feed list in, get grouped list out), I will prefer a library I can call from within a .NET program.
The whole thing should be relatively quick to learn for somebody with almost no previous knowledge in information retrieval. This will save me maybe 5-6 hours of manual work, and I don't want to spend more time than that in setting up an automated solution. OK, maybe up to 50% longer if I get the chance to learn something awesome :)
The question: What should I use, a library, or an algorithm? Which ones should I consider? If what I need is a library, how do I recognize one which is capable of delivering results based on spelling alone, as opposed to relying on context or dictionary use?
edit To clarify, I am not looking for actual semantic relatedness the way search or recommendation engines need it. I need to catch typos. So, I am looking for a metric by which mouse and rodent have zero similarity, but mouse and house have a very high similarity. And I am afraid that tools like Lucene use a metric which gets these two examples wrong (for my purposes).
Basically you are looking to cluster terms according to Semantic Relatedness.
One (hard) way to do it is following Markovitch and Gabrilovitch approach.
A quicker way will be consisting of the following steps:
download wikipedia dump and an open source Information Retrieval library such as Lucene (or Lucene.NET).
Index the files.
Search each term in the index - and get a vector - denoting how relevant the term (the query) is for each document. Note that this will be a vector of size |D|, where |D| is the total number of documents in the collection.
Cluster your vectors in any clustering algorithm. Each vector represents one term from your initial list.
If you are interested only in "visual" similarity (words are written similar to each other) then you can settle for levenshtein distance, but it won't be able to give you semantic relatedness of terms.For example, you won't be able to relate between "fall" and "autumn".
As stated in the title, I'm simply looking for algorithms or solutions one might use to take in the twitter firehose (or a portion of it) and
a) identify questions in general
b) for a question, identify questions that could be the same, with some degree of confidence
Thanks!
(A)
I would try to identify questions using machine learning and the Bag of Words model.
Create a labeled set of twits, and label each of them with a binary
flag: question or not question.
Extract the features from the training set. The features are traditionally words, but at least for any time I tried it - using bi-grams significantly improved the results. (3-grams were not helpful for my cases).
Build a classifier from the data. I usually found out SVM gives better performance then other classifiers, but you can use others as well - such as Naive Bayes or KNN (But you will probably need feature selection algorithm for these).
Now you can use your classifier to classify a tweet.1
(B)
This issue is referred in the world of Information-Retrieval as "duplicate detection" or "near-duplicate detection".
You can at least find questions which are very similar to each other using Semantic Interpretation, as described by Markovitch and Gabrilovich in their wonderful article Wikipedia-based Semantic Interpretation for Natural Language Processing. At the very least, it will help you identify if two questions are discussing the same issues (even though not identical).
The idea goes like this:
Use wikipedia to build a vector that represents its semantics, for a term t, the entry vector_t[i] is the tf-idf score of the term i as it co-appeared with the term t. The idea is described in details in the article. Reading the 3-4 first pages are enough to understand it. No need to read it all.2
For each tweet, construct a vector which is a function of the vectors of its terms. Compare between two vectors - and you can identify if two questions are discussing the same issues.
EDIT:
On 2nd thought, the BoW model is not a good fit here, since it ignores the position of terms. However, I believe if you add NLP processing for extracting feature (for examples, for each term, also denote if it is pre-subject or post-subject, and this was determined using NLP procssing), combining with Machine Learning will yield pretty good results.
(1) For evaluation of your classifier, you can use cross-validation, and check the expected accuracy.
(2) I know Evgeny Gabrilovich published the implemented algorithm they created as an open source project, just need to look for it.
I have a question that is somewhat high level, so I'll try to be as specific as possible.
I'm doing a lot of research that involves combining disparate data sets with header information that refers to the same entity, usually a company or a financial security. This record linking usually involves header information in which the name is the only common primary identifier, but where some secondary information is often available (such as city and state, dates of operation, relative size, etc). These matches are usually one-to-many, but may be one-to-one or even many-to-many. I have usually done this matching by hand or with very basic text comparison of cleaned substrings. I have occasionally used a simple matching algorithm like a Levenshtein distance measure, but I never got much out of it, in part because I didn't have a good formal way of applying it.
My guess is that this is a fairly common question and that there must be some formalized processes that have been developed to do this type of thing. I've read a few academic papers on the subject that deal with theoretical appropriateness of given approaches, but I haven't found any good source that walks through a recipe or at least a practical framework.
My question is the following:
Does anyone know of a good source for implementing multi-dimensional fuzzy record matching, like a book or a website or a published article or working paper?
I'd prefer something that had practical examples and a well defined approach.
The approach could be iterative, with human checks for improvement at intermediate stages.
(edit) The linked data is used for statistical analysis. As such, a little bit of noise is OK, but there is a strong preference for fewer "incorrect matches" over fewer "incorrect non-matches".
If they were in Python that would be fantastic, but not necessary.
One last thing, if it matters, is that I don't care much about computational efficiency. I'm not implementing this dynamically and I'm usually dealing with a few thousand records.
One common method that shouldn't be terribly expensive for "a few thousand records" would be cosine similarity. Although most often used for comparing text documents, you can easily modify it to work with any kind of data.
The linked Wikipedia article is pretty sparse on details, but following links and doing a few searches will get you some good info. Potentially an implementation that you can modify to fit your purposes. In fact, take a look at Simple implementation of N-Gram, tf-idf and Cosine similarity in Python
A simpler calculation, and one that might be "good enough" for your purposes would be a Jaccard index. The primary difference is that typically cosine similarity takes into account the number of times a word is used in a document and in the entire set of documents, whereas the Jaccard index only cares that a particular word is in the document. There are other differences, but that one strikes me as the most important.
The problem is that you have an array of distances, at least one for each column, and you want to combine those distances in an optimal way to indicate whether a pair of records are the same thing or not.
This is a problem of classification, there are many ways to do it, but logistic regression is one of simpler methods. To train a classifer, you will need to label some pairs of records as either matches or not.
The dedupe python library helps you do this and other parts of the difficult task of record linkage. The documentation has a pretty good overview of how to approach the problem of record linkage comprehensively.