how do news outlets like google news automatically classify and rank documents about emerging topics, like "obama's 2011 budget"?
i've got a pile of articles tagged with baseball data like player names and relevance to the article (thanks, opencalais), and would love to create a google news-style interface that ranks and displays new posts as they come in, especially emerging topics. i suppose that a naive bayes classifier could be trained w/ some static categories, but this doesn't really allow for tracking trends like "this player was just traded to this team, these other players were also involved."
No doubt, Google News may use other tricks (or even a combination thereof), but one relatively cheap trick, computationally, to infer topics from free-text would exploit the NLP notion that a word gets its meaning only when connected to other words.
An algorithm susceptible of discovering new topic categories from multiple documents could be outlined as follow:
POS (part-of-speech) tag the text
We probably want to focus more on nouns and maybe even more so on named entities (such as Obama or New England)
Normalize the text
In particular replace inflected words by their common stem. Maybe even replace some adjectives by a corresponding Named Entity (ex: Parisian ==> Paris, legal ==> law)
Also, remove noise words and noise expressions.
identify some words from a list of manually maintained "current / recurring hot words" (Superbowl, Elections, scandal...)
This can be used in subsequent steps to provide more weight to some N-grams
Enumerate all N-grams found in each documents (where N is 1 to say 4 or 5)
Be sure to count, separately, the number of occurrences of each N-gram within a given document and the number of documents which cite a given N-gram
The most frequently cited N-grams (i.e. the ones cited in the most documents) are probably the Topics.
Identify the existing topics (from a list of known topics)
[optionally] Manually review the new topics
This general recipe can also be altered to leverage other attributes of the documents and the text therein. For example the document origin (say cnn/sports vs. cnn/politics ...) can be used to select domain specific lexicons. Another example the process can more or less heavily emphasize the words/expressions from the document title (or other areas of the text with a particular mark-up).
The main algorithms behind Google News have been published in the academic literature by Google researchers:
Original paper.
Talk: Google News Personalization: Scalable Online Collaborative Filtering
Blog discussion.
Related
I have a database of several thousands of utterances. Each record (utterance) is a text representing a problem description, which a user has submitted to a service desk. Sometimes also the service desk agent's response is included. The language is highly technical, and it contains three types of tokens:
words and phrases in Language 1 (e.g. English)
words and phrases in Language 2 (e.g. French, Norwegian, or Italian)
machine-generated output (e.g. listing of files using unix command ls -la)
These languages are densely mixed. I often see that in one conversation, a sentence in Language 1 is followed by Language 2. So it is impossible to divide the data into two separate sets, corresponding to utterances in two languages.
The task is to find similarities between the records (problem descriptions). The purpose of this exercise is to understand whether some bugs submitted by users are similar to each other.
Q: What is the standard way to proceed in such a situation?
In particular, the problem lies in the fact that the words come from two different corpora (corpuses), while in addition, some technical words (like filenames, OS paths, or application names) will not be found in any.
I don't think there's a "standard way" - just things you could try.
You could look into word-embeddings that are aligned between langauges – so that similar words across multiple languages have similar vectors. Then ways of building a summary vector for a text based on word-vectors (like a simple average of all a text's words' vectors), or pairwise comparisons based on word vectors (like "Word Mover's Distance"), may still work with mixed-language texts (even mixes of languages within one text).
That a single text, presumably about a a single (or closely related) set of issues, has mixed language may be a blessing rather than a curse: some classifiers/embeddings you train from such texts might then be able to learn the cross-language correlations of words with shared topics. But also, you could consider enhancing your texts with extra synthetic auto-translated text, for any monolingual ranges, to ensure downstream embeddings/comparisons get closer to your ideal of language-obliviousness.
Thank you for the suggestions. After several experiments I developed a method which is simple and works pretty well. Rather than using existing corpora, I created my own corpus based on all the utterances available in my multilingual database. Without translating them. The database has 130,000 utterances, including 3,5 million of words (in three languages: English, French and Norwegian) and 150,000 unique words. The phrase similarity based on the meaning space constructed this way works surprisingly well. I have tested this method on production and the results are good. I also see a lot of space for improvement, and will continue to polish it. I also wrote this article An approach to categorize multi-lingual phrases, describing all the steps in more detail. Critics or improvements welcome.
I'm embarking on a project for a non-profit organization to help process and classify 1000's of reports annually from their field workers / contractors the world over. I'm relatively new to NLP and as such wanted to seek the group's guidance on the approach to solve our problem.
I'll highlight the current process, and our challenges and would love your help on the best way to solve our problem.
Current process: Field officers submit reports from locally run projects in the form of best practices. These reports are then processed by a full-time team of curators who (i) ensure they adhere to a best-practice template and (ii) edit the documents to improve language/style/grammar.
Challenge: As the number of field workers increased the volume of reports being generated has grown and our editors are now becoming the bottle-neck.
Solution: We would like to automate the 1st step of our process i.e., checking the document for compliance to the organizational best practice template
Basically, we need to ensure every report has 3 components namely:
1. States its purpose: What topic / problem does this best practice address?
2. Identifies Audience: Who is this for?
3. Highlights Relevance: What can the reader do after reading it?
Here's an example of a good report submission.
"This document introduces techniques for successfully applying best practices across developing countries. This study is intended to help low-income farmers identify a set of best practices for pricing agricultural products in places where there is no price transparency. By implementing these processes, farmers will be able to get better prices for their produce and raise their household incomes."
As of now, our approach has been to use RegEx and check for keywords. i.e., to check for compliance we use the following logic:
1 To check "states purpose" = we do a regex to match 'purpose', 'intent'
2 To check "identifies audience" = we do a regex to match with 'identifies', 'is for'
3 To check "highlights relevance" = we do a regex to match with 'able to', 'allows', 'enables'
The current approach of RegEx seems very primitive and limited so I wanted to ask the community if there is a better way to solving this problem using something like NLTK, CoreNLP.
Thanks in advance.
Interesting problem, i believe its a thorough research problem! In natural language processing, there are few techniques that learn and extract template from text and then can use them as gold annotation to identify whether a document follows the template structure. Researchers used this kind of system for automatic question answering (extract templates from question and then answer them). But in your case its more difficult as you need to learn the structure from a report. In the light of Natural Language Processing, this is more hard to address your problem (no simple NLP task matches with your problem definition) and you may not need any fancy model (complex) to resolve your problem.
You can start by simple document matching and computing a similarity score. If you have large collection of positive examples (well formatted and specified reports), you can construct a dictionary based on tf-idf weights. Then you can check the presence of the dictionary tokens. You can also think of this problem as a binary classification problem. There are good machine learning classifiers such as svm, logistic regression which works good for text data. You can use python and scikit-learn to build programs quickly and they are pretty easy to use. For text pre-processing, you can use NLTK.
Since the reports will be generated by field workers and there are few questions that will be answered by the reports (you mentioned about 3 specific components), i guess simple keyword matching techniques will be a good start for your research. You can gradually move to different directions based on your observations.
This seems like a perfect scenario to apply some machine learning to your process.
First of all, the data annotation problem is covered. This is usually the most annoying problem. Thankfully, you can rely on the curators. The curators can mark the specific sentences that specify: audience, relevance, purpose.
Train some models to identify these types of clauses. If all the classifiers fire for a certain document, it means that the document is properly formatted.
If errors are encountered, make sure to retrain the models with the specific examples.
If you don't provide yourself hints about the format of the document this is an open problem.
What you can do thought, is ask people writing report to conform to some format for the document like having 3 parts each of which have a pre-defined title like so
1. Purpose
Explains the purpose of the document in several paragraph.
2. Topic / Problem
This address the foobar problem also known as lorem ipsum feeling text.
3. Take away
What can the reader do after reading it?
You parse this document from .doc format for instance and extract the three parts. Then you can go through spell checking, grammar and text complexity algorithm. And finally you can extract for instance Named Entities (cf. Named Entity Recognition) and low TF-IDF words.
I've been trying to do something very similar with clinical trials, where most of the data is again written in natural language.
If you do not care about past data, and have control over what the field officers write, maybe you can have them provide these 3 extra fields in their reports, and you would be done.
Otherwise; CoreNLP and OpenNLP, the libraries that I'm most familiar with, have some tools that can help you with part of the task. For example; if your Regex pattern matches a word that starts with the prefix "inten", the actual word could be "intention", "intended", "intent", "intentionally" etc., and you wouldn't necessarily know if the word is a verb, a noun, an adjective or an adverb. POS taggers and the parsers in these libraries would be able to tell you the type (POS) of the word and maybe you only care about the verbs that start with "inten", or more strictly, the verbs spoken by the 3rd person singular.
CoreNLP has another tool called OpenIE, which attempts to extract relations in a sentence. For example, given the following sentence
Born in a small town, she took the midnight train going anywhere
CoreNLP can extract the triple
she, took, midnight train
Combined with the POS tagger for example; you would also know that "she" is a personal pronoun and "took" is a past tense verb.
These libraries can accomplish many other tasks such as tokenization, sentence splitting, and named entity recognition and it would be up to you to combine all of these tools with your domain knowledge and creativity to come up with a solution that works for your case.
I'm new to natural language process so I apologize if my question is unclear. I have read a book or two on the subject and done general research of various libraries to figure out how i should be doing this, but I'm not confident yet that know what to do.
I'm playing with an idea for an application and part of it is trying to find product mentions in unstructured text (e.g. tweets, facebook posts, emails, websites, etc.) in real-time. I wont go into what the products are but it can be assumed that they are known (stored in a file or database). Some examples:
"starting tomorrow, we have 5 boxes of #hersheys snickers available for $5 each - limit 1 pp" (snickers is the product from the hershey company [mentioned as "#hersheys"])
"Big news: 12-oz. bottles of Coke and Pepsi on sale starting Fri." (coca-cola is the product [aliased as "coke"] from coca-cola company and Pepsi is the product from the PepsiCo company)
"#OMG, i just bought my dream car. a mustang!!!!" (mustang is the product from Ford)
So basically, given a piece of text, query the text to see if it mentions a product and receive some indication (boolean or confidence number) that it does mention the product.
Some concerns I have are:
Missing products because of misspellings. I thought maybe i could use a string similarity check to catch these.
Product names that are also English words or things would get caught. Like mustang the horse versus mustang the car
Needing to keep a list of alternative names for products (e.g. "coke" for "coco-cola", etc.)
I don't really know where to start with this but any help would be appreciated. I've already looked at NLTK and SciKit and didn't really gleam how to do this from there. If you know of examples or papers that explain this, links would be helpful. I'm not specific to any language at this point. Java preferably but Python and Scala are acceptable.
The answer that you chose is not really answering your question.
The best approach you can take is using Named Entity Recognizer(NER) and POS tagger (grab NNP/NNPS; Proper nouns). The database there might be missing some new brands like Lyft (Uber's rival) but without developing your own prop database, Stanford tagger will solve half of your immediate needs.
If you have time, I would build the dictionary that has every brands name and simply extract it from tweet strings.
http://www.namedevelopment.com/brand-names.html
If you know how to crawl, it's not a hard problem to solve.
It looks like your goal is to classify linguistic forms in a given text as references to semantic entities (which can be referred to by many different linguistic forms). You describe a number of subtasks which should be done in order to get good results, but they nevertheless are still independent tasks.
Misspellings
In order to deal with potential misspellings of words, you need to associate these possible misspellings to their canonical (i.e. correct) form.
Phonetic similarity: Many reasons for "misspellings" is opacity in the relationship between the word's phonetic form (i.e. how it sounds) and its orthographic form (i.e. how it's spelled). Therefore, a good way to address this is to index terms phonetically so that e.g. innovashun is associated with innovation.
Form similarity: Additionally, you could do a string similarity check, but you may introduce a lot of noise into your results which you would have to address because many distinct words are in fact very similar (e.g. chic vs. chick). You could make this a bit smarter by first morphologically analyzing the word and then using a tree kernel instead.
Hand-made mappings: You can also simply make a list of common misspelling → canonical_form mappings. This would work well for "exceptions" not handled by the above methods.
Word-sense disambiguation
Mustang the car and Mustang the horse are the same form but refer to entirely different entities (or rather classes of entities, if you want to be pedantic). In fact, we ourselves as humans can't tell which one is meant unless we also know the word's context. One widely-used way of modelling this context is distributional lexical semantics: Defining a word's semantic similarity to another as the similarity of their lexical contexts, i.e. the words preceding and succeeding them in text.
Linguistic aliases (synonyms)
As stated above, any given semantic entity can be referred to in a number of different ways: bathroom, washroom, restroom, toilet, water closet, WC, loo, little boys'/girls' room, throne room etc. For simple meanings referring to generic entities like this, they can often be considered to be variant spellings in the same way that "common misspellings" are and can be mapped to a "canonical" form with a list. For ambiguous references such as throne room, other metrics (such as lexical-distributional methods) can also be included in order to disambiguate the meaning, so that you don't relate e.g. I'm in the throne room just now! to The throne room of the Buckingham Palace is beautiful.
Conclusion
You have a lot of work to do in order to get where you want to go, but it's all interesting stuff and there are already good libraries available for doing most of these tasks.
In an app that i'm creating, I want to add functionality that groups news stories together. I want to group news stories about the same topic from different sources into the same group. For example, an article on XYZ from CNN and MSNBC would be in the same group. I am guessing its some sort of fuzzy logic comparison. How would I go about doing this from a technical standpoint? What are my options? We haven't even started the app yet, so we aren't limited in the technologies we can use.
Thanks, in advance for the help!
This problem breaks down into a few subproblems from a machine learning standpoint.
First, you are going to want to figure out what properties of the news stories you want to group based on. A common technique is to use 'word bags': just a list of the words that appear in the body of the story or in the title. You can do some additional processing such as removing common English "stop words" that provide no meaning, such as "the", "because". You can even do porter stemming to remove redundancies with plural words and word endings such as "-ion". This list of words is the feature vector of each document and will be used to measure similarity. You may have to do some preprocessing to remove html markup.
Second, you have to define a similarity metric: similar stories score high in similarity. Going along with the bag of words approach, two stories are similar if they have similar words in them (I'm being vague here, because there are tons of things you can try, and you'll have to see which works best).
Finally, you can use a classic clustering algorithm, such as k-means clustering, which groups the stories together, based on the similarity metric.
In summary: convert news story into a feature vector -> define a similarity metric based on this feature vector -> unsupervised clustering.
Check out Google scholar, there probably have been some papers on this specific topic in the recent literature. A lot of these things that I just discussed are implemented in natural language processing and machine learning modules for most major languages.
The problem can be broken down to:
How to represent articles (features, usually a bag of words with TF-IDF)
How to calculate similarity between two articles (cosine similarity is the most popular)
How to cluster articles together based on the above
There are two broad groups of clustering algorithms: batch and incremental. Batch is great if you've got all your articles ahead of time. Since you're clustering news, you've probably got your articles coming in incrementally, so you can't cluster them all at once. You'll need an incremental (aka sequential) algorithm, and these tend to be complicated.
You can also try http://www.similetrix.com, a quick Google search popped them up and they claim to offer this service via API.
One approach would be to add tags to the articles when they are listed. One tag would be XYZ. Other tags might describe the article subject.
You can do that in a database. You can have an unlimited number of tags for each article. Then, the "groups" could be identified by one or more tags.
This approach is heavily dependent upon human beings assigning appropriate tags, so that the right articles are returned from the search, but not too many articles. It isn't easy to do really well.
Pretty common situation, I'd wager. You have a blog or news site and you have plenty of articles or blags or whatever you call them, and you want to, at the bottom of each, suggest others that seem to be related.
Let's assume very little metadata about each item. That is, no tags, categories. Treat as one big blob of text, including the title and author name.
How do you go about finding the possibly related documents?
I'm rather interested in the actual algorithm, not ready solutions, although I'd be ok with taking a look at something implemented in ruby or python, or relying on mysql or pgsql.
edit: the current answer is pretty good but I'd like to see more. Maybe some really bare example code for a thing or two.
This is a pretty big topic -- in addition to the answers people come up with here, I recommend tracking down the syllabi for a couple of information retrieval classes and checking out the textbooks and papers assigned for them. That said, here's a brief overview from my own grad-school days:
The simplest approach is called a bag of words. Each document is reduced to a sparse vector of {word: wordcount} pairs, and you can throw a NaiveBayes (or some other) classifier at the set of vectors that represents your set of documents, or compute similarity scores between each bag and every other bag (this is called k-nearest-neighbour classification). KNN is fast for lookup, but requires O(n^2) storage for the score matrix; however, for a blog, n isn't very large. For something the size of a large newspaper, KNN rapidly becomes impractical, so an on-the-fly classification algorithm is sometimes better. In that case, you might consider a ranking support vector machine. SVMs are neat because they don't constrain you to linear similarity measures, and are still quite fast.
Stemming is a common preprocessing step for bag-of-words techniques; this involves reducing morphologically related words, such as "cat" and "cats", "Bob" and "Bob's", or "similar" and "similarly", down to their roots before computing the bag of words. There are a bunch of different stemming algorithms out there; the Wikipedia page has links to several implementations.
If bag-of-words similarity isn't good enough, you can abstract it up a layer to bag-of-N-grams similarity, where you create the vector that represents a document based on pairs or triples of words. (You can use 4-tuples or even larger tuples, but in practice this doesn't help much.) This has the disadvantage of producing much larger vectors, and classification will accordingly take more work, but the matches you get will be much closer syntactically. OTOH, you probably don't need this for semantic similarity; it's better for stuff like plagiarism detection. Chunking, or reducing a document down to lightweight parse trees, can also be used (there are classification algorithms for trees), but this is more useful for things like the authorship problem ("given a document of unknown origin, who wrote it?").
Perhaps more useful for your use case is concept mining, which involves mapping words to concepts (using a thesaurus such as WordNet), then classifying documents based on similarity between concepts used. This often ends up being more efficient than word-based similarity classification, since the mapping from words to concepts is reductive, but the preprocessing step can be rather time-consuming.
Finally, there's discourse parsing, which involves parsing documents for their semantic structure; you can run similarity classifiers on discourse trees the same way you can on chunked documents.
These pretty much all involve generating metadata from unstructured text; doing direct comparisons between raw blocks of text is intractable, so people preprocess documents into metadata first.
You should read the book "Programming Collective Intelligence: Building Smart Web 2.0 Applications" (ISBN 0596529325)!
For some method and code: First ask yourself, whether you want to find direct similarities based on word matches, or whether you want to show similar articles that may not directly relate to the current one, but belong to the same cluster of articles.
See Cluster analysis / Partitional clustering.
A very simple (but theoretical and slow) method for finding direct similarities would be:
Preprocess:
Store flat word list per article (do not remove duplicate words).
"Cross join" the articles: count number of words in article A that match same words in article B. You now have a matrix int word_matches[narticles][narticles] (you should not store it like that, similarity of A->B is same as B->A, so a sparse matrix saves almost half the space).
Normalize the word_matches counts to range 0..1! (find max count, then divide any count by this) - you should store floats there, not ints ;)
Find similar articles:
select the X articles with highest matches from word_matches
This is a typical case of Document Classification which is studied in every class of Machine Learning. If you like statistics, mathematics and computer science, I recommend that you have a look at the unsupervised methods like kmeans++, Bayesian methods and LDA. In particular, Bayesian methods are pretty good at what are you looking for, their only problem is being slow (but unless you run a very large site, that shouldn't bother you much).
On a more practical and less theoretical approach, I recommend that you have a look a this and this other great code examples.
A small vector-space-model search engine in Ruby. The basic idea is that two documents are related if they contain the same words. So we count the occurrence of words in each document and then compute the cosine between these vectors (each terms has a fixed index, if it appears there is a 1 at that index, if not a zero). Cosine will be 1.0 if two documents have all terms common, and 0.0 if they have no common terms. You can directly translate that to % values.
terms = Hash.new{|h,k|h[k]=h.size}
docs = DATA.collect { |line|
name = line.match(/^\d+/)
words = line.downcase.scan(/[a-z]+/)
vector = []
words.each { |word| vector[terms[word]] = 1 }
{:name=>name,:vector=>vector}
}
current = docs.first # or any other
docs.sort_by { |doc|
# assume we have defined cosine on arrays
doc[:vector].cosine(current[:vector])
}
related = docs[1..5].collect{|doc|doc[:name]}
puts related
__END__
0 Human machine interface for Lab ABC computer applications
1 A survey of user opinion of computer system response time
2 The EPS user interface management system
3 System and human system engineering testing of EPS
4 Relation of user-perceived response time to error measurement
5 The generation of random, binary, unordered trees
6 The intersection graph of paths in trees
7 Graph minors IV: Widths of trees and well-quasi-ordering
8 Graph minors: A survey
the definition of Array#cosine is left as an exercise to the reader (should deal with nil values and different lengths, but well for that we got Array#zip right?)
BTW, the example documents are taken from the SVD paper by Deerwester etal :)
Some time ago I implemented something similiar. Maybe this idea is now outdated, but I hope it can help.
I ran a ASP 3.0 website for programming common tasks and started from this principle: user have a doubt and will stay on website as long he/she can find interesting content on that subject.
When an user arrived, I started an ASP 3.0 Session object and recorded all user navigation, just like a linked list. At Session.OnEnd event, I take first link, look for next link and incremented a counter column like:
<Article Title="Cookie problem A">
<NextPage Title="Cookie problem B" Count="5" />
<NextPage Title="Cookie problem C" Count="2" />
</Article>
So, to check related articles I just had to list top n NextPage entities, ordered by counter column descending.