For an application that we built, we are using a simple statistical model for word prediction (like Google Autocomplete) to guide search.
It uses a sequence of ngrams gathered from a large corpus of relevant text documents. By considering the previous N-1 words, it suggests the 5 most likely "next words" in descending order of probability, using Katz back-off.
We would like to extend this to predict phrases (multiple words) instead of a single word. However, when we are predicting a phrase, we would prefer not to display its prefixes.
For example, consider the input the cat.
In this case we would like to make predictions like the cat in the hat, but not the cat in & not the cat in the.
Assumptions:
We do not have access to past search statistics
We do not have tagged text data (for instance, we do not know the parts of speech)
What is a typical way to make these kinds of multi-word predictions? We've tried multiplicative and additive weighting of longer phrases, but our weights are arbitrary and overfit to our tests.
For this question, you need to define what it is you consider to be a valid completion -- then it should be possible to come up with a solution.
In the example you've given, "the cat in the hat" is much better than "the cat in the". I could interpret this as, "it should end with a noun" or "it shouldn't end with overly common words".
You've restricted the use of "tagged text data" but you could use a pretrained model, (e.g. NLTK, spacy, StanfordNLP) to guess the parts of speech and make an attempt to restrict predictions to only complete noun-phrases (or sequence ending in noun). Note that you would not necessarily need to tag all documents fed into the model, but only those phrases you're keeping in your autocomplete db.
Alternately, you could avoid completions that end in stopwords (or very high frequency words). Both "in" and "the" are words that occur in almost all English documents, so you could experimentally find a frequency cutoff (can't end in a word that occurs in more than 50% of documents) that help you filter. You could also look at phrases -- if the end of the phrase is drastically more common as a shorter phrase, then it doesn't make sense to tag it on, as the user could come up with it on their own.
Ultimately, you could create a labeled set of good and bad instances and try to create a supervised re-ranker based on word features -- both ideas above could be strong features in a supervised model (document frequency = 2, pos tag = 1). This is typically how search engines with data can do it. Note that you don't need search statistics or users for this, just a willingness to label the top-5 completions for a few hundred queries. Building a formal evaluation (that can be run in an automated manner) would probably help when trying to improve the system in the future. Any time you observe a bad completion, you could add it to the database and do a few labels -- over time, a supervised approach would get better.
I am building a document classifier to categorize documents.
So first step is to represent each documents as "features vector" for the training purpose.
After some research, I found that I can use either the Bag of Words approach or N-gram approach to represent a document as a vector.
The text in each document (scanned pdfs and images) is retrieved using an OCR, thus some words contain errors. And I don't have previous knowledge about the language used in these documents (can't use stemming).
So as far as I understand I have to use the n-gram approach. or are there other approaches to represent a document ?
I would also appreciate if someone could link me to an N-Gram guide in order to have a clearer picture and understand how it works.
Thanks in Advance
Use language detection to get document's language (my favorite tool is LanguageIdentifier from Tika project, but many others are available).
Use spell correction (see this question for some details).
Stem words (if you work in Java environment, Lucene is your choice).
Collect all N-grams (see below).
Make instances for classification by extracting n-grams from particular documents.
Build classifier.
N-gram models
N-grams are just sequences of N items. In classification by topic you normally use N-grams of words or their roots (though there are models based on N-grams of chars). Most popular N-grams are unigrams (just word), bigrams (2 serial words) and trigrams (3 serial words). So, from sentence
Hello, my name is Frank
you should get following unigrams:
[hello, my, name, is, frank] (or [hello, I, name, be, frank], if you use roots)
following bigrams:
[hello_my, my_name, name_is, is_frank]
and so on.
At the end your feature vector should have as much positions (dimensions) as there are words in all your text plus 1 for unknown words. Every position in instance vector should somehow reflect number of corresponding words in instance text. This may be number of occurrences, binary feature (1 if word occurs, 0 otherwise), normalized feature or tf-idf (very popular in classification by topic).
Classification process itself is the same as for any other domain.
I'm trying to devise a method that will be able to classify a given number of english words into 2 sets - "rare" and "common" - the reference being to how much they are used in the language.
The number of words I would like to classify is bounded - currently at around 10,000, and include everything from articles, to proper nouns that could be borrowed from other languages (and would thus be classified as "rare"). I've done some frequency analysis from within the corpus, and I have a distribution of these words (ranging from 1 use, to tops about 100).
My intuition for such a system was to use word lists (such as the BNC word frequency corpus, wordnet, internal corpus frequency), and assign weights to its occurrence in one of them.
For instance, a word that has a mid level frequency in the corpus, (say 50), but appears in a word list W - can be regarded as common since its one of the most frequent in the entire language. My question was - whats the best way to create a weighted score for something like this? Should I go discrete or continuous? In either case, what kind of a classification system would work best for this?
Or do you recommend an alternative method?
Thanks!
EDIT:
To answer Vinko's question on the intended use of the classification -
These words are tokenized from a phrase (eg: book title) - and the intent is to figure out a strategy to generate a search query string for the phrase, searching a text corpus. The query string can support multiple parameters such as proximity, etc - so if a word is common, these params can be tweaked.
To answer Igor's question -
(1) how big is your corpus?
Currently, the list is limited to 10k tokens, but this is just a training set. It could go up to a few 100k once I start testing it on the test set.
2) do you have some kind of expected proportion of common/rare words in the corpus?
Hmm, I do not.
Assuming you have a way to evaluate the classification, you can use the "boosting" approach to machine learning. Boosting classifiers use a set of weak classifiers combined to a strong classifier.
Say, you have your corpus and K external wordlists you can use.
Pick N frequency thresholds. For example, you may have 10 thresholds: 0.1%, 0.2%, ..., 1.0%.
For your corpus and each of the external word lists, create N "experts", one expert per threshold per wordlist/corpus, total of N*(K+1) experts. Each expert is a weak classifier, with a very simple rule: if the frequency of the word is higher than its threshold, they consider the word to be "common". Each expert has a weight.
The learning process is as follows: assign the weight 1 to each expert. For each word in your corpus, make the experts vote. Sum their votes: 1 * weight(i) for "common" votes and (-1) * weight(i) for "rare" votes. If the result is positive, mark the word as common.
Now, the overall idea is to evaluate the classification and increase the weight of experts that were right and decrease the weight of the experts that were wrong. Then repeat the process again and again, until your evaluation is good enough.
The specifics of the weight adjustment depends on the way how you evaluate the classification. For example, if you don't have per-word evaluation, you may still evaluate the classification as "too many common" or "too many rare" words. In the first case, promote all the pro-"rare" experts and demote all pro-"common" experts, or vice-versa.
Your distribution is most likely a Pareto distribution (a superset of Zipf's law as mentioned above). I am shocked that the most common word is used only 100 times - this is including "a" and "the" and words like that? You must have a small corpus if that is the same.
Anyways, you will have to choose a cutoff for "rare" and "common". One potential choice is the mean expected number of appearances (see the linked wiki article above to calculate the mean). Because of the "fat tail" of the distribution, a fairly small number of words will have appearances above the mean -- these are the "common". The rest are "rare". This will have the effect that many more words are rare than common. Not sure if that is what you are going for but you can just move the cutoff up and down to get your desired distribution (say, all words with > 50% of expected value are "common").
While this is not an answer to your question, you should know that you are inventing a wheel here.
Information Retrieval experts have devised ways to weight search words according to their frequency. A very popular weight is TF-IDF, which uses a word's frequency in a document and its frequency in a corpus. TF-IDF is also explained here.
An alternative score is the Okapi BM25, which uses similar factors.
See also the Lucene Similarity documentation for how TF-IDF is implemented in a popular search library.
We were set an algorithm problem in class today, as a "if you figure out a solution you don't have to do this subject". SO of course, we all thought we will give it a go.
Basically, we were provided a DB of 100 words and 10 categories. There is no match between either the words or the categories. So its basically a list of 100 words, and 10 categories.
We have to "place" the words into the correct category - that is, we have to "figure out" how to put the words into the correct category. Thus, we must "understand" the word, and then put it in the most appropriate category algorthmically.
i.e. one of the words is "fishing" the category "sport" --> so this would go into this category. There is some overlap between words and categories such that some words could go into more than one category.
If we figure it out, we have to increase the sample size and the person with the "best" matching % wins.
Does anyone have ANY idea how to start something like this? Or any resources ? Preferably in C#?
Even a keyword DB or something might be helpful ? Anyone know of any free ones?
First of all you need sample text to analyze, to get the relationship of words.
A categorization with latent semantic analysis is described in Latent Semantic Analysis approaches to categorization.
A different approach would be naive bayes text categorization. Sample text with the assigned category are needed. In a learning step the program learns the different categories and the likelihood that a word occurs in a text assigned to a category, see bayes spam filtering. I don't know how well that works with single words.
Really poor answer (demonstrates no "understanding") - but as a crazy stab you could hit google (through code) for (for example) "+Fishing +Sport", "+Fishing +Cooking" etc (i.e. cross join each word and category) - and let the google fight win! i.e. the combination with the most "hits" gets chosen...
For example (results first):
weather: fish
sport: ball
weather: hat
fashion: trousers
weather: snowball
weather: tornado
With code (TODO: add threading ;-p):
static void Main() {
string[] words = { "fish", "ball", "hat", "trousers", "snowball","tornado" };
string[] categories = { "sport", "fashion", "weather" };
using(WebClient client = new WebClient()){
foreach(string word in words) {
var bestCategory = categories.OrderByDescending(
cat => Rank(client, word, cat)).First();
Console.WriteLine("{0}: {1}", bestCategory, word);
}
}
}
static int Rank(WebClient client, string word, string category) {
string s = client.DownloadString("http://www.google.com/search?q=%2B" +
Uri.EscapeDataString(word) + "+%2B" +
Uri.EscapeDataString(category));
var match = Regex.Match(s, #"of about \<b\>([0-9,]+)\</b\>");
int rank = match.Success ? int.Parse(match.Groups[1].Value, NumberStyles.Any) : 0;
Debug.WriteLine(string.Format("\t{0} / {1} : {2}", word, category, rank));
return rank;
}
Maybe you are all making this too hard.
Obviously, you need an external reference of some sort to rank the probability that X is in category Y. Is it possible that he's testing your "out of the box" thinking and that YOU could be the external reference? That is, the algorithm is a simple matter of running through each category and each word and asking YOU (or whoever sits at the terminal) whether word X is in the displayed category Y. There are a few simple variations on this theme but they all involve blowing past the Gordian knot by simply cutting it.
Or not...depends on the teacher.
So it seems you have a couple options here, but for the most part I think if you want accurate data you are going to need to use some outside help. Two options that I can think of would be to make use of a dictionary search, or crowd sourcing.
In regards to a dictionary search, you could just go through the database, query it and parse the results to see if one of the category names is displayed on the page. For example, if you search "red" you will find "color" on the page and likewise, searching for "fishing" returns "sport" on the page.
Another, slightly more outside the box option would be to make use of crowd sourcing, consider the following:
Start by more or less randomly assigning name-value pairs.
Output the results.
Load the results up on Amazon Mechanical Turk (AMT) to get feedback from humans on how well the pairs work.
Input the results of the AMT evaluation back into the system along with the random assignments.
If everything was approved, then we are done.
Otherwise, retain the correct hits and process them to see if any pattern can be established, generate a new set of name-value pairs.
Return to step 3.
Granted this would entail some financial outlay, but it might also be one of the simplest and accurate versions of the data you are going get on a fairly easy basis.
You could do a custom algorithm to work specifically on that data, for instance words ending in 'ing' are verbs (present participle) and could be sports.
Create a set of categorization rules like the one above and see how high an accuracy you get.
EDIT:
Steal the wikipedia database (it's free anyway) and get the list of articles under each of your ten categories. Count the occurrences of each of your 100 words in all the articles under each category, and the category with the highest 'keyword density' of that word (e.g. fishing) wins.
This sounds like you could use some sort of Bayesian classification as it is used in spam filtering. But this would still require "external data" in the form of some sort of text base that provides context.
Without that, the problem is impossible to solve. It's not an algorithm problem, it's an AI problem. But even AI (and natural intelligence as well, for that matter) needs some sort of input to learn from.
I suspect that the professor is giving you an impossible problem to make you understand at what different levels you can think about a problem.
The key question here is: who decides what a "correct" classification is? What is this decision based on? How could this decision be reproduced programmatically, and what input data would it need?
I am assuming that the problem allows using external data, because otherwise I cannot conceive of a way to deduce the meaning from words algorithmically.
Maybe something could be done with a thesaurus database, and looking for minimal distances between 'word' words and 'category' words?
Fire this teacher.
The only solution to this problem is to already have the solution to the problem. Ie. you need a table of keywords and categories to build your code that puts keywords into categories.
Unless, as you suggest, you add a system which "understands" english. This is the person sitting in front of the computer, or an expert system.
If you're building an expert system and doesn't even know it, the teacher is not good at giving problems.
Google is forbidden, but they have almost a perfect solution - Google Sets.
Because you need to unterstand the semantics of the words you need external datasources. You could try using WordNet. Or you could maybe try using Wikipedia - find the page for every word (or maybe only for the categories) and look for other words appearing on the page or linked pages.
Yeah I'd go for the wordnet approach.
Check this tutorial on WordNet-based semantic similarity measurement. You can query Wordnet online at princeton.edu (google it) so it should be relatively easy to code a solution for your problem.
Hope this helps,
X.
Interesting problem. What you're looking at is word classification. While you can learn and use traditional information retrieval methods like LSA and categorization based on such - I'm not sure if that is your intent (if it is, then do so by all means! :)
Since you say you can use external data, I would suggest using wordnet and its link between words. For instance, using wordnet,
# S: (n) **fishing**, sportfishing (the act of someone who fishes as a diversion)
* direct hypernym / inherited hypernym / sister term
o S: (n) **outdoor sport, field sport** (a sport that is played outdoors)
+ direct hypernym / inherited hypernym / sister term
# S: (n) **sport**, athletics
(an active diversion requiring physical exertion and competition)
What we see here is a list of relationships between words. The term fishing relates to outdoor sport, which relates to sport.
Now, if you get the drift - it is possible to use this relationship to compute a probability of classifying "fishing" to "sport" - say, based on the linear distance of the word-chain, or number of occurrences, et al. (should be trivial to find resources on how to construct similarity measures using wordnet. when the prof says "not to use google", I assume he means programatically and not as a means to get information to read up on!)
As for C# with wordnet - how about http://opensource.ebswift.com/WordNet.Net/
My first thought would be to leverage external data. Write a program that google-searches each word, and takes the 'category' that appears first/highest in the search results :)
That might be considered cheating, though.
Well, you can't use Google, but you CAN use Yahoo, Ask, Bing, Ding, Dong, Kong...
I would do a few passes. First query the 100 words against 2-3 search engines, grab the first y resulting articles (y being a threshold to experiment with. 5 is a good start I think) and scan the text. In particular I"ll search for the 10 categories. If a category appears more than x time (x again being some threshold you need to experiment with) its a match.
Based on that x threshold (ie how many times a category appears in the text) and how may of the top y pages it appears in you can assign a weigh to a word-category pair.
for better accuracy you can then do another pass with those non-google search engines with the word-category pair (with a AND relationship) and apply the number of resulting pages to the weight of that pair. Them simply assume the word-category pair with highest weight is the right one (assuming you'll even have more than one option). You can also multi assign a word to a multiple category if the weights are close enough (z threshold maybe).
Based on that you can introduce any number of words and any number of categories. And You'll win your challenge.
I also think this method is good to evaluate the weight of potential adwords in advertising. but that's another topic....
Good luck
Harel
Use (either online, or download) WordNet, and find the number of relationships you have to follow between words and each category.
Use an existing categorized large data set such as RCV1 to train your system of choice. You could do worse then to start reading existing research and benchmarks.
Appart from Google there exist other 'encyclopedic" datasets you can build of, some of them hosted as public data sets on Amazon Web Services, such as a complete snapshot of the English language Wikipedia.
Be creative. There is other data out there besides Google.
My attempt would be to use the toolset of CRM114 to provide a way to analyze a big corpus of text. Then you can utilize the matchings from it to give a guess.
My naive approach:
Create a huge text file like this (read the article for inspiration)
For every word, scan the text and whenever you match that word, count the 'categories' that appear in N (maximum, aka radio) positions left and right of it.
The word is likely to belong in the category with the greatest counter.
Scrape delicious.com and search for each word, looking at collective tag counts, etc.
Not much more I can say about that, but delicious is old, huge, incredibly-heavily tagged and contains a wealth of current relevant semantic information to draw from. It would be very easy to build a semantics database this way, using your word list as a basis from scraping.
The knowledge is in the tags.
As you don't need to attend the subject when you solve this 'riddle' it's not supposed to be easy I think.
Nevertheless I would do something like this (told in a very simplistic way)
Build up a Neuronal Network which you give some input (a (e)book, some (e)books)
=> no google needed
this network classifies words (Neural networks are great for 'unsure' classification). I think you may simply know which word belongs to which category because of the occurences in the text. ('fishing' is likely to be mentioned near 'sports').
After some training of the neural network it should "link" you the words to the categories.
You might be able to put use the WordNet database, create some metric to determine how closely linked two words (the word and the category) are and then choose the best category to put the word in.
You could implement a learning algorithm to do this using a monte carlo method and human feedback. Have the system randomly categorize words, then ask you to vote them as "match" or "not match." If it matches, the word is categorized and can be eliminated. If not, the system excludes it from that category in future iterations since it knows it doesn't belong there. This will get very accurate results.
This will work for the 100 word problem fairly easily. For the larger problem, you could combine this with educated guessing to make the process work faster. Here, as many people above have mentioned, you will need external sources. The google method would probably work the best, since google's already done a ton of work on it, but barring that you could, for example, pull data from your facebook account using the facebook apis and try to figure out which words are statistically more likely to appear with previously categorized words.
Either way, though, this cannot be done without some kind of external input that at some point came from a human. Unless you want to be cheeky and, for example, define the categories by some serialized value contained in the ascii text for the name :P
When entering a question, stackoverflow presents you with a list of questions that it thinks likely to cover the same topic. I have seen similar features on other sites or in other programs, too (Help file systems, for example), but I've never programmed something like this myself. Now I'm curious to know what sort of algorithm one would use for that.
The first approach that comes to my mind is splitting the phrase into words and look for phrases containing these words. Before you do that, you probably want to throw away insignificant words (like 'the', 'a', 'does' etc), and then you will want to rank the results.
Hey, wait - let's do that for web pages, and then we can have a ... watchamacallit ... - a "search engine", and then we can sell ads, and then ...
No, seriously, what are the common ways to solve this problem?
One approach is the so called bag-of-words model.
As you guessed, first you count how many times words appear in the text (usually called document in the NLP-lingo). Then you throw out the so called stop words, such as "the", "a", "or" and so on.
You're left with words and word counts. Do this for a while and you get a comprehensive set of words that appear in your documents. You can then create an index for these words:
"aardvark" is 1, "apple" is 2, ..., "z-index" is 70092.
Now you can take your word bags and turn them into vectors. For example, if your document contains two references for aardvarks and nothing else, it would look like this:
[2 0 0 ... 70k zeroes ... 0].
After this you can count the "angle" between the two vectors with a dot product. The smaller the angle, the closer the documents are.
This is a simple version and there other more advanced techniques. May the Wikipedia be with you.
#Hanno you should try the Levenshtein distance algorithm. Given an input string s and a list of of strings t iterate for each string u in t and return the one with the minimum Levenshtein distance.
http://en.wikipedia.org/wiki/Levenshtein_distance
See a Java implementation example in http://www.javalobby.org/java/forums/t15908.html
To augment the bag-of-words idea:
There are a few ways you can also pay some attention to n-grams, strings of two or more words kept in order. You might want to do this because a search for "space complexity" is much more than a search for things with "space" AND "complexity" in them, since the meaning of this phrase is more than the sum of its parts; that is, if you get a result that talks about the complexity of outer space and the universe, this is probably not what the search for "space complexity" really meant.
A key idea from natural language processing here is that of mutual information, which allows you (algorithmically) to judge whether or not a phrase is really a specific phrase (such as "space complexity") or just words which are coincidentally adjacent. Mathematically, the main idea is to ask, probabilistically, if these words appear next to each other more often than you would guess by their frequencies alone. If you see a phrase with a high mutual information score in your search query (or while indexing), you can get better results by trying to keep these words in sequence.
From my (rather small) experience developing full-text search engines: I would look up questions which contain some words from query (in your case, query is your question).
Sure, noise words should be ignored and we might want to check query for 'strong' words like 'ASP.Net' to narrow down search scope.
http://en.wikipedia.org/wiki/Index_(search_engine)#Inverted_indices'>Inverted indexes are commonly used to find questions with words we are interested in.
After finding questions with words from query, we might want to calculate distance between words we are interested in in questions, so question with 'phrases similarity' text ranks higher than question with 'discussing similarity, you hear following phrases...' text.
Here is the bag of words solution with tfidfvectorizer in python 3
#from sklearn.feature_extraction.text import CountVectorizer
from sklearn.feature_extraction.text import TfidfVectorizer
import nltk
nltk.download('stopwords')
s=set(stopwords.words('english'))
train_x_cleaned = []
for i in train_x:
sentence = filter(lambda w: not w in s,i.split(","))
train_x_cleaned.append(' '.join(sentence))
vectorizer = TfidfVectorizer(binary=True)
train_x_vectors = vectorizer.fit_transform(train_x_cleaned)
print(vectorizer.get_feature_names_out())
print(train_x_vectors.toarray())
from sklearn import svm
clf_svm = svm.SVC(kernel='linear')
clf_svm.fit(train_x_vectors, train_y)
test_x = vectorizer.transform(["test phrase 1", "test phrase 2", "test phrase 3"])
print (type(test_x))
clf_svm.predict(test_x)