Word frequency algorithm for natural language proc

问题:

Without getting a degree in information retrieval, I'd like to know if there exists any algorithms for counting the frequency that words occur in a given body of text. The goal is to get a "general feel" of what people are saying over a set of textual comments. Along the lines of Wordle.

What I'd like:

• ignore articles, pronouns, etc ('a', 'an', 'the', 'him', 'them' etc)
• preserve proper nouns
• ignore hyphenation, except for soft kind

Reaching for the stars, these would be peachy:

• handling stemming & plurals (e.g. like, likes, liked, liking match the same result)
• grouping of adjectives (adverbs, etc) with their subjects ("great service" as opposed to "great", "service")

I've attempted some basic stuff using Wordnet but I'm just tweaking things blindly and hoping it works for my specific data. Something more generic would be great.

回答1:

You'll need not one, but several nice algorithms, along the lines of the following.

• ignoring pronouns is done via a stoplist.
• preserving proper nouns? You mean, detecting named entities, like Hoover Dam and saying "it's one word" or compound nouns, like programming language? I'll give you a hint: that's tough one, but there exist libraries for both. Look for NER (Named entitiy recognition) and lexical chunking. OpenNLP is a Java-Toolkit that does both.
• ignoring hyphenation? You mean, like at line breaks? Use regular expressions and verify the resulting word via dictionary lookup.
• handling plurals/stemming: you can look into the Snowball stemmer. It does the trick nicely.
• "grouping" adjectives with their nouns is generally a task of shallow parsing. But if you are looking specifically for qualitative adjectives (good, bad, shitty, amazing...) you may be interested in sentiment analysis. LingPipe does this, and a lot more.

I'm sorry, I know you said you wanted to KISS, but unfortunately, your demands aren't that easy to meet. Nevertheless, there exist tools for all of this, and you should be able to just tie them together and not have to perform any task yourself, if you don't want to. If you want to perform a task yourself, I suggest you look at stemming, it's the easiest of all.

If you go with Java, combine Lucene with the OpenNLP toolkit. You will get very good results, as Lucene already has a stemmer built in and a lot of tutorial. The OpenNLP toolkit on the other hand is poorly documented, but you won't need too much out of it. You might also be interested in NLTK, written in Python.

I would say you drop your last requirement, as it involves shallow parsing and will definetly not impove your results.

Ah, btw. the exact term of that document-term-frequency-thing you were looking for is called tf-idf. It's pretty much the best way to look for document frequency for terms. In order to do it properly, you won't get around using multidimenional vector matrices.

... Yes, I know. After taking a seminar on IR, my respect for Google was even greater. After doing some stuff in IR, my respect for them fell just as quick, though.

回答2:

Welcome to the world of NLP ^_^

All you need is a little basic knowledge and some tools.

There are already tools that will tell you if a word in a sentence is a noun, adjective or verb. They are called part-of-speech taggers. Typically, they take plaintext English as input, and output the word, its base form, and the part-of-speech. Here is the output of a popular UNIX part-of-speech tagger on the first sentence of your post:

$ echo "Without getting a degree in information retrieval, I'd like to know if there exists any algorithms for counting the frequency that words occur in a given body of text." | tree-tagger-english 
# Word  POS     surface form
Without IN  without
getting VVG get
a   DT  a
degree  NN  degree
in  IN  in
information NN  information
retrieval   NN  retrieval
,   ,   ,
I   PP  I
'd  MD  will
like    VV  like
to  TO  to
know    VV  know
if  IN  if
there   EX  there
exists  VVZ exist
any DT  any
algorithms  NNS algorithm
for IN  for
counting    VVG count
the DT  the
frequency   NN  frequency
that    IN/that that
words   NNS word
occur   VVP occur
in  IN  in
a   DT  a
given   VVN give
body    NN  body
of  IN  of
text    NN  text
.   SENT    .

As you can see, it identified "algorithms" as being the plural form (NNS) of "algorithm" and "exists" as being a conjugation (VBZ) of "exist." It also identified "a" and "the" as "determiners (DT)" -- another word for article. As you can see, the POS tagger also tokenized the punctuation.

To do everything but the last point on your list, you just need to run the text through a POS tagger, filter out the categories that don't interest you (determiners, pronouns, etc.) and count the frequencies of the base forms of the words.

Here are some popular POS taggers:

TreeTagger (binary only: Linux, Solaris, OS-X)
GENIA Tagger (C++: compile your self)
Stanford POS Tagger (Java)

To do the last thing on your list, you need more than just word-level information. An easy way to start is by counting sequences of words rather than just words themselves. These are called n-grams. A good place to start is UNIX for Poets. If you are willing to invest in a book on NLP, I would recommend Foundations of Statistical Natural Language Processing.

回答3:

Here is an example of how you might do that in Python, the concepts are similar in any language.

>>> import urllib2, string
>>> devilsdict = urllib2.urlopen('http://www.gutenberg.org/files/972/972.txt').read()
>>> workinglist = devilsdict.split()
>>> cleanlist = [item.strip(string.punctuation) for item in workinglist]
>>> results = {}
>>> skip = {'a':'', 'the':'', 'an':''}
>>> for item in cleanlist:
      if item not in skip:
        try:
          results[item] += 1
        except KeyError:
          results[item] = 1

>>> results
{'': 17, 'writings': 3, 'foul': 1, 'Sugar': 1, 'four': 8, 'Does': 1, "friend's": 1, 'hanging': 4, 'Until': 1, 'marching': 2 ...

The first line just gets libraries that help with parts of the problem, as in the second line, where urllib2 downloads a copy of Ambrose Bierce's "Devil's Dictionary" The next lines make a list of all the words in the text, without punctuation. Then you create a hash table, which in this case is like a list of unique words associated with a number. The for loop goes over each word in the Bierce book, if there is already a record of that word in the table, each new occurrence adds one to the value associated with that word in the table; if the word hasn't appeared yet, it gets added to the table, with a value of 1 (meaning one occurrence.) For the cases you are talking about, you would want to pay much more attention to detail, for example using capitalization to help identify proper nouns only in the middle of sentences, etc., this is very rough but expresses the concept.

To get into the stemming and pluralization stuff, experiment, then look into 3rd party work, I have enjoyed parts of the NLTK, which is an academic open source project, also in python.

回答4:

I wrote a full program to do just this a while back. I can upload a demo later when I get home.

Here is a the code (asp.net/c#): http://naspinski.net/post/Findingcounting-Keywords-out-of-a-Text-Document.aspx

回答5:

The first part of your question doesn't sound so bad. All you basically need to do is read each word from the file (or stream w/e) and place it into a prefix tree and each time you happen upon a word that already exists you increment the value associated with it. Of course you would have an ignore list of everything you'd like left out of your calculations as well.

If you use a prefix tree you ensure that to find any word is going to O(N) where N is the maximum length of a word in your data set. The advantage of a prefix tree in this situation is that if you want to look for plurals and stemming you can check in O(M+1) if that's even possible for the word, where M is the length of the word without stem or plurality (is that a word? hehe). Once you've built your prefix tree I would reanalyze it for the stems and such and condense it down so that the root word is what holds the results.

Upon searching you could have some simple rules in place to have the match return positive in case of the root or stem or what have you.

The second part seems extremely challenging. My naive inclination would be to hold separate results for adjective-subject groupings. Use the same principles as above but just keep it separate.

Another option for the semantic analysis could be modeling each sentence as a tree of subject, verb, etc relationships (Sentence has a subject and verb, subject has a noun and adjective, etc). Once you've broken all of your text up in this way it seems like it might be fairly easy to run through and get a quick count of the different appropriate pairings that occurred.

Just some ramblings, I'm sure there are better ideas, but I love thinking about this stuff.

回答6:

The algorithm you just described it. A program that does it out of the box with a big button saying "Do it"... I don't know.

But let me be constructive. I recommend you this book Programming Collective Intelligence. Chapters 3 and 4 contain very pragmatic examples (really, no complex theories, just examples).

回答7:

U can use the worldnet dictionary to the get the basic information of the question keyword like its past of speech, extract synonym, u can also can do the same for your document to create the index for it. then you can easily match the keyword with index file and rank the document. then summerize it.

回答8:

Everything what you have listed is handled well by spacy.

1. Ignore some words - use stop words
2. Extract subject - use part of speech tagging to identify it (works out of the box). After a sentence is parsed, find "ROOT" - the main verb of the sentence. By navigating the parse tree you will find a noun that relates to this verb. It will be the subject.
3. Ignore hyphenation - their tokenizer handles hyphens in most cases. It can be easily extended to handle more special cases.

If the list of topics is pre-determined and not huge, you may even go further: build a classification model that will predict the topic. Let's say you have 10 subjects. You collect sample sentences or texts. You load them into another product: prodigy. Using it's great interface you quickly assign subjects to the samples. And finally, using the categorized samples you train the spacy model to predict the subject of the texts or sentences.