Many supervised machine learning methods rely on the features of each word in the text to predict if this current word present the pertinent information (named entity for example) or not. From these classification algorithms, we can cite CRF, MaxENt, SVM, etc.Some others methods utilize the entire sentence or phrase to learn automatically a pattern that serves to extract the pertinent information.
It depend to your problem and its features. You must define the input and output. and test it practically, by using one of machine learning tools like Weka, Rapidminer or Matlab.
Word-based models are usually more flexible but provides limited discriminative power. For example, two sentences containing one same word do not always have the same meaning. Phrase/sentence-based models are more discriminating but may be sensitive to synonyms. If I have a large-scale corpora and a sufficiently large number of training data, I would like to use a phrase/sentence-based model. If I only have limited training data, I may consider using a word-based model.
I don't think a so strict general rule exist. It depends on the problem (what do you have to predict). Moreover, if the focus is on "results" be prepared to a lot of pre-processing on data whose sense (why pre-processing data in that way performs better with a given model / sampling technique) is not always clear. For instance, this is the approach of the winner of a Kaggle competition hosted by Facebook where you had to predict keywords and tags from millions of text questions
http://www.kaggle.com/c/facebook-recruiting-iii-keyword-extraction/forums/t/6650/share-your-approach/36520#post36520
Based on my experience, the methods you use to extract features or utilize them for both words or phrases are the key factors. For example, IDF of a word may be a feature; however, there must be an assumption on how to find an equivalent feature for phrases ( or maybe n-grams). On the other hand, they have different distributions among documents and collections, and thats a key point to consider, should you be using distributions ? Smoothing ? Will the collection help you get effective results?
As mentioned, it would help digging deeper in defining the problem, then decide which is which.
I have to point out that I am not going to claim a strict general rule, and so I am using terms like "usually", "not always", and "may be". I totally agree with Osama and Gino that the answer to the question depends on a lot of factors including problem formulation, types of documents and/or topics to be classified, etc.
??
https://www.researchgate.net/post/Which_method_can_provide_the_better_results_The_machine_learning_technique_that_utilizes_word-level_features_or_phrase-level_features_dataset2#548b0a85d11b8b8f668b45dc
- Badges
- Science topic