You definitely need NLP. For example:

EX P ER T S E XC H A NG E

may either mean

EXPERT SEX CHANGE

or

EXPERTS EXCHANGE

depending on context.

This problem is called word segmentation or text segmentation, and it has been extensively studied.

Note that in languages such as Chinese, Japanese, and Korean, all text in a sentence normally appears with no whitespace. Word segmentation is therefore a prerequisite for machine interpretation of CJK text.

Actually this is not an AI problem, this is pure programming question solved by dynamic programming.

If you simply want to connect the tokens into a list of correctly spelled words then you proceed as follows (Top down style dynamic programming):

public class Connector {

    public static void main(String[] args) {

        List<String> tokens = new ArrayList<String>(); // input
        Connector connector = new Connector();

        @SuppressWarnings("unused")
        List<String> ret = connector.connect(tokens);
    }

    Map<Integer, List<String>> cache;
    List<String> tokens;

    public List<String> connect(List<String> tokens) {
        this.cache = new HashMap<Integer, List<String>>();
        this.tokens = tokens;
        List<String> ret = connect(tokens, 0);
        return ret;
    }


    List<String> connect(List<String> tokens, int start) {
        if (cache.containsKey(start)) {
            return cache.get(start);
        } else {
            List<String> results = new ArrayList<String>();
            connectHelper(tokens, start, results);
            cache.put(start, results);
            return results;
        }
    }


    void connectHelper(List<String> tokens, int start, List<String> results) {
        int n = tokens.size();
        for(int i = start + 1; i < n; i++) {
            String part = concat(tokens, start, i);
            if (dictContains(part)) {
                List<String> restParts = connect(tokens, i);
                for(String oneRestPart : restParts) {
                    results.add(part + " " + oneRestPart);
                }
            }
        }
    }

    static private String concat(List<String> tokens, int low, int hi) {
        StringBuilder sb = new StringBuilder();
        for(int i = low; i < hi; i++) {
            sb.append(tokens.get(i));
        }
        return sb.toString();
    }
}

The problem is called ‘language modeling’

It is trivial to connect letters into some words by taking any dictionary (list of valid words) and finding a way that fits. However, in practice the main problem in this task usuall is finding the best words that fit those letters. Fortunately, this is a problem that others have as well – it occurs in, for example, speech recognition and thus is rather well researched and has resources and various ways to try and solve it. In this field, this problem is generally called ‘language modeling’.

Statistical methods

The currently standard approach for this problem is to statistically evaluate the likelihood of each subfragment based on a large number of examples of how the language looks in real world. If we often see fragments like “… to therapist” and rarely see fragments like “… to the rapist” then we may assume that those letters should be merged together.

A good classic textbook descriptions are available in Speech and Language Processing by Dan Jurafsky and Jim Martin (chapter 6 in the 1st edition, chapter 4 in the 2nd edition) and Foundations of Statistical Natural Language Processing by Chris Manning and Hinrich Schütze (chapter 6).

An often used library for doing language modeling is SRILM – http://www.speech.sri.com/projects/srilm/manpages/ and I believe that projects such as python NLTK (http://www.nltk.org/) also have tools for that. In any case, the particular model (weights of word combinations) is different for each language and domain, so this may be something that you have to adapt yourself.