Tag Archives: text analysis

Text analysis in Lapsang

Lapsang is my personal programming project that will download article titles from an RSS-feed and then recommends which ones are probably the most interesting for the user. Initially it will only use the title for analysis (later versions may use the actual article content, url, poster, etc.). For every title the user must tell if it sounds attractive to read and based on this input the program will learn the interests of the user. As soon as the user rates a title, the scores of the individual words of this title will be adjusted. For attractive titles the word score will increase and for uninteresting titles it will decrease. Based on this word scores the program can give a recommendation for new titles. As soon as a title contains words that the program saw before it will use the word scores to calculate a recommendation.

Not all words in a title will be used for scoring. Stop words like ‘a’, ‘the’, ‘and’, etc. can be ignored as they add no significant value. How about singular and plural, should ‘language’ be considered the same word and thus have the same score as ‘languages’? And how to handle verbs, how should I deal with present and past? For simplicity matters I will assume all titles will be in English as multilingual text analysis is way over my head for now :-) Probably it is sufficient to store only the stem of a word. There are various stemming algorithms available and I found a C# implementation for Porter stemming that I could integrate in my own tokenizer. As this text analysis is getting more and more complex I decided to take a step back and stop my own implementation and have a look at already available open source libraries.

Lucene.NET (a port of the Java search engine) is a mature project and contains various text analyzers. Maybe other parts of this library are useful in my program, too, but first I will focus on the text analysis. I downloaded the binaries from the Lucene.NET website to start experimenting. The library is literally ported from the Java version (easier to maintain for them, they say) so the API feels very Java-ish and not so .NET-ish.
The download contains the core Luce.Net.dll and various additional libraries for more advanced analysis purposes.

I created a temporary project and added references to the Lucene.Net.dll and the Snowball.Net.dll (for language-specific analysis using stemming). The following code shows the results using various analyzers to tokenize a text:

using System;
using System.IO;

using Lucene.Net.Analysis;

namespace TryTextAnalysis
class MainClass
static void Main (string[] args)
	string title = 	"My husband is a programmer; I have no idea what that means.";
	ShowTokens(title, new WhitespaceAnalyzer());
	ShowTokens(title, new SimpleAnalyzer());
	ShowTokens(title, new StopAnalyzer());
	ShowTokens(title, new Lucene.Net.Analysis.Standard.StandardAnalyzer());
	ShowTokens(title, new Lucene.Net.Analysis.Snowball.SnowballAnalyzer("English", StopAnalyzer.ENGLISH_STOP_WORDS));

static void ShowTokens(string text, Analyzer analyzer)
	TokenStream stream = analyzer.TokenStream("text", new StringReader(text));
	while (true)
		Token token = stream.Next();
		if (token == null)
		Console.Write(" [{0}]", token.TermText());

The results when you run this program are as follows:

My husband is a programmer; I have no idea what that means.
 [My] [husband] [is] [a] [programmer;] [I] [have] [no] [idea] [what] [that] [means.]
 [my] [husband] [is] [a] [programmer] [i] [have] [no] [idea] [what] [that] [means]
 [my] [husband] [programmer] [i] [have] [idea] [what] [means]
 [my] [husband] [programmer] [i] [have] [idea] [what] [means]
 [my] [husband] [programm] [i] [have] [idea] [what] [mean]

The SnowBallAnalyzer suits my needs so now I can start coding the word scoring algorithm.