Hacker News Reader

Comment by SkyMarshal | Hacker News Reader

Interesting, I've been working on a similar idea but different implementation. Out of curiosity, how do you select the summary content for a URL? That can be a problem.

For example, I tested tldr.it on a recent HN submission, Programmers: How to Make the Systems Guys Love You (http://news.ycombinator.com/item?id=1800839), and it seems to select the first paragraph or so of the article (depending on whether you select the short, medium, or long tldr).

However, so many authors these days start out with a long, rambling preamble, instead of the venerable Who/What/When/Where/(Why|How), that the first few paragraphs don't necessarily provide a real tldr-type summary.

Which is the case with that particular article - the first few paragraphs don't help me decide whether to read it or not, and the entire article is quite long, with some good nuggets further down in it.

View on HN

I extract it using a pretty clever algorithm then run it through a few things to summarize it. The extraction is nowhere near perfect; it performs best on the major news sites thus far. I didn't really have time to polish it as much as I'd like, but it seems to work well especially on FOXNews and NYTimes (and Blogspot articles).

Do you mind summarizing a bit how you generate snippets?

We were discussing this on MetaOptimize recently (http://metaoptimize.com/qa/questions/2815/how-are-search-eng...), but I'm curious to hear about alternate approaches.

I build a summariser for classifier4j (http://classifier4j.cvs.sourceforge.net/viewvc/classifier4j/...).

It's 5 years old now, but the summaries it generates are competitive quality-wise with most things out there (eg, the MS Word summarizer). Unfortunately I don't have an online demo working atm (like I said - 5 years old)

My algorithm is something I made up, and from memory it works like this:

1) Remove HTML, stem, remove stopwords etc

2) Sort unique words by popularity in the text

3) Split the original text on sentence boundaries.

4) Include each sentence that first mentions the next most popular word, until the summary is the maximum length requested.

Like most things, it's surprising how well a simple algorithm like that works.

There are ports for C#, and Googling just then apparently someone has done a python port too.

The first couple of articles I tried:

http://www.sfgate.com/cgi-bin/article.cgi?f=/c/a/2010/10/18/...

http://bstj.bell-labs.com/

suffered from complete content extraction failure.

The third one seemed to work pretty well, but there were lots of spurious newlines in the output which made it really hard to read.

Nice idea but needs another 48 hours of polish :-)

PS I tried both those sfgate and Bell Systems pages through viewtext.org, and that got the content for both just fine.

Maybe you could pipe requests through their API:

http://viewtext.org/help/api

Cool, thanks for posting, will definitely keep tabs. Good luck with it. Will post mine when i've got a working site.

Reply on news.ycombinator.com