Ask HN: Review ContextSense - deduces context and sentiment for any URL/text | Hacker News Reader

Ask HN: Review ContextSense - deduces context and sentiment for any URL/text

wingify.com

20 pointsparaschopra25 comments

http://en.wikipedia.org/wiki/The_Holocaust

Slightly Negative (0.33)

Extract From the same link (text only as shown below) Highly Positive 0.81!

>The usual German term for the extermination of the Jews during the Nazi period was the euphemistic phrase Endlösung der Judenfrage (the "Final Solution of the Jewish Question"). In both English and German, "Final Solution" is widely used as an alternative to "Holocaust".[16] For a time after World War II, German historians also used the term Völkermord ("genocide"), or in full, der Völkermord an den Juden ("the genocide of the Jewish people"), while the prevalent term in Germany today is either Holocaust or increasingly Shoah.

NLP is H A R D, I know and defining a range value for sentiment even harder.

Yup, I agree -- though more data provides more oppurtunities to make a better guess.

Hey guys, my startup (Wingify) has exposed an API for determining sentiment and context from any URL or piece of text. I'll be glad if you could review it.

ContextSense was made to demonstrate our contextual targeting capabilities -- and sentiment aspect was added to avoid traditional blunders of displaying ads on pages/news on catastrophies. Best to try ContextSense with text heavy URLs. Also try it with news items (both positive and negative).

I will be happy to provide API access in case any of you is interested in trying it out.

Cool! I tried something like this, it classified blog posts into 100 different moods... it had ~8% accuracy (It also classified as positive or negative, but I don't remember the accuracy, I'll have to dig it up at some point...). Good luck with it.

Do you remember what was it? I will be really curious to have a look. Thanks.

It was accurate around 56% of the time classifying as positive, negative, or neutral. I used a naive Bayesian classifier and guessed the highest rather than showing the percent positive.

I searched for http://school.discoveryeducation.com/schooladventures/slaver...

Few human practices have provoked such deep and widespread outrage as the practice of one human being enslaving another. So why has slavery survived for thousands of years? How did it become so important to civilization? Explore the ways that slavery has been woven into the fabric of societies in America and around the world.

Sentiment: Highly Positive (0.77)

Seems off in cases where the tone of the text is neutral but the subject matter is negative.

The comments so far refer to the "sentiment deducer" aspect of the service. They've brought up some examples of how it's not working right now, but they missed an important question: what in the world is "positive sentiment" abstracted from context, and why would I want to know if a site had it?

An issue we haven't brought up is tags. Current tags for HN: "ago, points, comments, com, hours, hour, hacker, discuss, minutes, tornado." I see some good ones, but it's quite noisy (and it seems to have grabbed the tornado story as a main topic--scraping the title?). This feature seems most useful to me, and most easy to improve.

No, not just a title but the whole website. The algo currently works on TF-IDF of the tags, and HN is full of points, ago, hours, etc. repeated all over it. This increases the TF. We would be putting an upper limit on it though.

I didn't mean the title tag--I meant the first story title (the SEO kids talk about google fixating on the first h1 tag of a site, I though it might be a similar idea). I have no idea what this upper limit change would do, but good luck.

I put this in:

http://en.wikipedia.org/wiki/C*-algebra

The results are better than I would've guessed, but not really so great (this article is pretty degenerate I think, versus more "human-interest" type stories).

One thing I'm curious out: the list of topics is here:

People (7.88) Algebra (4.47) Research Groups (3.41) Software (2.42) Math (2.03) Functional Analysis (1.53) Journals (1.01) Science (0.71) Publications (0.64) Logic and Foundations (0.6)

...does looking at the content give you any sense for why "People" comes in first?

Yes,the false positive rate is a not the least and we are working on decreasing it!

I guess what I'm asking is: is "People" just a default answer?

No, People is not a default answer. It is due to underlying taxonomy data. Working on refining the database.

Reading the other comments, it looks like your biggest problems are false positives. It might not be a bad idea to put together some kind of feedback mechanism - in the event that I disagree with the sentiment/tags/categories that I could provide better ones and you could eventually use that data to refine your algorithm.

That being said, this looks really cool and full of potential, now I just need a reason to be auto-generating that kind of meta data.

I think positive/negative is too shallow. I actually do enjoy this, but I think you need to rate sentiments on other vectors, so if I rate http://teddziuba.com/ I get something like:

Anger: 25 Smugness: 50 Vulgarity: 15 Boredom: 5

etc.

I'm not even sure that's a good example or if it's clear, I just think positive/negative is an oversimplification of the sentiment of most sites.

Whatever you do, Ted Dzubia will break any scale.

Have it recommend changes. Tell us what we can do to improve.

Would be nice if it didn't need http in front of a url.

Can't get it to work. Won't load.

Which browser? Platform? And which URL?

Firefox 3.5.3 on WinXP 64 with the URL that comes up when you load the site. The status spinners just keep spinning forever.

Thanks for the info. Strangely, with exactly same configuration (except WinXP 32 -- which shouldn't make any difference), this is working for me.

What Firefox extensions? Works great here on XP 64.

None that should interfere... Firebug, Site Launcher, Colorzilla.

I typically have cookies disable, but I turned them on and it made no difference.

Reply on news.ycombinator.com