I work at Newssift as the Director of Language Technology. We appreciate any and all feedback on the Newssift beta as we work to refine the tool for our users. I’d like to respond and clarify some of the issues you raise:

- “The sentiment analysis is useless. It’s another example on how applying a sentiment to a full article instead of analyzing individually each comment doesn’t really provide any value to the reader or the decision maker.”

LB: This comment betrays a fundamental misunderstanding of the difference between the two semantic elements in our Sentiment engine – polarity and opinion. It may be the case that we have made this distinction insufficiently clear to users, but it is not the case that the results are “useless”. I think the disconnect is that “polarity”, which is the objective global positive or negative effect of an event, is virtually always applied on the document level whereas “opinion” can have both a document and an individual-level application.

“Many of the tools listed in the article use only two (1. morphology and 2. grammatical) out of four (3. logic and 4. disambiguation) known semantic approaches. Statistical or other rule based approaches are used to approximate the 3rd and 4th methods. This can result in poor resolution generally but on sentiment in particular as is the case with Lexalytics. Consider the sentence “The seats in my lousy Chevy are actually really great”. Lexalytics sees “lousy” and “great” and correctly identifies them as adjectives (e.g. the grammatical approach) but fails to attribute them correctly to the nouns in the sentence (e.g.
a semantic logic approach). The “lousy” adjective is easy since it is right in front of Chevy. It’s that pesky “great” which is as far away and to the right of “seats” as possible. Without a true semantic logic step Lexalytics would only tally up one positive adjective and one negative adjective and score it neutral. But that is not what this consumer is really saying. Overall they think their car is lousy and only one aspect of the car is great. That is not neutral in my book, it is negative.”

LB: This example is antithetical to this writer’s previous point – it was claimed that document-level sentiment does not work yet an example of a sentence was taken out of context to demonstrate that point. As a matter of fact the lexical weight of those elements in the example would very likely be overshadowed by further positive discussions of the “Chevy” in the greater context of an article.

- The categorization seems to work only for economic and financial articles (and not even that well on borderline content).

LB: If the writer is referring to Business Topics “categorization”, this is exactly the point.. Newssift focuses on sources and articles that are relevant to a business audience. Articles that are not relevant to Business Topics, do not get classified into Business Topics. We do not refer to this as either a recall or a source issue.

- The entity recognition has a so-so performance considering that FT for sure owns, and probably manages well, up-to-date lists. It is missing, probably to keep the precision high, many organizations and people. At the same time, just try a search for “GE rating” and look how many not normalized duplicates you see in the list of people. Or it lists the same organization in many different formats, sometimes incorrectly. A search for Microsoft gets you Microsoft Corporation but also Microsoft Windows, Microsoft Visual Studio, Microsoft Photosynth, etc. These are products not organizations.
- They are using a basic html cleaner: people and companies are taken from the peripheral content too and not only (as it should) from the main and more relevant section of the content.

LB: I am unable to duplicate this search experience most likely because of our improvements in annotation and source filtering. Newssift is in beta and our technology team continues to test entity extraction accuracy and our numbers continue to improve. We also consider consumer feedback , from people such as yourself, as we refine and improve Newssift.

Dear Techcrunch, please remove spam.

Maybe you are right but aren’t you curious to take a look? I just need a 30 minutes demo.

Thanks for your message anyway and I really hope you will be successful. We need more and more examples of smarter search and analysis websites.

Here is my first analysis:
- The sentiment analysis is useless. It’s another example on how applying a sentiment to a full article instead of analyzing individually each comment doesn’t really provide any value to the reader or the decision maker.
- Many of the tools listed in the article use only two (1. morphology and 2. grammatical) out of four (3. logic and 4. disambiguation) known semantic approaches. Statistical or other rule based approaches are used to approximate the 3rd and 4th methods. This can result in poor resolution generally but on sentiment in particular as is the case with Lexalytics. Consider the sentence “The seats in my lousy Chevy are actually really great”. Lexalytics sees “lousy” and “great” and correctly identifies them as adjectives (e.g. the grammatical approach) but fails to attribute them correctly to the nouns in the sentence (e.g.
a semantic logic approach). The “lousy” adjective is easy since it is right in front of Chevy. It’s that pesky “great” which is as far away and to the right of “seats” as possible. Without a true semantic logic step Lexalytics would only tally up one positive adjective and one negative adjective and score it neutral. But that is not what this consumer is really saying. Overall they think their car is lousy and only one aspect of the car is great. That is not neutral in my book, it is negative.
- The categorization seems to work only for economic and financial articles (and not even that well on borderline content).
- The entity recognition has a so-so performance considering that FT for sure owns, and probably manages well, up-to-date lists. It is missing, probably to keep the precision high, many organizations and people. At the same time, just try a search for “GE rating” and look how many not normalized duplicates you see in the list of people. Or it lists the same organization in many different formats, sometimes incorrectly. A search for Microsoft gets you Microsoft Corporation but also Microsoft Windows, Microsoft Visual Studio, Microsoft Photosynth, etc. These are products not organizations.
- They are using a basic html cleaner: people and companies are taken from the peripheral content too and not only (as it should) from the main and more relevant section of the content.

I want to challenge here any content provider who wants to replicate what FT did. Please contact us at http://www.expertsystem.net and in 4 months, with 4 people and 1 core technology we can deliver a system that is better performing than FT at a fraction (really) of the cost.

Endeca is able to leverage abundance of metadata, created by Nstein’s text mining to deliver fast and accurate search results.

The semantic features of this are really cool – definitely a lot of possibility here, and it’s executed well.

Now, they just need a logo that doesn’t look like it’s from the mid-90’s

http://health.okguru.com

Make it simple stupid!

I tried searching for an article about the bank of Canada from today, and there is nothing. They need to add more sources to this to make it relevant. From the looks of it they’ve got a few newswire and the financial times (what a shocker there) but nothing of relevancy. Either they expand their reach or i’ll stick with Google and Business.com as they index more than a couple sources.

I found your entry interesting do I’ve added a Trackback to it on my weblog …

A very similar site here.. wotnews.com