english

Cormac Herley, Microsoft Research: "So Long, And No Thanks for the Externalities: The Rational Rejection of Security Advice by Users" (PDF)

via Greg Linden

Language Model Generated Injection Attacks: Cool/Disturbing LingPipe Application

read: Shellcode that looks like English.

Amazing.

When SMART tells you about possibly defective hard disk blocks, you might want to know what volumes are affected so you can perform extra precautions or rewrite the data to the the disk to reallocate the affected blocks.

The excellent Bad block HOWTO for smartmontools tells you how to do this, but is too easy to get confused by all the different block numbers, sizes and offsets you have to calculate.

Therefore, I have written a small script that performs the calculations and determines the LVM2 volume that contains a given block number: badblocklocate.py.

Can determine LVM2 logical volume names from block numbers given on the command line, or can call smartctl to automatically determine defective blocks. Please see the comments at the top of the file for usage information.

For now, it works for LVM2 logical volumes only (though classical partitions can be determined through the error message.) If time permits, i might extend it that it also determines which file contains the defective blocks.

Domain Adaptation (i.e.: you train a statistical classifier on one type of text but want to use it on a different type of text) is one of my research interests.

Recently I found a great survey of publications on domain adaptation:
A Literature Survey on Domain Adaptation of Statistical Classifiers,
by Jing Jiang, who is a PhD candidate at UIUC and has written some interesting papers on domain adaptation herself, such as
Instance weighting for domain adaptation in NLP.

Machine Learning researcher John Langford about "The Privacy Problem" - the need to collect vast amounts of data for machine learning and data mining and the privacy issues that come with such big data collections.

(I also think it is not just the data collection that poses privacy problems, it is also machine learning and data mining techniques that enable us to gather information about people that would otherwise be hidden in the sea of data.)

Peter Norvig: Warning Signs in Experimental Design and Interpretation

Mistakes one can make in conductung and interpretingstatistical experiments and how you can spot them.

Also comes with two very interesting graphical examples about how "humans are very good at detecting patterns, but rather poor at detecting randomness".

Intelligent email clients:

Dredze et. al.: Feature Design for Transfer Learning
(learning to recognize mails that need a reply)


Neustaedter et.al.: The Social Network and Relationship Finder:
Social Sorting for Email Triage

Boone: Concept Features in Re:Agent, an Intelligent Email Agent

(there's a lot more, these are just the ones I stumbled upon.)

I wonder why machine learning features don't play any role in real-world email clients apart from spam classification.

Fernando Pereira: "Zellig Harris, natural language processing, and search"
(about the differences between general language and technical languages and their implications for NLP)

Bill Softky: "How Google translates without understanding"
(elReg article about Google's effort in Statistical Machine Translation)

"Computational linguistics in the lat 20 years essentially has been the Theory of the Wall Street Journal"
Ron Kaplan (CTO Powerset), in a talk at IMS Stuttgart this afternoon.

The application I'm working on is of course supposed to be localized to all kinds of languages (european ones first, but with a certain sinophile coworker, anything is possible). So, I was looking at internationalization/localization support for PHP, and arrived at two contenders:

• the gettext PHP extension (Tutorial)


• PEAR::Translation2

The Stanford NLP Group released a Named Entity Recognition software, based on Conditional Random Fields and implemented in Java.

It is pretty fast and also acheives quite good performance with the included models.

For integration into IBM's UIMA text analysis frameword, I have written an Analysis Engine component that wraps the Stanford NE Regonizer.
You can download it here: stanford-ner-uima.zip
Just like the Recognizer itself it is licensed under the GPL.

Please let me know if it is useful for you.

I wish I had known this a few hours ago... Eclipse Workspace Re-Builder Plug-in

Use pdftk: pdftk file1.pdf file2.pdf cat output outfile.pdf

Etzioni (1996!): "I view the World Wide Web as an information food chain. The maze of pages and hyperlinks that comprise the Web are at the very bottom of the chain. The WebCrawlers and Alta Vistas of the world are information herbivores; they graze on Web pages and regurgitate them as searchable indices. Today, most Web users feed near the bottom of the information food chain, but the time is ripe to move up. Since 1991, we have been building information carnivores, which intelligently hunt and feast in Unix, on the Internet, and on the Web".

Nice metaphor.
(Save for the "bots" and "agents" rhethoric that follows.)

Etzioni is building "information omnivores" now.

If you experience slow out-of-cache responses with the stock BIND (9.2.3) nameserver on SuSE Linux Enterprise Server (SLES9), try to disable to ipv6 kernel module. This seems to be related to a similar issue in CentOS