On Jobs and NLP Degrees...

Thomas posted an interesting quandary recently. I'll summarize it this way: How does a person choose which M.S. program in NLP to attend? As far as Thomas and I are aware, there are no rankings for computational linguistics/NLP programs; so, is word of mouth all anyone has to go on? Does anyone out there know of any resources for helping someone like Thomas?

Does the NLP community out there care to contribute words of wisdom to the next generation of CL/NLP newbies?

You may wish to read my own discussion of what I perceive to be the difference between CL and NLP here.

Here was my advice to Thomas. You're free to attack it viciously.

I think the crucial question is about your goals: do you want to be an academic working on high level problems like parsing and discourse (in which case you're looking at getting a PhD), or do you want to get a job in industry (a PhD is good in industry, but there are plenty of NLP jobs for Master's level, even some for Bachelors)?

If industry is your answer, the school you choose won't really matter that much; it's the skills you develop. I'd strongly advise you to develop competency with machine learning, if you haven't already. You don't have to be great at it, just competent. That's a highly marketable skill set now, and will be for the foreseeable future. General competence with statistics and corpus linguistics is highly valued.

So, I'd ask each program where stats and ML fit into their programs (or how much flexibility they give you for taking electives).

And, just for the record, SUNY Buffalo has an M.S. in CL too. Not too late to apply. You can kinda surf Lake Erie (gotta be better than Georgia surfing).

(pssst, context for the surfing reference can be found on Thomas’ profile).

I scanned the last 10 or so NLP related non-academic job postings on The Linguist List and found a fair bit of consistency in the skills they were asking for. Above all else, they all wanted good programming skills. If you search Monster.com for "computational linguistics" I think you'll see an even greater emphasis on programming skills.

Here's a representative sample of the "requirements" from those Linguist List job postings. Taken all together, they may look intimidating, but this is a mash-up of ten+ postings. It's just meant to sketch what industry is looking for.

• Experience in one or more of the following: MS SQL Database Server; Internet Information Services/Apache Tomcat; Windows operating systems;.NET; Java.
• Strong programming skills in at least two of the following programming languages: Python, C++, Java and Perl
• Multimodal statistical algorithms for language processing and modeling in both speech and handwriting applications
• Develop tools for efficiently processing corpora of speech and/or sketch/handwriting data;
• Work with a team of researchers and developers to successfully integrate research components and validate functionality;
• Experience desired with statistical language modeling for either speech or handwriting applications (e.g., familiarity with CMU-Cambridge LM toolkit, SRILM toolkit, ATT FST toolkit, MALLET, Libbow, etc.);
• Strong algorithmic skills and analytical background;
• Demonstrated success in working in a fast-paced environment;
• Ability to work effectively and successfully either independently and/or in a collaborative team environment.
• Experience in the creation and exploitation of domain and task ontologies in text analytics
  
• Strong background in statistical modelling required.
  
• Knowledge of machine translation or natural language processing techniques
  
• Ability to perform linguistic data analysis.
  
• Proficiency in one or more scripting languages (Perl, Python, Ruby) or programming languages, particularly C++, is a plus.
  
• MS or PhD in Computational Linguistics or related field.
  
• Work experience in production-quality NLP systems.
  
• Familiarity with Unix/Linux operating system environment is a plus.
  
• Experience in machine learning, information retrieval, or data mining are all pluses.
  
• Experience in the building of domain-specific ontologies is useful
  
• Experience in statistical analysis and machine learning
  
• Development, analysis, and support of grammar engine rules for English
  
• Experience in corpus or text analysis, conversation analysis, or computational linguistics
  
• Experienced architect/developer to design scalable enterprise application friendly implementations of spell checking, sentiment, named entity extraction

Bacon Strength

Having only just recently taken the NetFlix plunge, I had been ignoring the flurry of interest amongst computational linguists about Recommender Systems. I am now fully aware of the profound need and utility of improving said systems. Somehow, NetFlix got from the set [Blue Velvet, Chinatown, Midnight Cowboy] to the recommendation The Wild Bunch. There must be a sub-culture growing around the absurdity and humor derivable from such recommendations. Imagine you decided to follow such recommendation religiously. Honestly, how long would it take you to get to Glitter? Scary thought, huh? Now you realize how crucial Recommender Systems are to the survival of humankind.

It seems to me that an automated version of Six Degrees of Kevin Bacon ought to work AT LEAST this well, right? You simply recommend any movie that shares a cast member with a rated movie. The closer two movies are in a Kevin Bacon network, the more strongly you recommend it. Let's call this Bacon Strength. Hmmmmm, wait a second, I might be on to something ... this could be bigger than Google ... why am I telling YOU people about this ... the idea is mine, do you hear! MINE!!!!

Plus, I'm completely amazed that at least four Chuck Norris movies are available for immediate online viewing, but only the first season of the new Dr. Who. wtf?

Speaking English

Steven Levitt, Freako-economist, posted this tempting morsel recently:

I got an email the other day from a blog reader who tells me that there are now more non-native English speakers than native English speakers.

Having silly expectations of writers, I foolishly assumed Levitt would tell us all WHERE this fact held true. If he is referring to The U.S., then it's quite a remarkable claim. China, not so much. He seems to be claiming that some change has occurred where a once predominately English speaking country is no longer so. Unfortunately, his post never answers this, rather he is just looking for a cute way to transition from a story about Malaysian baby names to a modestly humorous email about Jello. It's a blogger's prerogative to tease readers into reading on, so no harm done.

But, I can't help wondering just what was he referring to in his introductory sentence? Has Malaysia ever been predominantly English speaking? As far as I know, no. The current Ethnologue report says this: "National or official language: Malay. Also includes Burmese, Chinese Sign Language, Eastern Panjabi (43,000), Malayalam (37,000), Sylheti, Telugu (30,000)."

No English.

So, can any of you, dear readers, come up with a once predominately English speaking country that is no longer so? A nice little challenge.

Google Linguistics

Erin made the following well-taken point in a comment to this earlier post:

This appeal to the authority of Google is troublesome in linguistics, since we often refer to Google results for evidence for hypotheses about usage. That is documents indexed by Google as a data source, rather than its search results as authoritative figure, of course, but this may not be obvious to the average Joe. :\

I have used Google repeatedly to find instances of constructions that I could not find using standard corpus linguistics methods with hand compiled corpora like the BNC. Typically I’m looking for any instance, just to prove people really do say the thing I’m claiming is possible. For example, I needed to find some examples of passivized complements embedded under 60 different barrier verbs following this pattern:

a. I banned John from being examined by the doctor.
b. I banned John from getting examined by the doctor.

Many of the verbs I wanted to search for are low frequency in the BNC (e.g., barricade, derail, hamper, etc) so the likelihood of finding examples of passivized complements using say a Tgrep2 search is low. So, I ventured into the scary land of Google Linguistics. I used the search query “verbed * from being” and “verbed * from getting” Within a short time, I had multiple examples for most of the verbs I was looking for. I can’t imagine performing this task more efficiently with any other tool. Google really worked well under those circumstances.

Let me note that I have not used Google hit counts or page counts to derive any statistics regarding frequency of occurrence, though. When I do this sort of thing, I’m careful to use my common sense to decide if a return is from a native speaker or not, and often what I do is skim a page to see if there are any obvious ESL errors. Also, I use my own intuition regarding the acceptability of a usage (by pure coincidence, Peter Ludlow from U. Toronto will be here in Buffalo this week giving a talk on the role of linguistic intuitions).

One of the more thorough discussions of the use of search engines in linguistics research is Adam Kilgarriff’s “Googleology is bad science”, a squib from Computational Linguistics (2007, v33, 1)

He writes that the web is attractive to linguists because it is “enormous, free, immediately available, and largely linguistic”. But, he points out four major flaws:

1. search engines do not lemmatise or part-of-speech tag
2. search syntax is limited
3. there are constraints on numbers of queries and numbers of hits per query
4. search hits are for pages, not for instances.

Kilgarriff offers this alternative: “work like the search engines, downloading and indexing substantial proportions of the web, but to do so transparently, giving reliable figures, and supporting language researchers’ queries”

The squib goes on to detail how we might go about doing that in a principled way. It’s well worth the read.

"According to Google,..."

Being both a poker player and former writing teacher, I am better acquainted than most with just how stupid the average person is. The fear that this day would come has lurked in my mind for some time, but today, I re-discovered the ugly truth that people just don't understand even the most basic tenants of reason, research, skepticism, and critical thinking.

Through a series of blog links, I happened on to the comment thread for a popular TV/radio talk show host's web page (I refuse to link to it). The topic regarded one of the current presidential candidates' alleged ethnicity (clearly false/ridiculous hypotheses peppered the thread). I have long since been accustomed to idiocy regarding high profile public figures, so none of this interested me, until I skimmed past one commenter whose attempt at validating the allegation started with "According to google,..." and proceeded to quote some unspecified website. This would be a classic case of argument from authority were it not for the fact that the mere Google search engine alone was being treated as the authority in question.

If Google returns it, it must be true.

There is a scary group of idiots out there who, deep in their hearts, believe that Google magically filters their search returns for QUALITY. Hence, Google is being treated as a primary source.

"Burn down the mission, if we're gonna stay alive..."

The Ling-O-Sphere Revisited

In December I posted about an idea regarding my desire to see a linguistics blog aggregator that "automatically checks a given set of linguistics websites, then updates a topic cloud which clusters posts according to relevance for a particular topic" (see my full post and relevant comments here ).

I see now that William Cohen at his Cranial Darwinism blog has recently posted two new academic papers on the automatic discovery of blog topics (aka, latent topic modeling) as well automatic methods of modeling blog influence. Daume has posted on related topics in the past as well (see here for one relevant post).

Having skimmed the first paper a bit, I see lots of scary words and phrases like "Latent Dirichlet Allocation" and "probabilistic framework"; I'm neck deep in finishing my dissertation (or failing to finish it; I'll be able to distinguish the two in about 3 weeks), so my interest in struggling through challenging papers is low, but they look well worth the read ... someday ... sigh.

On The Cognitive Properties of Skin

After posting on the voiceless phone call story below, I began searching around for more information on how the device actually works. Failing to find any relevant patents pending (suspicious, I thought) I began searching for information on Michael Callahan, the wunderkind who appears to be the principle inventor, though many are probably involved.

After some searching, the most specific information I have yet found on the technology behind the voiceless phone was found in this article from the University of Illinois at Urbana-Champaign Engineering department website. Note the passage I have emphasized:

“Once we hit upon the idea of direct input, we were off and running,” explained Thomas Coleman, a project team member. The young researchers discovered that information sent from the brain can be accurately measured through the conductive properties of the skin. Typically, according to Coleman, these measurements are obtained through rigid metallic electrodes which neither respond to natural movements of the body nor to increasing skin moisture. They often become very uncomfortable under prolonged use.

"Our system uses proprietary technology to gather neurological information through encapsulated conductive gel pads, shielding the embedded electrode from the skin,” Coleman said. ‘The Audeo’ device we developed applies gentle pressure over the vocal cords, while the form-fitting band automatically adjusts in diameter, accommodating head and neck movements to maintain efficient contact.”

From there, team members created a computer program, which reads the intercepted neurological signals, and communicates a ‘response,’ both on the screen and as an audio signal. Initial work centered on determining the differences between a ‘yes’ and a ‘no’ response, which could be recognized by the computer. The software has since been enhanced to effectively ‘learn’ and adapt to the user’s neurological signals without the need of extensive training. The equipment analyzes the user during a one-time calibration process and generates a personalized user identity. (my emphasis; quote marks had to be manually inserted to replace funny characters, but i tried to represent the original faithfully)

This is how far removed from serious neuroscience I am. I had no clue. I realized some information could be gathered from the skin, like Galvanic skin response, but I must say I’m shocked to learn that phonemic information regarding unarticulated utterances can be retrieved from the skin around a person’s neck. Clearly, there is more to this story. I’ll keep digging.