Stanford NER Tool -- training for a new domain - stanford-nlp

What is the amount of sentences needed to effectively train the CRF for a domain like restauarants (restaurant names, addresses, cusines) or music (artist name, song name genre).

As a point of reference, I believe the CoNLL training data for (location, organization, person, misc) NER has around 14,000 sentences.

It depends a lot of the kind of data you will be tagging and how variable it will be. I've worked on a project also involving restaurant and music domains. In my case we would be handling user queries, which tend to be short and don't present that much variability (particularly for restaurant, but not for music, which is a very noisy domain).
For the restaurant domain, training it with ~2k sentences was fine, but of course, if you can get more data, your model will be much more accurate.
For music, the situation is a little bit more tricky since song/band names can be virtually anything. In this case, only data alone might not be enough to get an acceptable accuracy. In my project we used ~5k for music and many features and some additional post-processing to get things right.

Related

What matching algorithm could I use?

I would need some help because I don't know what algorithm i could use for the following (I use python) :
Steve is 25 and he buys everyday orange juice
Maria is 23 and she likes to buy smoothies
Steve & Maria tastes are pretty much the same.
Juan is 16 and he only drinks sodas
Juan tastes are not the same as Steve and Maria.
====================================================
I would like to use a matching algorithm that will detect the users who have the same drink preference and a close age. To continue with the example, Steve and Maria would be matched together but not Juan. Which one should I use ?
I agree with #klutt that your task is pretty vague. There are two approaches that come to mind, but not knowing more details about your problem does limit the details I can provide in my answer that would help you. I am interpreting the question as if you are taking in raw text and might want to process more sentences that have very similar semantic and syntactical structure.
An algorithmic approach:
Assuming that your word choices are static in their semantic meaning (Maria is 23 ... Steve is 25), we can parse each sentence and identify tokens like is or and or same and essentially perform lexical analysis on the text... from here, you could continue thinking about how you would go about matching and so forth... but this is rather complicated...
Neural Network approach:
If you are taking in raw text in the form of sentences, it's a problem that's not straight forward to solve using a top-down algorithmic approach.
You could take an approach with neural networks that trains a model to solve your problem, but then again what you seem to be asking is quite complex since there are multiple "facts" within each sentence that are not semantically related. For example, your second sentence identifies that Maria is 23 but at the end of that sentence there is a comparison between Steve and Maria. And your first sentence only identifies Steve as 25.
Even if you chunk raw text into sentences, you would have to have a very fine tuned neural network architecture and a lot of training data to get remotely close to your goal.
Now, both of those solutions are very complex... but if you wanted to create an application that collects this data (via a form or prompt) and puts it into a structured format (like a json or xml object) to organize and store the data in memory (perhaps writing out to a database or file for persistent storage), that might be a good route to go down.
This can serve as a good lesson in how to think about data as well. It is one thing if you have a pool of thousands of sentences, just raw data that you need to organize for quantitative purposes (classic qualitative -> quantitative problems). It is another thing if you are going to be collecting this data. If you are going to be collecting data, having a program that collects and organizes names, ages, and drink preferences (and then organizes that data within certain data structures), then we can talk about matching algorithms.
I will also add here that if you do have structured data, Collaborative filtering (mentioned by Shridhar) is a great starting place.
Collaborative filtering best suits your needs.
In the newer, narrower sense, collaborative filtering is a method of
making automatic predictions (filtering) about the interests of a user
by collecting preferences or taste information from many users
(collaborating). The underlying assumption of the collaborative
filtering approach is that if a person A has the same opinion as a
person B on an issue, A is more likely to have B's opinion on a
different issue than that of a randomly chosen person. For example, a
collaborative filtering recommendation system for television tastes
could make predictions about which television show a user should like
given a partial list of that user's tastes (likes or dislikes).[3]
Note that these predictions are specific to the user, but use
information gleaned from many users. This differs from the simpler
approach of giving an average (non-specific) score for each item of
interest, for example based on its number of votes.

How to create a gazetteer based Named Entity Recognition(NER) system?

I have tried my hands on many NER tools (OpenNLP, Stanford NER, LingPipe, Dbpedia Spotlight etc).
But what has constantly evaded me is a gazetteer/dictionary based NER system where my free text is matched with a list of pre-defined entity names, and potential matches are returned.
This way I could have various lists like PERSON, ORGANIZATION etc. I could dynamically change the lists and get different extractions. This would tremendously decrease training time (since most of them are based on maximum entropy model so they generally includes tagging a large dataset, training the model etc).
I have built a very crude gazetteer based NER system using a OpenNLP POS tagger, from which I used to take out all the Proper nouns (NP) and then look them up in a Lucene index created from my lists. This however gives me a lot of false positives. For ex. if my lucene index has "Samsung Electronics" and my POS tagger gives me "Electronics" as a NP, my approach would return me "Samsung Electronics" since I am doing partial matches.
I have also read people talking about using gazetteer as a feature in CRF algorithms. But I never could understand this approach.
Can any of you guide me towards a clear and solid approach that builds NER on gazetteer and dictionaries?
I'll try to make the use of gazetteers more clear, as I suspect this is what you are looking for. Whatever training algorithm used (CRF, maxent, etc.) they take into account features, which are most of the time:
tokens
part of speech
capitalization
gazetteers
(and much more)
Actually gazetteers features provide the model with intermediary information that the training step will take into account, without explicitly being dependent on the list of NEs present in the training corpora. Let's say you have a gazetteer about sport teams, once the model is trained you can expand the list as much as you want without training the model again. The model will consider any listed sport team as... a sport team, whatever its name.
In practice:
Use any NER or ML-based framework
Decide what gazetteers are useful (this is maybe the most crucial part)
Affect to each gazetteer a relevant tag (e.g. sportteams, companies, cities, monuments, etc.)
Populate gazetteers with large lists of NEs
Make your model take into account those gazetteers as features
Train a model on a relevant corpus (it should containing many NEs from gazetteers)
Update your list as much as you want
Hope this helps!
You can try this minimal bash Named-Entity Recognizer:
https://github.com/lasigeBioTM/MER
Demo: http://labs.fc.ul.pt/mer/

How do services like fflick work? What algorithms do they use?

Services like fflick, mombo etc perform analysis of tweets about movies. They seem to process hundreds of thousands of tweets.
How do they match a tweet with a movie? For example, lets say there is a movie called "unknown". How do they determine if a tweet talks about unknown-the movie or unknown something else?
How are they able to collect so many tweets? Streaming API?
Do they maintain a list of movie names and check each tweet against this list, to find out if a tweet is referencing a particular movie?
The following are just my guesses.
Certainly a list of movie names is required. That's a necessary first step in trimming the tweets down to a subset that could possibly refer to a movie.
A movie title is either recognisable from the words themselves (e.g. "Terminator 2") or it requires the author to disambiguate it (e.g. "Unknown" -- or "Gone With the Wind", which could be referring to either the movie or the book). In the latter case, a variety of clues will be provided. Perhaps most obviously:
Anything that follows a phrase like "Just saw" or "Watched" is highly likely to be a movie name. Less so anything following "Read".
If the the name of the director or an actor in the film is mentioned, it's likely to be referring to the movie.
Twitter content is heavily skewed towards the latest thing, so the probability that a movie is being discussed drops as the time since the movie hit the theatres increases.
If a tweet is in response to a another tweet known with high probability to be referring to a particular movie, then it is probably about the same movie.
I expect that criteria like the above are used to assign probabilities for classification according to some weights, and that the usual techniques have been applied to tweak the weights to give good predictions. I would expect a supervised machine learning approach: essentially, have some humans classify a few hundred tweets, then optimise the weights for performance on some subset of this dataset, and finally test how well the chosen weights work for classifying the remainder of the dataset (this is to check that overfitting has not occurred).

Media recommendation engine - Single user system - How to start

I want to implement a media recommendation engine. I saw a similar posts on this, but I think my requirements are bit different from those, so posting here.
Here is the deal.
I want to implement a recommendation engine for media players like VLC, which would be an engine that has to care for only single user. Like, it would be embedded in a media player on a PC which is typically used by single user. And it will start learning the likes and dislikes of the user and gradually learns what a user likes. Here it will not be able to find similar users for using their data for recommendation as its a single user system. So how to go about this?
Or you can consider it as a recommendation engine that has to be put in say iPods, which has to learn about a single user and recommend music/Movies from the collections it has.
I thought of start collecting the genre of music/movies (maybe even artist name) that user watches and recommend movies from the most watched Genre, but it look very crude, isn't it?
So is there any algorithms I can use or any resources I can refer up to?
Regards,
MicroKernel :)
What you're trying to do is quite challenging... particularly because it's still in the research stage and a lot of PHDs from reputable universities across the world are trying to get a good solution for that.
SO here are some things that you might need:
Data that you can analyze:
Lots, and lots, and lots of data!
It could be meta data about the media (name, duration, title, author, style, etc.)
Or you can try to do some crazy feature extraction from the media itself.
References to correlate the data to.
Since you can't get other users, you always need the user feedback.
If you don't want to annoy your user to death with feedback questions, then make your application connect to a central server so you can compare users.
An algorithm that can model your data sufficiently well.
If you have no experience at all, then try k-nearest neighbor (the simplest one).
Collaborative filtering
Pearson Correlation
Matrix Factorization/Decomposition
Singular value decomposition (SVD)
Ensemble learning <-- Allows you to combine multiple algorithms and take advantage of their strengths.
The winners of the NetFlix prize said this:
Predictive accuracy is substantially
improved when blending multiple
predictors. Our experience is that
most efforts should be concentrated in
deriving substantially different
approaches, rather than refining a
single technique. Consequently, our
solution is an ensemble of many
methods.
Conclusion:
There is no silver bullet for recommendation engines and it takes years of exploration to find a good combination of algorithms that produce sufficient results. :)

Algorithms to find stuff a user would like based on other users likes

I'm thinking of writing an app to classify movies in an HTPC based on what the family members like.
I don't know statistics or AI, but the stuff here looks very juicy. I wouldn't know where to start do.
Here's what I want to accomplish:
Compose a set of samples from each users likes, rating each sample attribute separately. For example, maybe a user likes western movies a lot, so the western genre would carry a bit more weight for that user (and so on for other attributes, like actors, director, etc).
A user can get suggestions based on the likes of the other users. For example, if both user A and B like Spielberg (connection between the users), and user B loves Batman Begins, but user A loathes Katie Holmes, weigh the movie for user A accordingly (again, each attribute separately, for example, maybe user A doesn't like action movies so much, so bring the rating down a bit, and since Katie Holmes isn't the main star, don't take that into account as much as the other attributes).
Basically, comparing sets from user A similar to sets from user B, and come up with a rating for user A.
I have a crude idea about how to implement this, but I'm certain some bright minds have already thought of a far better solution already, so... any suggestions?
Actually, after a quick research, it seems a Bayesian filter would work. If so, would this be the better approach? Would it be as simple as just "normalizing" movie data, training a classifier for each user, and then just classify each movie?
If your suggestion includes some brain melting concepts (I'm not experienced in these subjects, specially in AI), I'd appreciate it if you also included a list of some basics for me to research before diving into the meaty stuff.
Thanks!
Matthew Podwysocki had some interesting articles on this stuff
http://codebetter.com/blogs/matthew.podwysocki/archive/2009/03/30/functional-programming-and-collective-intelligence.aspx
http://codebetter.com/blogs/matthew.podwysocki/archive/2009/04/01/functional-programming-and-collective-intelligence-ii.aspx
http://weblogs.asp.net/podwysocki/archive/2009/04/07/functional-programming-and-collective-intelligence-iii.aspx
This is similar to this question where the OP wanted to build a recommendation system. In a nutshell, we are given a set of training data consisting of users ratings to movies (1-5 star rating for example) and a set of attributes for each movie (year, genre, actors, ..). We want to build a recommender so that it will output for unseen movies a possible rating. So the inpt data looks like:
user movie year genre ... | rating
---------------------------------------------
1 1 2006 action | 5
3 2 2008 drama | 3.5
...
and for an unrated movie X:
10 20 2009 drama ?
we want to predict a rating. Doing this for all unseen movies then sorting by predicted movie rating and outputting the top 10 gives you a recommendation system.
The simplest approach is to use a k-nearest neighbor algorithm. Among the rated movies, search for the "closest" ones to movie X, and combine their ratings to produce a prediction.
This approach has the advantage of being very simple to easy implement from scratch.
Other more sophisticated approaches exist. For example you can build a decision tree, fit a set of rules on the training data. You can also use Bayesian networks, artificial neural networks, support vector machines, among many others... Going through each of these wont be easy for someone without the proper background.
Still I expect you would be using an external tool/library. Now you seem to be familiar with Bayesian Networks, so a simple naive bayes net, could in fact be very powerful. One advantage is that it allow for prediction under missing data.
The main idea would be somewhat the same; take the input data you have, train a model, then use it to predict the class of new instances.
If you want to play around with different algorithms in simple intuitive package which requires no programming, I suggest you take a look at Weka (my 1st choice), Orange, or RapidMiner. The most difficult part would be to prepare the dataset to the required format. The rest is as easy as choosing what algorithm and applying it (all in a few clicks!)
I guess for someone not looking to go into too much details, I would recommend going with the nearest neighbor method as it is intuitive and easy to implement.. Still the option of using Weka (or one of the other tools) is worth looking into.
There are a few algorithms that are good for this:
ARTMAP: groups via probability against each other (this isn't fast but its the best thing for your problem IMO)
ARTMAP holds a group of common attributes and determines likelyhood of simliarity via a percentages.
ARTMAP
KMeans: This seperates out the vectors by the distance that they are from each other
KMeans: Wikipedia
PCA: will seperate the average of all the values from the varing bits. This is what you would use to do face detection, and background subtraction in Computer Vision.
PCA
The K-nearest neighbor algorithm may be right up your alley.
Check out some of the work of the top teams for the netflix prize.

Resources