Crf Application: Fill & Download for Free

Some of the applications of sequential labeling tasks that I have heard of:Computer Vision: Face and gesture analysis for emotion recognition, object recognition (tree-shaped CRF), image segmentation (grid-shaped CRF).Bioinformatics: Protein structure prediction.NLP: Word alignment in machine translation, proper-noun coreference.

For an up-to-date introduction to knowledge extraction techniques I would recommend you take a look at the recent work in Google's "Knowledge Vault" project. As you can read in their recent paper (Page on cmu.edu), this project is trying to do exactly what you are asking: build a probabilistic model of "knowledge" starting off mainly from unstructured text but also using webpage structure, links, and other auxiliary information. References in that paper will point you in many interesting directions. Note that this approach is very related to the already existing Knowledge Graph, which in turn is related to the idea of Linked data.As for particular techniques you might want to get started with, I recommend Latent Semantic Analysis and Conditional Random Fields (CRF). You should also take a look at the The Stanford NLP (Natural Language Processing) Group toolkit. And you can even finish up by looking at the very hot applications of Deep Learning to this problem (see Deep Learning for NLP)

Since you are an undergrad student, I think something that Gupta mentioned is worthwhile for you to try. As far as I know, many of the parsing models are based on the tree structure which can apply top-down/bottom-up approaches.For simplicity, I will suggest you implement something that is linear structure cause I think tree structure or arbitrary graph structure may be too difficult for you.Some projects that I think you can work on within your ability (all of them are linear structures):Named Entity Recognition using Linear-chain Conditional Random Field (CRF): this task is already well studied by many researchers and using linear-chain CRF model can already achieve a quite promising performance, though you can use some other complicate models.Sentiment Classification by Linear-chain CRF: this task can also be done by linear-chain CRF and you can use some data like Twitter dataset and I think there might be somewhere that you can download the data.Other tasks/applications that you can search by using Linear-chain CRF. I think many and many applications can use. (Web Opinion Mining, Music Analysis, Relational Learning, etc) There’s a lot can be done by linear-chain CRF.Since I think linear-chain structure might be easier for you to get familiar with CRF model also the dynamic programming idea behind. Furthermore, it’s good to know about CRF. Here is a very nice and clean tutorial on linear-chain CRF. That’s also where I started to learn NLP.Furthermore, the nice thing is that you don’t really need to care about the detail implementation (especicallly the math although I recommend to know the idea) since there are many people has already implemented fast and clean linear-chain CRF tool. You can directly use it and develop your own application.I list some of them here:CRFSuite: A fast implementation of CRFCRF++: Also a CRF toolkit.StatNLP: Aimed for developing a tool for any graphical models. But this one is more related to researcher instead of engineer. They haven’t released any versions yet, but I heard that it will be soon.Please leave your comments. I will come to reply for your questions.