Dataset and papers for baseline I'm doing a project about Topic Detection and Tracking in text. I need to perform a baseline so I can compare existing results with mine.
I read some papers where they use datasets that are not so easy to get, for various reasons. For example:


*

*They use some sort of API to get text from a website at a certain time;

*They use a dataset from many years ago and are nowhere to be found;

*They use a dataset of which they ask hundreds or even thousands of dollars for.


Of course, to see if my recreation of their baseline was successful, I need the exact same dataset so I can compare mine with their results. I was wondering if someone can point me to a research paper (more is even better) where they use a dataset which is free and still available for download. And where they use it with different classifiers (Naïve Bayes, kNN, SVM, Decision Trees, etc.) showing their corresponding results.
 A: In my dissertation I compare performance of various classification algorithms for identifying articles of interest for a neuroscience-related knowledge base. I made the text of the articles used for my experiment available via github, and you're welcome to use them! The document-level corpus uses the text in medline records, which can be obtained via the PubMed search engine, and can also be accessed programmatically (e.g., via python scripts). The document ids in my data set correspond to the unique pubmed identifiers in their data set!
A: Some of the semi-standard data sets for information retrieval are:


*

*The 20 newsgroups data set, from usenet newsgroups. 

*The various versions of the Reuters Corpus (news articles), including


*

*RCV1/2/TRC2 (available from NIST if you fill out a form: see here). David D. Lewis distributes some data derived from the corpus, which might help you get started quickly too. 

*Reuters-21578 is available here. Here's a head-to-head comparison of a bunch of algorithms on it Joachims (1998)


*Enron Emails: emails from the FERC's investigation in Enron's malfesience. Available here, accompanied by a list of papers that have used the data set

*Data sets from TREC are often not publicly available, but might be accessible if you're at an institution with an LDC subscription.


There are some "specialized" data sets (e.g., biomedical text annotated with gene names or protein names); others are tagged with different types of entities, like temporal expressions.
If you want Twitter data, they've been very reluctant to let people re-distribute tweets (especially now that they sell bulk access to them), but people have distributed code that makes downloading them, via the Twitter API, pretty painless. 
It really depends on what you want to do. 
