| Christina Sauper | Aria Haghighi | Regina Barzilay |
|---|---|---|
| csauper@csail.mit.edu | aria42@gmail.com | regina@csail.mit.edu |
In this paper, we investigate how modeling content structure can benefit text analysis applications such as extractive summarization and sentiment analysis. This follows the linguistic intuition that rich contextual information should be useful in these tasks. We present a framework which combines a supervised text analysis application with the induction of latent content structure. Both of these elements are learned jointly using the EM algorithm. The induced content structure is learned from a large unannotated corpus and biased by the underlying text analysis task. We demonstrate that exploiting content structure yields significant improvements over approaches that rely only on local context.
Full Text: http://groups.csail.mit.edu/rbg/code/content_structure/sauper-emnlp-10.pdf
This code is available for research use only.
These instructions mainly pertain to the Amazon and Yelp data sets for the multi-aspect phrase extraction task.
Source code is included, but to just get started running the system quickly, I
recommend installing maven2, then compiling with mvn compile in the base
directory. After that, run the system as follows (substituting your desired
config file and memory usage):
java -ea -server -mx5G -Djava.ext.dirs=lib -cp target/classes phrase.jointtopic.Main amazon-conf.yaml
amazon-conf.yaml
yelp-conf.yaml
These are the config files which specify parameters for the model. See GlobalOptions.java for more potential options.
conf/amazon.list
conf/yelp.list
Lists of token files (see below for format). Those marked as
TRAIN_LABELED will be used as labeled input documents; those marked as
TEST_LABELED will be used at test time.
We performed experiments on three separate corpora, a set of Amazon HDTV reviews (59 labeled, 12.8k unlabeled), a set of Yelp restaurant reviews (96 labeled, 31k unlabeled), and a set of IGN DVD reviews (665 labeled).
*.tok
Tokenized file, one word per line. The columns of this file are as follows:
word sentence # word #(sent) start-char end-char start-char(sent) end-char(sent)
*.ann
Annotations corresponding to the tokenized file; one word's label per line. Some annotation files are tagged with begin / inside / end tokens; ability to automatically strip these is controlled by an option in the config file.
A demo system is online at http://condensr.com. Backend code is online at https://github.com/csauper/condensr.
The demo system puts together results from multi-aspect phrase extraction on Yelp reviews with Google maps and other restaurant metadata in order to display a concise summary of restaurants based on a search in a particular area.
Each restaurant summary presents the following:
- Basic restaurant information (name, address, location)
- Automatically extracted and categorized highlights from all Yelp reviews for the restaurant in several categories (food, ambiance, service, value, overall)
- Automatically determined sentiment classification (positive, negative, neutral)
- Supporting highlights to expand on the main points