ClusType
Source code for SIGKDD'15 paper ClusType: Effective Entity Recognition and Typing by Relation Phrase-Based Clustering (Slides).
Given a text corpus (e.g., a collection of news articles), it performs automatically entity extraction and typing using distant supervision (i.e., examples from external knowledge bases like Freebase). For example, from the sentence "The best BBQ Iβve tasted in Phoenix " the system will recognize BBQ as food and phoenix as location. More background can be found in our WWW'16 tutorial.
ClusType works on coarse-grained entity types (e.g., Person, Location, Organization); for more fine-grained entity typing, please use AFET (Ren et al., EMNLP'16).
Data
- NYT:
- Yelp:
- Corpus: 230k Yelp reviews sampled from Yelp Dataset (download)
- Seed entities: entity linking result by DBpediaSpotlight (download)
- Tweet:
System Output & Evaluation
The system output on NYT dataset can be downloaded from here. We evaluated the result over ~1k (20,874 annotated entity mentions) gold standard set. Sample output on 50k Yelp reviews can be download from here.
Evaluate the result:
python src/evaluation.py -ResultPath -GroundTruthPath
Dependencies
- python 2.7
- numpy, scipy, scikit-learn, lxml, TextBlob and related corpora
$ sudo pip install numpy scipy sklearn lxml textblob
$ sudo python -m textblob.download_corpora
Default Run
$ ./run.sh
Run.sh - File path setup
We take Yelp dataset as an example.
Input: text corpus path.
RawText='data/yelp/yelp_230k.txt'
- format: "docId \TAB document \n"
Input: type mapping file path.
TypeFile='data/yelp/type_tid.txt'
- format: "type name \TAB typeId \n". "NIL" means "Not-of-Interest"
Input: mapping between Freebase and DBpedia entities.
FreebaseMap='data/freebase_links.nt'
- Download Freebase-to-DBpedia mapping file. Place it under "data/" directory
Output: output file from candidate generation (format: "docId \TAB segmented sentence \n").
SegmentOutFile='result/segment.txt'
- Segments are separated by ",". Entity mention candidates are marked with ":EP". Relation phrases are marked with ":RP".
Output: entity linking result (please download the corresponding seed entity files).
SeedFile='data/yelp/seed_yelp.txt'
- Format: "docId \TAB entity name \TAB Original Freebase Type \TAB Refined Type \TAB Freebase EntityID \TAB Similarity Score \TAB Relative Rank \n".
- NOTE: Our entity linking module calls DBpediaSpotLight Web service, which has limited querying speed. This process can be largely accelarated by installing the tool on your local machine Link.
Output: entity mentions found in each document.
ResultFile='result/yelp/results.txt'
- Format: "docId \TAB entity mention \TAB entity type \n".
Output: In-text annotation of entity mentions.
ResultFileInText='result/yelp/resultsInText.txt'
Run.sh - Model parameters
Threshold on significance score for candidate generation.
significance="2"
Switch on capitalization feature for candidate generation.
capitalize="1"
Maximal phrase length for candidate generation.
maxLength='4'
Minimal support of phrases for candidate generation.
minSup='30'
Number of relation phrase clusters.
NumRelationPhraseClusters='500'
Reference
@inproceedings{ren2015clustype,
title={Clustype: Effective entity recognition and typing by relation phrase-based clustering},
author={Ren, Xiang and El-Kishky, Ahmed and Wang, Chi and Tao, Fangbo and Voss, Clare R and Han, Jiawei},
booktitle={Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining},
pages={995--1004},
year={2015},
organization={ACM}
}