No description, website, or topics provided.
Java Python
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
.settings
data
lib
models
resource
src
.gitignore
.project
.pydevproject
CATENA.png
README.md
experiment_result.xlsx
extractVectors.py
pom.xml

README.md

CATENA

CAusal and TEmporal relation extraction from NAtural language texts

CATENA is a sieve-based system to perform temporal and causal relation extraction and classification from English texts, exploiting the interaction between the temporal and the causal model. The system requires pre-annotated text with EVENT and TIMEX3 tags according to the TimeML annotation standard, as these annotation are used as features to extract the relations.

Requirements

  • Java Runtime Environment (JRE) 1.7.x or higher

Maven

CATENA is now available on Maven Central. Please add the following dependency in your pom.xml.

<dependency>
  <groupId>com.github.paramitamirza</groupId>
  <artifactId>CATENA</artifactId>
  <version>1.0.2</version>
</dependency>

To build the fat (executable) JAR:

  • Install the WS4J library in your local Maven repo, e.g., mvn install:install-file -Dfile=./lib/ws4j-1.0.1.jar -DgroupId=edu.cmu.lti -DartifactId=ws4j -Dversion=1.0.1 -Dpackaging=jar
  • Run mvn package to build the executable JAR file (in target/CATENA-<version>.jar).

Text processing tools:

Other libraries:

Other resources:

  • Temporal and causal signal lists, available in resource/. This folder must be placed within the root folder of the project.
  • Classification models, available in models/, including: catena-event-timex.model, catena-event-dct.model, catena-event-event.model and catena-causal-event-event.model.

Usage

! The input file(s) must be in the TimeML annotation format or CoNLL column format (one token per line) !

usage: Catena
 -i,--input <arg>        Input TimeML file/directory path
 -f,--col                (optional) Input files are in column format (.col)
 -tl,--tlinks <arg>      (optional) Input file containing list of gold temporal links
 -cl,--clinks <arg>      (optional) Input file containing list of gold causal links
 -gl,--gold              (optional) Gold candidate pairs to be classified are given
 -y,--clinktype          (optional) Output the type of CLINK (ENABLE, PREVENT, etc.) from the rule-based sieve
        
 -x,--textpro <arg>      TextPro directory path
 -l,--matelemma <arg>    Mate tools' lemmatizer model path   
 -g,--matetagger <arg>   Mate tools' PoS tagger model path
 -p,--mateparser <arg>   Mate tools' parser model path      
 
 -t,--ettemporal <arg>   CATENA model path for E-T temporal classifier    
 -d,--edtemporal <arg>   CATENA model path for E-D temporal classifier                       
 -e,--eetemporal <arg>   CATENA model path for E-E temporal classifier
 -c,--eecausal <arg>     CATENA model path for E-E causal classifier
 
 -b,--train              (optional) Train the models
 -m,--tempcorpus <arg>   (optional) Directory path (containing .tml or .col files) for training temporal classifiers
 -u,--causcorpus <arg>   (optional) Directory path (containing .tml or .col files) for training causal classifier     

For example

java -Xmx2G -jar ./target/CATENA-1.0.2.jar -i ./data/example_COL/ --col --tlinks ./data/TempEval3.TLINK.txt --clinks ./data/Causal-TimeBank.CLINK.txt -l ./models/CoNLL2009-ST-English-ALL.anna-3.3.lemmatizer.model -g ./models/CoNLL2009-ST-English-ALL.anna-3.3.postagger.model -p ./models/CoNLL2009-ST-English-ALL.anna-3.3.parser.model -x ./tools/TextPro2.0/ -d ./models/catena-event-dct.model -t ./models/catena-event-timex.model -e ./models/catena-event-event.model -c ./models/catena-causal-event-event.model -b -m ./data/Catena-train_COL/ -u ./data/Causal-TimeBank_COL/

CoNLL column format

The input document must be in tab-separated 'one-token-per-line' format, with each column as: | token | token-id | sentence-id | lemma | event-id | event-class | event-tense+aspect+polarity | timex-id | timex-type | timex-value | signal-id | causal-signal-id | pos-tag | chunk | lemma | pos-tag | dependencies | main-verb |

  • event-id and event-class: TimeML event ID and attributes
  • timex-id and timex-type and timex-value: TimeML timex ID and attributes
  • signal-id and causal-signal-id: temporal and causal signal ID
  • event-tense+aspect+polarity: optional attributes of an event, if given O, CATENA will infer them automatically according to PoS tags and dependency relations
  • pos-tag: BNC tagset (default tagset uset to build the models) or Penn Treebank tagset
  • chunk:
  • dependencies: in the format of dep1:deprel1||dep2:deprel2||..., dependency relations are resulted from Mate-tools

See for example data/example_COL/.

Output format

The output will be a list of temporal and/or causal relations, one relation per line, in the format of:

filename  entity_1  entity_2  TLINK_type/CLINK/CLINK-R
  • TLINK_type: One of TLINK types according to TimeML, e.g., BEFORE, AFTER, SIMULTANEOUS
  • CLINK: entity_1 CAUSE entity_2
  • CLINK-R: entity_1 IS_CAUSED_BY entity_2

System architecture

alt tag

CATENA contains two main modules:

  1. Temporal module, a combination of rule-based and supervised classifiers, with a temporal reasoner module in between.
  2. Causal module, a combination of a rule-based classifier according to causal verbs, and supervised classifier taken into account syntactic and context features, especially causal signals appearing in the text.

The two modules interact, based on the assumption that the notion of causality is tightly connected with the temporal dimension: (i) TLINK labels for event-event pairs, resulting from the rule-based sieve + temporal reasoner, are used for the CLINK classifier, and (ii) CLINK labels are used as a post-editing method for correcting the wrongly labelled event pairs by the Temporal module.

Publication

Paramita Mirza and Sara Tonelli. 2016. CATENA: CAusal and TEmporal relation extraction from NAtural language texts. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, Osaka, Japan, December. [pdf]

Dataset

  • Training data for the Temporal module is taken from the TempEval-3 shared task, particularly the combination of TBAQ-cleaned (English training data) and TE3-platinum (English test data).
  • Training data for the Causal module is Causal-TimeBank, the TimeBank corpus annotated with causal information.
  • TimeBank-Dense corpus is used in one of the evaluation schemes for temporal relation extraction.
  • Causal-TempEval3-eval.txt (available in data/) is used in one of the evaluation schemes for causal relation extraction.

! Whenever making reference to this resource please cite the paper in the Publication section. !

Web Service

Soon!

Contact

For more information please contact Paramita Mirza (paramita135@gmail.com).