~$ git clone git@github.com:islambeltagy/pl-semantics.git
~$ cd pl-semantics

~/pl-semantics$ bin/mlnsem download

~/pl-semantics$ bin/mlnsem install

~/pl-semantics$ bin/mlnsem compile

~/pl-semantics$ bin/mlnsem gen prb		#Generate helping files for a toy dataset I call it prb
~/pl-semantics$ bin/mlnsem run prb		#Run the toy examples of prb

NOTE: You will need to downgrade "bison" to version 2.7 for Alchmey to compile

• Install wget

• Use Makefile.macosx for candc (create a symbolic link)

• Install prolog brew install homebrew/x11/swi-prolog

• To compile alchemy, you need gcc4.9. Install it using brew install homebrew/x11/gcc49

• Alchemy also needs Bison 2.3 (install it), Flex 2.5.4 (already installed) and Perl 5.8.8 (already installed)

• Last step to compile alchemy is to edit alchemy/src/makefile. You will find few comments in the file showing how to edit it.

• Use Java 7. If you want to use a later version of Java, you need to update the version of scala used.

• RTE and STS datasets we have now are FraCas, RTE1, RTE2, RTE3, MsrVid, MsrPar, Prob I will add RTE4-RTE7 and Trento dataset soon. Each dataset has an apprevition (listed in bin/mlnsem)

  • FraCas: frc
  • MsrVid: vid
  • MsrPar: par
  • RTEi: rte i
  • SICK as an RTE task: sick-rte
  • SICK as an STS task: sick-sts
  • Prob: prb #This dataset contains few examples I selected. Each example represents one specific Problem.

• First, some helping files need to be generated for each dataset using:

  ~/pl-semantics$ bin/mlnsem gen DataSetAppreviation

For example:

~/pl-semantics$ bin/mlnsem gen rte 2

• Then, run the system for this dataset:

  ~/pl-semantics$ bin/mlnsem run DataSetAppreviation

• Please check bin/mlnsem for more details

They are all listed in

src/main/scala/utcompling/mlnsemantics/util/Config.scala

Default values are good enough to run the system. Only one argument is not listed, which is the "range" argument.

Let's say you want to run the 4th, 5th, 6th, and 9th pairs of FraCas. Command is:

~/pl-semantics$ bin/mlnsem run frc 4-6,9

• Install Scala's plugin on your eclipse.

• Use sbt to generate eclipse project files:

• java -jar bin/sbt-launch*.jar

• When it starts, type: eclipse

• Two projects are generated pl-semantics and scala-logic. Import them to eclipse and you are done.

Entry point for running the system on Condor is the script

~/pl-semantics$ bin/condor.sh

The script provides the following functionalities:

• Get status of Condor jobs. ARGS are passed to condor_q

  ~/pl-semantics$ bin/condor status ARGS

• Remove all submitted Condor jobs

  ~/pl-semantics$ bin/condor remove

• Submit new jobs to Condor. EXP_DIR: is the directory used to store output from condor. STEP is number of pairs per job. ARGS are the usual arguments you want to pass to bin/mlnsem (execluding the leading "run" and the range argument), and execluding "task" that the script picks automatically based on the dataset. Output files are saved in pl-semantics/EXP_DIR/. Be careful, consecutive submissions of tasks with the same EXP_DIR will overwrite each others. The file: pl-semantics/EXP_DIR/config contains the command line arguments passed to the condor script.

  ~/pl-semantics$ bin/condor submit EXP_DIR STEP ARGS

for example: submit the sick-sts dataset to condor. Output from condor will be saved in condor/firstExp/ . Each Condor job contains 10 pairs of sentences. Run each pair for 30 seconds, and do not print any log. Use PSL for the STS task

~/pl-semantics$ bin/condor submit condor/firstExp 10 sick-sts -task sts -softLogic psl -timeout 30000 -log OFF 

It is actually simpler, because default values are set to make condor experiments easier depending on the dataset, so the command above can be

~/pl-semantics$ bin/condor submit condor/firstExp 10 sick-sts

• Prints a list of the tasks without submitting anything. This is helpful to check the number of tasks and arguments before submitting the tasks.

  ~/pl-semantics$ bin/condor print EXP_DIR STEP ARGS

• In some cases (for reasons I do not understand) some condor tasks break without notice. Calling the condor script with the argument "fix" checks all output files and make sure that all condor tasks terminated correctly. If some of them did not, resubmit them again. Do not call "fix" while some taks are already running . Make sure to use the same EXP_DIR. STEP and ARGS will be read from the "config" file

  ~/pl-semantics$ bin/condor fix EXP_DIR

• Collects results of individual tasks into one block, prints it, and store it in EXP_DIR/result

  ~/pl-semantics$ bin/condor collect EXP_DIR

• Call the weka script (explained below) with the approbriate arguments based on the experiment. Note that "collect" has to be called before "eval" because "eval" uses "collect"'s output that is saved in EXP_DIR/result

  ~/pl-semantics$ bin/condor eval EXP_DIR

We use WEKA for classification and Regression. We use SVM for classification for RTE, and Additive Regression for regression for STS. The script bin/weka.sh is to make this task easy.

The script bin/weka.sh can be called like this:

~/pl-semantics$ cat ACTUAL | bin/weka.sh COMMAND GoldStandardFile NumberOfColumns

where

• ACTUAL: is a file contains the system's output score in one long line.

• COMMAND: "classify" or "regress" depending on the task

• GoldStandardFile: the name is descriptive enough. A sample file is resources/sick/sick-rte.gs

• NumberOfColumns[optional, defaul = 1]: number of columns in ACTUAL. It is useful in case of STS which outputs two columns, and in multiple parses which outout squared numbers of parses. This argument is optional, and its default value is 1.

Example: after running the system on Condor, you can read the output using the condor script with argument "collect" as shown before. One easy way to get the classification/regression score is by piping the output from the condor scrip to the weka script as below:

~/pl-semantics$ bin/condor collect condor/firstExp  | tail -n 1  | bin/weka.sh regress resources/sick/sick-sts.gs  1 

~/pl-semantics$ bin/condor submit condor/firstExp 2 sick-rte //or could be " 10 sick-sts" for the sts dataset. No need to select timeout, task, inference tool, nor the log level, the appropriate default values are automatically picked for each dataset

~/pl-semantics$ bin/condor status //make sure all tasks are done

~/pl-semantics$ bin/condor collect condor/firstExp //collect and report number of errors

~/pl-semantics$ bin/condor fix condor/firstEx //fix in case collects report errors

~/pl-semantics$ bin/condor collect condor/firstExp //collect again after fixing it done

~/pl-semantics$ bin/condor eval condor/firstEx //use output from collect to call weka for regression or classification depending on the dataset

~/pl-semantics$ bin/mlnsem boxer OPTIONS

You can parse a single sentence with the -s option:

~/pl-semantics$ bin/mlnsem boxer -s "A dog walks." [OPTIONS]

or an entire file of sentences (one sentence per line):

~/pl-semantics$ bin/mlnsem boxer -f sentences.txt [OPTIONS]

Output options:

-draw true/false (default: true)        Prints a graphical DRT representation
-boxer true/false (default: true)       Prints logical form in Boxer notation
-drt true/false (default: true)         Prints the logical form as a DRS
-fol true/false (default: true)         Prints the logical form in first-order logic

Regarding tokenization: The code will automatically tokenize all input sentences, so it does not matter if the input is given tokenized or not. It will not do sentence-splitting, however, but this can be included if it would be useful.

• git clone git@github.com:raptros/tr-corpus-one.git

• Start the soap server with the correct arguments - $CANDC_HOME/bin/soap_server --models $CANDC_HOME/models --server localhost:12200 --candc-printer boxer will do the job

• ./run local trc1.MaxTransform data/rules.in data/sentences.txt data/out

• Experimented with prior -1.0

• In Sts.scala:

Employ Lucene for indexing and searching inference rules from external resources. See /utcompling/mlnsemantics/util/Lucene.scala for more details. Each rule has the following format: TAB TAB TAB <entail_score>

Put type constraints for variables via FindEventsProbabilisticTheoremProver class.

• In AlchemyTheoremProver.scala:

Function createSamePredRule generates inference rules for each pair of predicates having the same name. E.g., man_dt(indv0) => man_dh(indv0). This function also employs WordNet to find similar words.

Function makeEvidenceFile removes all evidence predicates that don't appear in inference rules. This can help to reduce domain size.

Function makeMlnFile removes all theme relation predicates and equality constraints in Hypothesis because there are no inference rules for them. To do that easier, I remove inner parentheses of conjunction expressions (see funtion _convert).

• In InferenceRuleInjectingProbabilisticTheoremProver.scala:

Function checkCompatibleType checks compatibility between lhs and rhs of an inference rule.

Function convertParaphraseToFOL converts paraphrase rules in text format to FOL.

Function getAllRelations also gets "agent" and "patient" relations.

• In MergeSameVarPredProbabilisticTheoremProver.scala:

Function getAllPreds puts entity type for BoxerNamed.

Do not merge same variable predicates.

• In HardAssumptionAsEvidenceProbabilisticTheoremProver.scala:

Change renameVars.

• Remove ("--candc-int-betas" -> "0.00075 0.0003 0.0001 0.00005 0.00001") parameter from /utcompling/mlnsemantics/datagen/CncLemmatizeCorpus.scala and /utcompling/scalalogic/discourse/impl/BoxerDiscourseInterpreter.scala because we get more parse fails with it.

• In utcompling/scalalogic/discourse/candc/boxer/expression/interpreter/impl/OccurrenceMarkingBoxerExpressionInterpreterDecorator.scala:

Get positions of words in sentence. Words with same name will be treated differently based on their position.

• Parser errors:

• Wrong POS tags. Several verbs are classified as nouns.
• Wrong semantic representations.

• Problem with negation.

• Problem with directional distributional similarity.

• Problem with singular and plural nouns.

• No vector representation for sentences "A woman fries eggs" and "A woman fries big eggs" in group 5.