Author: Joseph Le Roux <joseph.le.roux@gmail.com> Co-author: Corentin Ribeyre <corentin.ribeyre@gmail.com> Date: Fri Sep 13 15:13:04 2013

This is the README for lorg tools written at http://www.nclt.dcu.ie Questions, bug reports, feedback on the software can be logged via github We’ll do our best to get back to you but we won’t make any promises.

lorg_tools consist of 4 programs:

1. twostage_lorgparser: a PCFG-LA parser implementing the CKY algorithm and a coarse-to-fine strategy.
2. tb2gram: a trainer for PCFG-LAs
3. simple_lorgparser: a simple PCFG parser implementing the CKY algorithm (deprecated)
4. feature_extractor: a feature extractor from parse trees (experimental)

This software is free to use in non-commercial projects.

If you use these tools in support of an academic publication, please cite the paper [1].

1. You need boost – >= 1.46 – available at boost page or from your favorite package manager.
2. You may use Google malloc instead of your system implementation. It might speed up memory allocation. Autotools will automatically try to detect the availability of this library and use it, provided it is usable.
3. Intel Threading Building Blocks available at tbb page or from your favorite package manager

autoreconf –install && ./configure && make && make install

NB: You may need administrator privileges to install lorg tools system-wide

This program reads a treebank (in PTB format) and create a PCFG-LA, in a text format readable by the two-stage parser (cf. infra). It implements split-merge EM learning algorithm for PCFG_LAs. See [2] for more details on the algorithm.

1. For the impatient :

   tb2gram treebankfile_1 … treebankfile_n -o output_prefix

2. More information : tb2gram -h

   Various parameters can be changed and will give very different results. In particular:

   random-seed

   change the random seed for EM initialisation. It will help getting different grammars with the same parameters (useful if you want to test the relevance of your new parameters, for example a new tag set).

unknown-word-mapping

how rare words are classified during training to handle unknown words at parsing. See [1] for details. This parameter can be set to:

• generic : is the simplest, and will work for all languages
• BerkeleyEnglish : use the signatures from the Berkeley Parser (English)
• BaselineFrench : use the signatures from the Berkeley Parser (French)
• EnglishIG : use the signatures collected on PTB ranked by information gain (English). See [1] for more information.
• FrenchIG :use the signatures collected on FTB ranked by information gain (French)
• Arabic : use morphotactic signatures (Arabic)
• ArabicIG : use morphotactics and information gain (Arabic)
• ItalianIG : signatures collected from Italian corpora (experimental)

unknown-word-cutoff

cut-off for classifying rare words. Default is 5. Again see [1] for details

nbthreads

Number of threads. Default is 1 (seems to be useless with recent tbb implementations)

remove-functions

Remove grammatical functions from node names in the treebank before learning the grammar. Default is 1 (set).

hm and vm

set the horizontal (hm) an vertical (vm) markovization for rule extraction. Defaults are 1 for vm (node names are not modified) and 0 for hm (1 intermediate symbol by original symbol)

1. Examples

   a. Generating a grammar for English

   tb2gram $WSJ/0[2-9] $WSJ/1* $WSJ/2[01] -v –nbthreads 8 -w EnglishIG -u 1 -o english_grammar

   This will recursively read sections 02-21 for training data, be verbose, use (by default) 6 split/merge/smooth steps, 8 threads, the automatically acquired signatures for English, replace tokens occurring once in the training data with their signatures. Please note that it will also use the default markovization settings.

This program takes as input a grammar and text and outputs parse trees for this text. It implements a coarse-to-fine CKY PCFG-LA parser.

1. for the impatient :

   twostage_lorgparser input -g grammar_file -o output

   where the grammar file was created by the trainer, and input is a file of sentences (one per line).

   If input is not set, the parser will read from standard input and, accordingly, if output is not set the parser will write on standard output.

2. more information :

   twostage_lorgparser -h

3. You should use the same signatures (option –unknown-word-mapping) as in training. The parser will output a warning if this is not the case.
4. These are the parameters that you may want to change: unknown-word-mapping: should have the same value for training and parsing. beam-threshold: the probability threshold used for chart construction.

   stubbornness

   number of parse attempts with increasing lower beam-thresholds. The last attempt is performed without beam (potentially leading to huge forests). A negative value disables this feature.

   accurate

   a finer set of thresholds for the coarse-to-fine solution extraction (experimental)

   parser-type

   the algorithm used for solution extraction

   vit

   Viterbi (PCFG approximation)

   kmax

   k-best MaxRule output a list of solutions of length k. Use the –k to set the length of the list and –verbose to display solution scores.

   max

   MaxRule. This is equivalent to kmax with k set to 1 but note that it is more efficient. This is the default setting.

   maxn

   MaxRule with several grammars. If you use this parsing method, the command line would be something like:

   twostage_lorgparser input -g grammar_file -o output -F othergram_1 … -F othergram_n

   input-type

   input format (see section on [sentence files] )

   1. Example

      twostage_lorgparser wsj23.tagged -g english_grammar_smoothed6 -o wsj3.parsed -w EnglishIG –input-type tag –parser-type kmax –k 20

      This will parse the file wsj23.tagged, assuming that it is in the “tag” format (see below), using the grammar english_grammar_smoothed6, the maxrule algorithm, returning the 20 best parses for each sentence.

      [sentence files]: sec-5-2

(deprecated)

===================

This script will remove extra comment lines from the parser’s output to conform to evalb format.

If your version of boost is not installed system-wide or more generally if it is installed in a non-standard directory dir, be sure to add dir in your LD_LIBRARY_PATH. For example, add in your .bashrc: export LD_LIBRARY_PATH=dir:$LD_LIBRARY_PATH

If you add signatures for a new language, and you feel like new versions of lorg tools should have these signatures, please contact us via github.

We present the formats for the different types of files used by our software

Treebanks should be in PTB format. Trees can be on several lines. Encoding must be UTF-8.

These files must contain one sentence by line. There are 3 different types of input :

raw

The parser does its own tokenizing. This feature is only implememented for English and is highly experimental.

tok

The input is tokenized as in the corresponding treebank.

tag

The input is tokenized and each token has a list of predicted pos tags. Input looks like:

tok_1 ( TAG_11 … TAG_1n ) … tok_m ( TAG_m1 … TAG_mp )

Caution: Spaces before and after parentheses are mandatory.

For the last 2 types of input, strings “[“, “[[” , “]” and “]]” have a special meaning. There are treated like chunk delimiters with the following semantics. “[” or “[” mark the beginning of a chunk, while “]” or “]” mark the end of a chunk. Double symbols indicate strong frontiers while simple ones refer to weak frontiers.

IMPORTANT: –> You should escape the input parentheses if they have no special meaning –> You should escape the squared parentheses in the input and in the training set, if not

lat

The input is a lattice (or a dag, or an acyclic automaton). Each line is an edge of the form:

begin_postion end_position word [ optional list of postags ]

Sentences are separated by an empty line. This returns the best tree over the lattice, so it chooses one path. This might be useful to parse the output of a speech recognition system but can also be used for other purposes (for example, see [3]) This is still an experimental feature.

We have chosen text format over binary format, so grammars can be more easily amended using scripts. On the other hand, this makes our grammars quite large on disk. An annotated grammar is made of annotation information followed by annotation histories, followed by grammar rules, followed by lexical rules.

Grammar ::= Comment* AnnotationHistory+ InternalRule+ LexicalRule+ AnnotationHistory ::= NT Tree[integer]

• Comments starts with //
• An AnnotationInfo gives the number of annotations for a non-terminal symbol in the grammar.
• An AnnotationHistory gives the history of splits for a non-terminal symbol.
• In the future Annotation Histories and Annotation Information will be unified.
• NT is a string

There are 2 kinds of rules, lexical rules and internal rules, the latter being divided in binary and unary internal rules. See the following table for a BNF desciption of the format, where annotation is of integer type and probability of floating point type.

InternalRule ::= BinaryRule \/ UnaryRule BinaryRule ::= “int” NT NT NT binary_probability+ UnaryRule ::= “int” NT NT unary_probability+ binary_probability ::= (annotation,annotation,annotation,probability) unary_probability ::= (annotation,annotation,probability) LexicalRule ::= “lex” NT word lexical_probability+ lexical_probability ::= (annotation,probability)

[1] “Handling Unknown Words in Statistical Latent-Variable Parsing Models for Arabic, English and French”, Mohammed Attia, Jennifer Foster, Deirdre Hogan, Joseph Le Roux, Lamia Tounsi and Josef van Genabith, Proceedings of SPMRL 2010.

[2] “Improved Inference for Unlexicalized Parsing”, Slav Petrov and Dan Klein, HLT-NAACL 2007

[3] “Language-Independent Parsing with Empty Elements”, Shu Cai, David Chiang and Yoav Goldberg, ACL-2011 (Short Paper)

The following people have worked on the LORG parser: Joseph Le Roux, Deirdre Hogan, Jennifer Foster, Corentin Ribeyre, Lamia Tounsi and Wolfgang Seeker, Antoine Rozenknop.