[ Skip to the content ]

Institute of Formal and Applied Linguistics Wiki


[ Back to the navigation ]

This is an old revision of the document!


Table of Contents

TectoMT Tutorial

Welcome at TectoMT Tutorial. This tutorial should take about 2 hours.

What is TectoMT

TectoMT is a highly modular NLP (Natural Language Processing) software system implemented in Perl programming language under Linux. It is primarily aimed at Machine Translation, making use of the ideas and technology created during the Prague Dependency Treebank project. At the same time, it is also hoped to significantly facilitate and accelerate development of software solutions of many other NLP tasks, especially due to re-usability of the numerous integrated processing modules (called blocks), which are equipped with uniform object-oriented interfaces.

Prerequisities

In this tutorial, we assume

Installation and setup

    cd ~/BIG
    svn --username <username> co https://svn.ms.mff.cuni.cz/svn/tectomt_devel/trunk tectomt
    cd tectomt/install
    ./install.sh
    source ~/BIG/tectomt/config/init_devel_environ.sh

TectoMT Architecture

Blocks, scenarios and applications

In TectoMT, there is the following hierarchy of processing units (software components that process data):

This tutorial itself has its blocks in libs/blocks/Tutorial and the application in applications/tutorial.

Layers of Linguistic Structures

MT pyramid in terms of PDT layers

The notion of 'layer' has a combinatorial nature in TectoMT. It corresponds not only to the layer of language description as used e.g. in the Prague Dependency Treebank, but it is also specific for a given language (e.g., possible values of morphological tags are typically different for different languages) and even for how the data on the given layer were created (whether by analysis from the lower layer or by synthesis/transfer).

Thus, the set of TectoMT layers is a Cartesian product {S,T} x {English,Czech,…} x {W,M,P,A,T}, in which:

Blocks in block repository libs/blocks are located in directories indicating their purpose in machine translation.

Example: Block adding Czech morphological tags (pos, case, gender, etc.) can be found in libs/blocks/SCzechW_to_SCzechM/Simple_tagger.pm.

There are also other directories for other purpose blocks, for example blocks which only print out some information go to libs/Print. Our tutorial blocks are in libs/blocks/Tutorial/.

First application

Once you have TectoMT installed on your machine, you can find this tutorial in applications/tutorial/. After you cd into this directory, you can see our plain text sample data in sample.txt.

Most applications are defined in Makefiles, which describe sequence of blocks to be applied on our data. In our particular Makefile, four blocks are going to be applied on our sample text: sentence segmentation, tokenization, tagging and lemmatization. Since we have our input text in plain text format, the file is going to be converted into tmt format beforehand (the in target in the Makefile).

We can run the application:

make all

Our plain text data sample.txt have been transformed into tmt, an internal TectoMT format, and saved into sample.tmt. Then, all four blocks have been loaded and our data has been processed. We can now examine sample.tmt with a text editor (vi, emacs, etc).

Changing the scenario

We'll now add a syntax analysis (dependency parsing) to our scenario by adding three more blocks. Instead of

analyze:
        brunblocks -S -o \
                SEnglishW_to_SEnglishM::Sentence_segmentation_simple \
                SEnglishW_to_SEnglishM::Penn_style_tokenization \
                SEnglishW_to_SEnglishM::TagMxPost \
                SEnglishW_to_SEnglishM::Lemmatize_mtree \
        -- sample.tmt

we'll have:

analyze:
        brunblocks -S -o \
                SEnglishW_to_SEnglishM::Sentence_segmentation_simple \
                SEnglishW_to_SEnglishM::Penn_style_tokenization \
                SEnglishW_to_SEnglishM::TagTnT \
                SEnglishW_to_SEnglishM::Lemmatize_mtree \
                SEnglishM_to_SEnglishA::McD_parser_local \
                SEnglishM_to_SEnglishA::Fix_McD_Tree \
                SEnglishM_to_SEnglishA::Fill_afun_after_McD \
        -- sample.tmt

Note: Makefiles use tabulators to mark command lines. Make sure your lines start with tabulator (or two tabulators) and not, for example, with 4 spaces.

After running

make all

we can examine our sample.tmt again. Really, an analytical layer SEnglishA describing a dependency tree with analytical functions (<afun>) has been added to each bundle.

You can view the trees in sample.tmt with TrEd by typing

tmttred sample.tmt

Adding a new block

The linguistic structures in TectoMT are represented using the following object-oriented interface/types:

You can get TectoMT automatically execute your block code on each document or bundle by defining the main block entry point:

Each block must have exactly one entry point.

We'll now examine an example of a new block in file libs/blocks/Tutorial/Print_node_info.pm.

This block illustrates some of the most common methods for accessing objects:

Attributes of documents, bundles or nodes can be accessed by attribute getters and setters, for example:

Our tutorial block Print_node_info.pm is ready to use. You only need to add this block to our scenario, e.g. as a new Makefile target:

print_info:
        brunblocks -S -o Tutorial::Print_node_info -- sample.tmt

We can observe our new block behaviour:

make print_info

Try to change the block so that it prints out the information only for verbs. (You need to look at an attribute tag at the m level). The tagset used is Penn Treebank Tagset.

Advanced block: finite clauses

Motivation

It is assumed that finite clauses can be translated independently, which would reduce computational complexity or make parallel translation possible. We could even use hybrid translation - each finite clause could be translated by the most self-confident translation system. In this task, we are going to split the sentence into finite clauses.

Task

A block which, given an analytical tree (SEnglishA), fills each a-node with boolean attribute is_clause_head which is set to 1 if the a-node corresponds to a finite verb, and to 0 otherwise.

Instructions

There is a block template with hints in libs/blocks/Tutorial/Mark_heads.pm. You should edit the block so that the output of this block is the same a-tree, in addition with attribute is_clause_head attached to each a-node. There is also a printing block libs/blocks/Print_finite_clauses.pm which will print out the a-nodes grouped by clauses:

finite_clauses:
        brunblocks -S -o \
                Tutorial::Mark_heads \
                Tutorial::Print_finite_clauses \
        -- sample.tmt

You are going to need these methods:

Note: get_children() returns topological node children in a tree, while get_eff_children() returns node children in a linguistic sense. Mostly, these do not differ.

Advanced version: The output of our block might still be incorrect in special cases - we don't solve coordination (see the second sentence in sample.txt) and subordinate conjunctions.

Your turn: more tasks

SVO to SOV

Motivation: During translation from an SVO based language (e.g. English) to an SOV based language (e.g. Korean) we might need to change the word order from SVO to SOV.

Task: Change the word order from SVO to SOV.

Instructions:

Prepositions

Prepositions example

Motivation: In dependency approach a question “where to hang prepositions” arises. In praguian style (PDT), prepositions are heads of the subtree and the noun/pronoun is dependent on the preposition. However, another ordering might be preferable: The noun/pronoun might be the head of subtree, while the preposition would take the role of a modifier.

Task: The task is to rehang all prepositions as indicated at the picture. You may assume that prepositions have at most 1 child.

Instructions:

You are going to need these new methods:

Hint:

Advanced version: What happens in case of multiword prepositions? For example, because of, instead of. Can you handle it?

Further information


[ Back to the navigation ] [ Back to the content ]