Welcome at TectoMT Tutorial. This tutorial should take about 3 hours.

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 facilitate and significantly 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.

In this tutorial, we assume

• Your system is Linux
• Your shell is bash
• You have basic experience with bash and can read basic Perl

• Checkout SVN repository. If you are running this installation in computer lab in Prague, you have to checkout the repository into directory /home/BIG (because data quotas don't apply here):

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

• In tectomt/install/ run ./install.sh:

    cd tectomt/install
    ./install.sh

• In your .bashrc file, add line (or source this file every time before experimenting with TectoMT):

    source ~/BIG/tectomt/config/init_devel_environ.sh

• In your .bash_profile file, add line

    source .bashrc

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

• The basic units are blocks. They serve for some very limited, well defined, and often linguistically interpretable tasks (e.g., tokenization, tagging, parsing). Technically, blocks are Perl classes inherited from TectoMT::Block, each saved in a separate file. The blocks repository is in libs/blocks/.
• To solve a more complex task, selected blocks can be chained into a block sequence, called also a scenario. Technically, scenarios are instances of TectoMT::Scenario class, but in some situations (e.g. on the command line) it is sufficient to specify the scenario simply by listing block names separated with spaces.
• The highest unit is called application. Applications correspond to end-to-end tasks, be they real end-user applications (such as machine translation), or 'only' NLP-related experiments. Technically, applications are often implemented as Makefiles, which only glue the components existing in TectoMT. Some demo applications can be found in applications.

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

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:

• {S,T} distinguishes whether the data was created by analysis or transfer/synthesis (mnemonics: S and T correspond to (S)ource and (T)arget in MT perspective).
• {English,Czech…} represents the language in question
• {W,M,P,A,T…} represents the layer of description in terms of PDT 2.0 (W - word layer, M - morphological layer, A - analytical layer, T - tectogrammatical layer) or extensions (P - phrase-structure layer).

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/.

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).

• One physical tmt file corresponds to one document.
• A document consists of a sequence of bundles (<bundle>), mirroring a sequence of natural language sentences originating from the text. So, for one sentence we have one <bundle>.
• Each bundle contains tree shaped sentence representations on various linguistic layers. In our example sample.tmt we have morphological tree (SEnglishM) in each bundle. Later on, also an analytical layer (SEnglishA) will appear in each bundle as we proceed with our analysis.
• Trees are formed by nodes and edges. Attributes can be attached only to nodes. Edge's attributes must be equivalently stored as the lower node's attributes. Tree's attributes must be stored as attributes of the root node.

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::TagMxPost \
                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.

Blocks can also be parametrized. For syntax parser, we might want to use a smaller but faster model. To achieve this, replace the line

SEnglishM_to_SEnglishA::McD_parser_local \

with

SEnglishM_to_SEnglishA::McD_parser_local MT_PARAM_MCD_EN_MODEL=conll_mcd_order2_0.1.model \

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

tmttred sample.tmt

Try to click on some nodes to see their parameters (tag, lemma, form, analytical function etc).

Note: For more information about tree editor TrEd, see TrEd User's Manual.

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

• document - TectoMT::Document
• bundle - TectoMT::Bundle
• node - TectoMT::Node

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

• sub process_document - run this procedure on each document
• sub process_bundle - run this procedure on each bundle (sentence)

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:

• my @bundles = $document->get_bundles() - an array of bundles contained in the document
• my $root_node = $bundle->get_tree($layer_name) - the root node of the tree of the given type in the given bundle
• my @children = $node->get_children() - array of the node's children
• my @descendants = $node->get_descendants() - array of the node's children and their children and children of their children …
• my $parent = $node->get_parent() - parent node of the given node, or undef for root
• my $root_node = $node->get_root() - the root node of the tree into which the node belongs

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

• $node->get_attr($attr_name)
• $node->set_attr($attr_name, $attr_value)

Some interesting attributes on morphologic layer are form, lemma and tag. Some interesting attributes on analytical layer are afun (analytical function) and ord (surface word order). To reach form, lemma or tag from analytical layer, that is, when calling this attribute on an a-node, you use $a_node→get_attr('m/form') and the same way for lemma and tag. The easiest way to see the node attributes is to click on the node in TrEd:

tmttred sample.tmt

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.

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.

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.

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:

• my $root = $bundle->get_tree('tree_name')
• my $attr = $node->get_attr('attr_name')
• $node->set_attr('attr_name',$attr_value)
• my @eff_children = $node->get_eff_children()

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. If interested, see Figure 1 in btred tutorial.

Hint: Finite clauses in English usually require grammatical subject to be present.

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.

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:

• You can use block template in libs/blocks/BlockTemplate.pm.
• To find an object to a verb, look for objects among effective children of a verb ($child->get_attr('afun') eq 'Obj' ). That implies working on analytical layer.
• For debugging, a method returning surface word order of a node is useful: $node->get_attr('ord'). It can be used to print out nodes sorted by attribute ord.
• Once you have node $object and node $verb, use method $object->shift_before_node($verb). This method takes the whole subtree under node $object and re-counts the attributes ord (surface word order) so that all nodes in subtree under $object have smaller ord than $verb. That is, the method rearranges the surface word order from VO to OV.

Advanced version: This solution shifts object (or more objects) of a verb just in front of that verb node. So f.e.: Mr. Brown has urged MPs. changes to: Mr. Brown has MPs urged. You can try to change this solution, so the final sentence would be: Mr. Brown MPs has urged. You may need a method $node→shift_after_subtree($root_of_that_subtree). Subjects should have attribute 'afun' eq 'Sb'.

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:

• my @children = $node->get_children()
• my $parent = $node->get_parent()
• $node->set_parent($parent)

Hint:

• On analytical layer, you can use this test to recognize prepositions: $node->get_attr('afun') eq 'AuxP'
• To see the results, you can again use TrEd (tmttred sample.tmt)

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

• TectoMT Developer's Guide
• Questions? Ask kravalova at ufal.mff.cuni.cz
• Solutions to this tutorial tasks are in libs/blocks/Tutorial/*solution*.pm.
• TrEd, TrEd User's Manual - tree editor