Constructing a Turkish-English Parallel TreeBank

Turkish is an agglutinative and morphologically rich language with a free constituent order. Although statistical NLP research on Turkish has taken significant steps in recent years, much remains to be done. Especially for the annotated corpora, Turkish is still behind similar languages such as Czech, Finnish, or Hungarian. For example, EuroParl corpus [13], one of the biggest parallel corpora in statistical machine translation, contains 22 languages (but not Turkish). Although there exist some recent works to produce parallel corpora for Turkish-English pair, the produced corpus is only applicable for phrase-based training [22, 8].

In recent years, many efforts have been made to annotate parallel corpora with syntactic structure to build parallel treebanks. A parallel treebank is a parallel corpus where the sentences in each language are syntactically (if necessary morphologically) annotated, and the sentences and words are aligned. In the parallel treebanks, the syntactic annotation usually follows constituent and/or dependency structure. Well-known parallel treebank efforts are

• •

  Prague Czech-English dependency treebank annotated with dependency structure [6]

• •

  English-German parallel treebank, annotated with POS, constituent structures, functional relations, and predicate-argument structures [7]

• •

  Linköping English-Swedish parallel treebank that contains 1,200 sentences annotated with POS and dependency structures [1]

• •

  Stockholm multilingual treebank that contains 1,000 sentences in English, German and Swedish annotated with constituent structure [12]

In this study, we report our preliminary efforts in constructing an English-Turkish parallel treebank corpus for statistical machine translation. Our approach converts English parse trees into equivalent Turkish parse trees by applying several transformation heuristics. The main components of our strategy are (i) tree permutation, where we permute the children of a node; and (ii) leaf replacement, where we replace English word token at a leaf node.

This paper is organized as follows: In Section 2, we give the literature review for parallel treebank construction efforts in Turkish. In Section 3, we give a very brief overview on Turkish syntax. We give the details of our corpus construction strategy in Section 4 and explain our transformation heuristics in Section 5. Finally, we conclude in Section 6.

Turkish Treebank creation efforts started with the METU-Sabancı dependency Treebank. METU-Sabancı Treebank explicitly represents the head-dependent relations and functional categories. In order to adapt the corpus written in 1990’s Turkish to further studies, a subset of 7.262 sentences of the corpus was manually annotated morphologically and syntactically [2]. METU-Sabancı Treebank is then used in many Turkish NLP studies [11, 23, 19, 20, 9, 10].

METU-Sabancı Treebank is also subject to transformation efforts from dependency-structure to constituency-structure. Combinatory Categorical Grammar (CCG) is extracted from the METU-Sabancı Treebank with annotation of lexical categories [3]. Sub-lexical units revealing the internal structure of the words are used to generate a Lexical Grammar Formalism (LGF) for Turkish with the help of finite state machines [4, 5].

Swedish-Turkish parallel treebank is the first parallel Treebank effort for Turkish [16]. The treebank is a balanced syntactically annotated corpus containing both fiction and technical documents. In total, it consists of approximately 160,000 tokens in Swedish and 145,000 in Turkish. Parallel texts are linguistically annotated using different layers from part of speech tags and morphological features to dependency annotation.

English-Swedish-Turkish parallel treebank [15], mainly the successor of the Swedish-Turkish parallel treebank, consists of approximately 300,000 tokens in Swedish, 160,000 in Turkish and 150,000 in English. The majority of the original text is written in Swedish and translated to Turkish and/or English. For the syntactic description, dependency structure is chosen instead of the constituent structure. All data is automatically annotated with syntactic tags using MaltParser [17]. MaltParser is trained on the Penn Treebank for English, on the Swedish treebank Talbanken05 [18], and on the METU-Sabancı Turkish Treebank [2], respectively.

ParGram parallel treebank [21] is a joint effort for the construction of a parallel treebank involving ten languages (English, Georgian, German, Hungarian, Indonesian, Norwegian, Polish, Turkish, Urdu, Wolof) from six language families. The treebank is based on deep Lexical-Functional Grammars that were developed within the framework of the Parallel Grammar effort. ParGram treebank allows for the alignment of sentences at several levels: dependency structures, constituency structures and POS information.

Turkish is an agglutinative language with rich derivational and inflectional morphology through suffixes. Word forms usually have a complex yet fairly regular morphotactics.

Turkish sentences have an unmarked SOV order. However, depending on the discourse, constituents can be scrambled to emphasize, topicalize and focus certain elements. Case markings identify the syntactic functions of the constituents, [14].

In order to constrain the syntactic complexity of the sentences in the corpus, we selected from the Penn Treebank II 9560 trees which contain a maximum of 15 tokens. These include 8660 trees from the training set of the Penn Treebank, 360 trees from its development set and 540 trees from its test set. In the first phase of our work, we translated 4247 trees of the training set and all of those in the development and the test sets.

When we have a sufficiently rich corpus of parallel trees, our next step is to train a SMT learner to imitate the human translator who operates under our restricted set of operations. Naturally, human translators often base their transformation decisions on the whole tree. Still, having a common set of rules and heuristics helps the translators in both consistency and speed. In the following, we illustrate these heuristics.

Parallel treebank construction efforts increased significantly in the recent years. Many parallel treebanks are produced to build statistically strong language models for different languages. In this study, we report our preliminary efforts to build such a parallel corpus for Turkish-English pair. We translated and transformed a subset of parse trees of Penn Treebank to Turkish. We cover more than 50% of all sentences with a maximum length of 15-words including punctuation.

This work constitutes the preliminary step of parallel treebank generation. As a next step, we will focus on morphological analysis and disambiguation of Turkish words. After determining the correct morphological analysis of Turkish words, we will use the parts of these analyses to replace the leaf nodes that we intentionally left as “*NONE*”. As a future work, we plan to expand the dataset to include all Penn Treebank sentences.