TCL - a Lexicon of Turkish Discourse Connectives

Deniz Zeyrek, Kezban BaşıbüyükACL 2019

TL;DR

This paper introduces TCL, the first comprehensive lexicon of Turkish discourse connectives, built automatically using discourse-annotated corpora. This resource is crucial for advancing Turkish natural language processing, discourse analysis, and cross-linguistic studies by providing a structured representation of connectives' properties and meanings.

What Problem Does This Paper Address?

Discourse connectives (like "but," "because," "therefore") are vital for understanding how sentences and ideas link together in text, forming coherent discourse. While many languages have resources detailing these connectives, Turkish presents unique challenges due to its agglutinative nature, allowing for complex suffixal connectives alongside traditional lexical ones. Existing dictionaries are insufficient for capturing the nuances of Turkish discourse relations. This paper aims to fill this gap by creating a lexicon that details Turkish discourse connectives, their various forms, syntactic roles, and semantic senses, making them more accessible for computational linguistics and linguistic research.

Methods

The authors developed a lexicon, TCL (Turkish Discourse Connective Lexicon), following the structure of the German DiMLex. The core of their methodology involved:

  • Automatic Lexicon Construction: A custom search tool was developed to automatically extract discourse connectives and their properties from annotated corpora.
  • Data Integration: Information was gathered from three PDTB- (Penn Discourse Treebank) inspired Turkish corpora: TDB 1.0, TDB 1.1, and the Turkish section of TED-MDB. This multi-corpus approach ensures broader coverage of connectives and their senses.
  • Syntactic Categorization: A multi-source approach was used to determine the syntactic category of connectives, leveraging tools like UDPipe, the Turkish Linguistic Society website, TDK, and TRmorph. This was particularly important for handling ambiguous cases.
  • Suffixal Connective Handling: A specific method was devised to identify and represent suffixal connectives (converbs), accounting for their complex allomorphy based on vowel and consonant harmony.
  • DiMLex Formatting: The final lexicon entries were formatted according to the DiMLex XML standard, facilitating integration into a larger multilingual connective lexicon database.

Datasets

The following PDTB-inspired Turkish discourse-annotated corpora were used:

  • TDB 1.0 (Turkish Discourse Bank 1.0): A 400,000-word corpus of modern written Turkish annotated for explicit discourse connectives and their relations.
  • TDB 1.1 (Turkish Discourse Bank 1.1): A 40,000-word subset of TDB 1.0, annotated with all five PDTB relation types, including implicit relations and alternative lexicalizations.
  • TED-MDB (TED Multilingual Discourse Bank): The Turkish section of this corpus, containing TED talk transcripts.

Key Insights and Ideas

  • Turkish Connective Diversity: Turkish employs a rich set of discourse connectives, including not only single words but also complex suffixal forms (converbs) and phrasal expressions, making a comprehensive lexicon essential.
  • Importance of Multi-Corpus Approach: Using multiple corpora (TDB 1.0, TDB 1.1, TED-MDB) was crucial for capturing the full range of connectives and their various senses, as some were present in only one source.
  • Handling Suffixal Connectives: The paper highlights the challenge and necessity of explicitly representing the allomorphy of suffixal connectives, which are a hallmark of Turkish.
  • PDTB-Style Annotation for Turkish: The successful extension of PDTB-style annotation to Turkish demonstrates its applicability to morphologically rich languages, paving the way for more consistent cross-linguistic discourse analysis.
  • Lexicon as a Bridge: The TCL lexicon serves as a vital bridge between raw annotated corpora and practical applications in NLP, translation studies, and language learning.

Key Results

The paper successfully developed TCL, a lexicon containing 180 Turkish discourse connectives. For each connective, the lexicon details:

  • Orthographical variants (single/phrasal, continuous/discontinuous).
  • Canonical form.
  • Frequency of occurrence.
  • Syntactic category (e.g., CCONJ, Converb, ADP, ADV).
  • Semantic senses, mapped to PDTB-style relation types.
  • Representative examples from the corpora.

The sense distribution analysis showed that while the overall distribution of top-level sense classes in Turkish corpora (Expansion, Contingency, Comparison, Temporal) largely mirrors that of the PDTB 2.0, there are notable differences in proportions, particularly in the prevalence of Expansion and Contingency in Turkish.

Limitations and Future Work

  • Genre Sensitivity: The authors acknowledge that discourse connectives can be genre-sensitive. The current TCL is populated from texts of various genres, but future work will involve computing and incorporating genre-specific sense distributions into DiMLex.
  • Extending the Search Tool: The authors plan to extend the TCL search tool with new functionalities to further enhance its utility.
  • Multilingual Integration: The ultimate goal is to synchronize TCL with other connectives lexicons within the connective-lex.info database, fostering cross-linguistic research.

Summary updated

BibTeX
@inproceedings{zeyrek2019tcl,
  author = {Deniz Zeyrek and Kezban Başıbüyük},
  title = {{TCL - a Lexicon of Turkish Discourse Connectives}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
}
Figure