Proceedings of the First Workshop on Linguistic Resources for Natural Language Processing

Peter Machonis, Anabela Barreiro, Kristina Kocijan, Max SilberzteinCOLING 2018

TL;DR

This document presents the proceedings of the First Workshop on Linguistic Resources for Natural Language Processing (LR4NLP) held at COLING 2018. It showcases research on building and utilizing large-scale linguistic resources, bridging the gap between traditional linguistics and modern NLP applications, and highlights advancements in areas like semantic textual similarity, multilingual corpora, and speech recognition.

What Problem Does This Paper Address?

Historically, linguists and NLP software developers have operated in separate spheres. Linguists focused on creating comprehensive, application-agnostic linguistic resources, while NLP researchers concentrated on technical aspects of methods like statistical and neural network parsers, emphasizing recall, precision, and large training corpora. However, recent advancements have shown the increasing value of large-coverage linguistic resources for robust NLP software, leading to a convergence of these fields. This workshop aimed to bring together researchers from both communities to discuss the development and application of such resources.

Methods

The workshop featured a variety of methods and approaches across its sessions:

  • Linguistics vs. Statistics vs. Neural Networks: This session explored the interplay between linguistic knowledge, statistical methods, and neural networks. Techniques included semi-automatic analysis of speech collections to inform phonetics and parsing, evaluation of reference corpora, and comparative assessments of rule-based versus neural net systems for Semantic Textual Similarity.
  • NooJ Platform: Several presentations focused on the NooJ software, a platform used for developing linguistic resources like electronic dictionaries, morphological and syntactic grammars, and for parsing and annotating corpora. Methods involved transforming lexicon grammar dictionaries into NooJ formats and using NooJ to detect and describe linguistic phenomena (e.g., verb aspectual derivatives).
  • Monolingual Resources: This session covered the creation and utilization of large-scale monolingual resources. Methods included systematic derivation of language usage patterns from lexical-semantic representations, automatic morpho-syntactic tagging of corpora, and novel summarization techniques like "Telegraphic Summarization" that select text segments across sentences.
  • Multilingual Resources: Research in this session focused on building multilingual linguistic resources for Machine Translation (MT). Approaches involved developing parallel corpora for various language pairs, using embedding techniques to compare semantic frames across languages, constructing multilingual corpora annotated with translation relations, and leveraging OCR for automatic classification of dictionary examples.

Datasets

The workshop papers utilized and discussed a range of datasets and corpora:

  • Open American National Corpus: Used for evaluating reference corpora.
  • Large phrase-structure grammars: Mentioned as a resource for text parsing and generation.
  • Large dictionaries of multiword units and expressions: Used in NLP software.
  • Digital speech collections: Used for phonetic analysis.
  • Journalistic corpus (Argentine Rioplatense Spanish): Analyzed for adjective usage over a thirty-year span.
  • STYLUS (SysTematicallY Derived Language USe): A resource derived from argument realizations for English verb classes.
  • Contemporary Amharic Corpus: Automatically tagged for morpho-syntactic information, collected from 25,199 documents.
  • Gold corpus of 200 English short stories: Annotated for summarization.
  • Speech corpus of Tigrinya: Created for speech recognition.
  • Parallel corpora for five Ethiopian Languages: Amharic, Tigrigna, Afan-Oromo, Wolaytta, and Geez.
  • FrameNet and SALSA corpora: Used for comparing frame semantics.
  • TED Talks corpus: A parallel multilingual oral corpus in English, French, and Chinese.
  • Cesselin Japanese-French dictionary: An open-source dictionary used for improving annotation.
  • Aligned bilingual Portuguese-English corpus: Used for analyzing Portuguese contractions.
  • Code-mixed English-Hindi corpus: A newly created parallel corpus.

Key Insights and Ideas

  • Bridging the Divide: The workshop underscored the growing synergy between linguistic theory and NLP practice, demonstrating that rich linguistic resources are crucial for building more robust and nuanced NLP applications.
  • The Power of NooJ: The NooJ platform emerged as a versatile tool for both linguists and computational linguists, facilitating the creation of diverse linguistic resources and their application in NLP tasks.
  • Multilingualism is Key: A significant portion of the research focused on building multilingual resources, highlighting the challenges and opportunities in cross-lingual NLP, particularly for low-resource languages.
  • Beyond Sentence-Level Analysis: Papers explored methods for analyzing phenomena at finer granularities, such as phrasal verbs, derivational processes in verbs, and the nuances of contractions.
  • The Value of Annotated Data: The consistent theme was the critical role of annotated corpora and dictionaries in advancing NLP capabilities, from semantic similarity to machine translation and speech recognition.

Key Results

  • Semantic Textual Similarity: Zhang & Moldovan assessed the limitations and strengths of neural net systems against rule-based systems on Semantic Textual Similarity, comparing performance against the SemEval 2012 benchmark.
  • Croatian Verb Derivatives: Kocijan et al. demonstrated how to detect and describe major derivational processes in Croatian verbs, exporting annotated chains for aspectual and derivational information.
  • French Locative Verbs: Boudhina & Fehri presented a rule-based system for disambiguating French locative verbs and translating them into Arabic.
  • Ethiopian Language MT: Abate et al. developed parallel corpora for seven Ethiopian languages and conducted MT experiments, noting the negative impact of morphological complexity on performance.
  • Cross-Lingual Frame Semantics: Sikos & Padó showed how multilingual vector representations of frames can address the need for accessing broad-coverage resources across language pairs.
  • Japanese-French Dictionary Analysis: Boitet et al. explored automatic classification of illustrative examples in a large Japanese-French dictionary obtained by OCR.
  • Code-Mixed Translation: Dhar et al. created a parallel English corpus from code-mixed English-Hindi data and presented an augmentation approach for MT.

Limitations and Future Work

  • Low-Resource Languages: The workshop highlighted the challenges in building resources for low-resource languages, with future work focusing on developing more efficient methods for their creation and utilization in MT.
  • Morphological Complexity: The impact of morphological complexity on MT performance was noted as a significant challenge, suggesting further research into handling such complexities.
  • Automatic Annotation: While progress was shown in automatic annotation (e.g., for Amharic and Tigrinya), the accuracy and robustness of these methods for diverse languages and domains remain areas for continued development.
  • Cross-Lingual Transfer: The potential for leveraging resources and methodologies across languages was a recurring theme, indicating future work in developing more effective cross-lingual transfer techniques.
  • Integration of Resources: The ultimate goal of integrating diverse linguistic resources into existing MT systems was emphasized, suggesting future work on system design and plug-and-play capabilities.

Summary updated

BibTeX
@inproceedings{machonis2018proceedings,
  author = {Peter Machonis and Anabela Barreiro and Kristina Kocijan and Max Silberztein},
  title = {{Proceedings of the First Workshop on Linguistic Resources for Natural Language Processing}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2018},
}
Figure