A Generic Framework for Multiword Expressions Treatment: from Acquisition to Applications

Carlos RamischACL 2012

TL;DR

This paper introduces a flexible, open-source framework (mwetoolkit) for automatically acquiring multiword expressions (MWEs) from text. The authors demonstrate its utility by integrating it into lexicography and machine translation, showing that automatic MWE acquisition can significantly improve these NLP applications.

What Problem Does This Paper Address?

Multiword expressions (MWEs) – like "rock 'n' roll" or "dry ice" – are ubiquitous in language and pose a significant challenge for Natural Language Processing (NLP) systems. Their meanings are often not directly derivable from their constituent words, and they exhibit limited syntactic flexibility, making them difficult to handle with standard linguistic rules. While MWEs are crucial for accurate language understanding and generation, many NLP applications either neglect them or rely on ad-hoc, system-specific solutions. This paper addresses the need for a unified, adaptable approach to automatically identify and acquire MWEs, and critically, to integrate this acquired knowledge into real-world NLP applications.

Methods

The core contribution is the mwetoolkit, a methodological framework designed for the automatic acquisition of MWEs. It's described as "open and flexible," meaning it's modular and can be customized. The pipeline involves:

  • Preprocessing: This stage can optionally enrich raw text with linguistic information like Part-of-Speech (POS) tags, lemmas, and dependency parses, if available for the target language.
  • Candidate Extraction: MWE candidates are identified using various methods, including n-grams, POS sequences, and syntactic patterns. The framework supports patterns that don't require strict contiguity of words.
  • Candidate Filtering: A range of techniques are employed to filter out non-MWEs, from simple frequency thresholds and stopword lists to sophisticated statistical association measures.
  • Validation: Candidates can be manually annotated or used to train machine learning models for further extraction.

A key aspect of the framework is its language independence and adaptability to different MWE types and domains. It doesn't rely on fixed-length expressions or strict adjacency, allowing for more general applicability.

Datasets

The paper reports on evaluations using several corpora:

  • English: Europarl corpus (13M words) for phrasal verb (PV) acquisition.
  • French: Europarl corpus (14.5M words) for noun compound (NC) acquisition.
  • Greek: Europarl corpus (26M words) for nominal MWE acquisition.
  • Portuguese: PLN-BR-FULL corpus (29M words) for complex predicate (CP) acquisition.

The authors also mention using the English–Portuguese Europarl v6 corpus (1.7M sentences, ~50M words) for training the SMT system. The created data sets for English, French, Greek, and Portuguese are noted as being freely available.

Key Insights and Ideas

  • Unified Framework is Key: The paper argues that a single, flexible framework that integrates multiple acquisition techniques is more beneficial than isolated methods. This allows for customization based on the specific MWE type and application.
  • Application-Oriented Evaluation Matters: Intrinsic evaluations of MWE acquisition (e.g., precision/recall of identified MWEs) are often vague and hard to generalize. Extrinsic evaluation, by integrating MWEs into real NLP applications, provides a more concrete measure of their impact and usefulness.
  • MWEs Improve NLP Applications: The experiments demonstrate that automatically acquired MWEs can indeed speed up and improve the quality of applications like computer-aided lexicography and statistical machine translation.
  • Flexibility is Paramount: The framework's ability to handle non-contiguous words and its language-agnostic design are highlighted as crucial for broad applicability.

Key Results

  • Lexicography:
    • English: 5,302 phrasal verb candidates extracted, 875 identified as true PVs.
  • French: 3,746 noun compound candidates obtained.
  • Greek: Lexicon of 815 nominal MWEs created.
  • Portuguese: 407,014 complex predicate candidates extracted, with 699 annotated as compositional and 74 as idiomatic verbal expressions.
  • Statistical Machine Translation (SMT):
    • Preliminary results on phrasal verb translation showed mixed improvements across different integration strategies, with some degrading performance. The authors attribute this to the need for more fine-grained evaluation and more data.
  • Table 3 shows baseline translation performance (0.53 Good, 0.36 Acceptable, 0.11 Incorrect) and how different integration strategies (TOK, PV?, PART, VERB, BILEX) performed, with some showing slight improvements or degradations.

Limitations and Future Work

  • SMT Integration Needs More Work: The authors acknowledge that the SMT experiments require further investigation to better understand the impact of different integration strategies.
  • Fine-grained Evaluation: The current SMT evaluation needs to consider more fine-grained classes of phrasal verbs rather than mixing them all.
  • More Data for SMT: More annotated data is needed to obtain more representative results in SMT.
  • Syntax-based Identification: The authors are currently applying syntax-based identification and analyzing word alignment for a better understanding of SMT integration.
  • Bilingual MWE Acquisition: They plan to pursue experiments on bilingual MWE acquisition from parallel and comparable resources.
  • MWE Simplification: A future direction is to experiment with MWE simplification (e.g., replacing "go back" with "regress") as a preprocessing step for SMT to improve translation quality.
  • Expert MT Systems: Evaluating strategies for integrating verbal MWEs in expert MT systems is also a future goal.
  • Language Specificity: While the framework is designed to be language-independent, it notes that languages using spaces to separate words are assumed, and additional preprocessing might be needed for languages like Chinese or Japanese, or for German compounds.

Summary updated

BibTeX
@inproceedings{ramisch2012generic,
  author = {Carlos Ramisch},
  title = {{A Generic Framework for Multiword Expressions Treatment: from Acquisition to Applications}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2012},
}
Figure