Multiword Expressions We Live by: A Validated Usage-based Dataset from Corpora of Written Italian

Francesca Masini, M. Silvia Micheli, Andrea Zaninello, Sara Castagnoli, Malvina NissimCLiC-it 2020 · OpenReview

Introduction

Multiword Expressions (MWEs)—including idioms like kick the bucket, compound nouns like cell phone, and fixed phrases like by and large—represent a significant challenge for Natural Language Processing (NLP). These combinations often exhibit semantic non-compositionality, where the meaning of the whole is not a simple sum of its parts. While English benefits from extensive MWE resources, languages like Italian have historically faced a data gap, relying on automated extraction methods prone to significant noise. This paper presents a manually validated, usage-based dataset of 1,682 Italian MWEs, providing a high-quality "gold standard" for the research community.

Methodology

The researchers developed a multi-step pipeline to extract and validate MWEs from two major Italian corpora: la Repubblica (journalistic, ~380M tokens) and PAISÀ (web-based, ~250M tokens).

  1. Candidate Extraction: Using the EXTra software, the team scanned the corpora for 122 specific Part-of-Speech (POS) sequences (e.g., Noun-Preposition-Noun). POS patterns were prioritized over complex syntactic relations to better capture fixed, adjacent word combinations.
  2. Statistical Ranking: Candidates were ranked using LogLikelihood (LL), identifying word combinations occurring together more frequently than chance.
  3. Filtering: A threshold of LL ≥ 7,500 was applied to balance precision and variety, yielding 7,045 initial candidates.
  4. Manual Validation: Two human annotators independently reviewed the candidates for fixedness and idiomaticity. A third annotator resolved disagreements, which occurred in fewer than 10% of cases.

Results and Pattern Analysis

The validation process revealed that only 24% (1,682) of automatically extracted candidates were true MWEs, highlighting the necessity of human oversight. While 122 POS patterns were initially targeted, the validated dataset instantiated only 82 of them.

Table 1: Most frequent POS patterns and examples in the validated dataset.

As shown in Table 1, the most frequent structure in Italian is Noun-Preposition-Noun (e.g., punto di vista), which accounts for a significant portion of complex nominals. Other dominant patterns include Verb-Article-Noun (valere la pena) and Verb-Preposition-Noun (scendere in campo).

Corpus Divergence and Linguistic Insights

A key finding of the study is the high degree of divergence between the two source corpora. Despite both representing mid-to-formal written Italian, there was only a 22% overlap in MWEs between la Repubblica and PAISÀ.

Table 2: Distribution of MWEs across the la Repubblica and PAISÀ corpora.

Table 2 illustrates this distribution: 982 MWEs were unique to the journalistic corpus, while 328 were unique to the web corpus. This suggests that MWE usage is highly sensitive to text genre and domain. Furthermore, the study found that over 70% of the lemmas in the dataset belong to the Vocabolario di Base (the 7,000 most common Italian words), indicating that MWEs are primarily constructed from common "generic" words like fare (to do/make) or prendere (to take).

Conclusion and Future Work

The release of this validated dataset provides a critical benchmark for Italian NLP tasks, such as MWE identification and machine translation. However, the authors note limitations: the current set only includes contiguous MWEs and does not yet account for discontinuous phrases (e.g., prendere una saggia decisione). Future iterations aim to expand the dataset to include non-contiguous occurrences and explore the semantic variability of inflected MWE forms.

Summary updated

BibTeX
@inproceedings{masini2020multiword,
  author = {Francesca Masini and M. Silvia Micheli and Andrea Zaninello and Sara Castagnoli and Malvina Nissim},
  title = {{Multiword Expressions We Live by: A Validated Usage-based Dataset from Corpora of Written Italian}},
  booktitle = {CLiC-it},
  year = {2020},
  url = {https://openreview.net/forum?id=9bn-7L9npZd},
}
Figure