The Role of Adverbs in Language Variety Identification: The Case of Portuguese Multi-Word Adverbs

Izabela Müller, Nuno Mamede, Jorge BaptistaNAACL 2024

TL;DR

This paper investigates the role of multi-word adverbs in distinguishing Brazilian Portuguese (PT-BR) from European Portuguese (PT-PT). By creating a comprehensive lexicon and analyzing journalistic corpora, they found a significant overlap between the two varieties, suggesting multi-word adverbs are not a strong sole indicator for dialect identification, though some unique expressions do exist.

What Problem Does This Paper Address?

Distinguishing between language varieties, like Brazilian and European Portuguese, is crucial for Natural Language Processing (NLP) tasks that require dialect-specific understanding. While many linguistic features can be used for this, multi-word adverbs (expressions of two or more words acting as a single adverbial unit, often with idiomatic meanings) have been largely overlooked. This paper addresses this gap by exploring whether these multi-word adverbs can effectively differentiate between PT-BR and PT-PT, and to what extent their usage is asymmetrical across the two varieties.

Methods

The authors first compiled a large lexicon of Portuguese multi-word adverbs, annotating each with diatopic information (indicating if it's specific to PT-BR, PT-PT, or common to both). This lexicon was built upon existing work and expanded to include Brazilian Portuguese specific expressions. They then applied this lexicon to analyze two comparable journalistic corpora: one for PT-BR and one for PT-PT, sourced from the Discrimination of Similar Languages (DSL) Corpus Collection. The analysis involved identifying occurrences of these multi-word adverbs within the corpora and examining their distribution and variety-specific markings. They also employed a linguistic framework for classifying adverbs based on their syntactic and semantic features.

Datasets

  • Lexicon of Portuguese Multi-word Adverbs: A custom lexicon containing 3,665 multi-word adverbs, annotated with variety information (PT-BR, PT-PT, or common).
  • DSL Corpus Collection (DSLCC): Specifically, the European Portuguese (PT-PT) and Brazilian Portuguese (PT-BR) partitions were used. These partitions consist of 18,000 sentences each.
  • PtTenTen2020 Corpus: Used for verifying occurrences and disambiguating specific adverbs.

Key Insights and Ideas

  • High Lexical Overlap: The most striking finding is the substantial similarity in multi-word adverb usage between PT-BR and PT-PT. This suggests that relying solely on multi-word adverbs might not be the most effective strategy for dialect identification.
  • Idiomaticity as a Challenge: The idiomatic nature of many multi-word adverbs makes them difficult to identify automatically, as their meaning isn't always compositional. This also leads to "false positive" matches where a sequence of words resembles an adverb but functions differently.
  • Subtle Differences Exist: Despite the overlap, the study did identify some adverbs unique to each variety, and also highlighted subtle differences in usage, such as the presence or absence of articles (e.g., "ao domicílio" vs. "a domicílio").
  • Orthographic Reform Ambiguity: The analysis revealed instances where spelling variations (e.g., "de facto" vs. "de fato") likely stem from uncertainties in applying the Portuguese orthographic reform, further complicating clear-cut distinctions.
  • Importance of Context: The paper emphasizes that identifying the true function of multi-word expressions often requires deep contextual analysis, as simple string matching can be misleading.

Key Results

  • Lexicon Composition: The lexicon of multi-word adverbs is predominantly common to both PT-BR and PT-PT (68.1%). Exclusively Brazilian adverbs constitute 26.4%, while exclusively European adverbs make up 5.5%.
  • Corpus Analysis:
    • The number of distinct multi-word adverb entries found in the corpora was remarkably similar: 715 for PT-PT and 668 for PT-BR.
  • The proportion of entries attributed to common Portuguese was high in both partitions (PT-PT: 87.9%, PT-BR: 92.8%).
  • Exclusively variety-specific entries were relatively small: PT-PT had 10.3% unique adverbs, while PT-BR had 6.9%.
  • The total number of adverb matches in the corpora was also very close: 5,695 for PT-PT and 5,700 for PT-BR.
  • False Positives: A significant portion of matches were "false positives" (78.3% in one analysis), indicating that the system identified sequences that were not true multi-word adverbs in context, or were part of longer expressions.
  • Spurious Occurrences: Instances of adverbs marked as exclusive to one variety appearing in the other were rare, with only a few cases identified and often explained by ambiguity or spelling variations.

Limitations and Future Work

  • Ambiguity and False Positives: The authors acknowledge that identifying multi-word adverbs accurately is challenging due to their ambiguity and potential overlap with other linguistic structures (like compound nouns or free prepositional phrases). This leads to a high rate of false positives.
  • Sub-optimal for Dialect Identification: The significant lexical overlap between PT-BR and PT-PT in multi-word adverbs makes this category "sub-optimal" for dialect identification tasks when used in isolation.
  • Future Work: The authors plan to provide annotated corpus partitions with newly identified multi-word adverbs. They believe this resource will be valuable for enhancing other dialect-sensitive NLP tasks and improving parsing accuracy. Further linguistic analysis is needed to improve the precision of identifying these ambiguous strings.

Summary updated

BibTeX
@inproceedings{muller2024role,
  author = {Izabela Müller and Nuno Mamede and Jorge Baptista},
  title = {{The Role of Adverbs in Language Variety Identification: The Case of Portuguese Multi-Word Adverbs}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2024},
}
Figure