TL;DR
This paper introduces a methodology and a new dataset for extracting paraphrastic variations between European Portuguese (EP) and Brazilian Portuguese (BP). By focusing on multiword and phrasal units, the authors aim to improve language technology applications like machine translation and text generation by enabling better adaptation between these two major varieties of Portuguese.
What Problem Does This Paper Address?
Paraphrasing, the ability to express the same idea using different words or structures, is fundamental to human communication and a crucial capability for advanced Natural Language Processing (NLP) systems. However, adapting these systems to handle the nuances of different language varieties, such as European Portuguese (EP) and Brazilian Portuguese (BP), presents a significant challenge. These varieties, while mutually intelligible, exhibit distinct lexical, syntactic, and stylistic differences. Existing approaches often rely on ad-hoc methods for alignment, lacking a systematic way to capture these specific cross-variety paraphrastic phenomena. This gap hinders the development of robust paraphrasing systems that can effectively handle variety adaptation, a key requirement for applications like cross-lingual communication, language learning, and content generation for diverse audiences.
Methods
The authors developed a methodology inspired by the "Logos Model," focusing on a semantico-syntactic analysis of paraphrastic units. They employed the CLUE-Aligner tool to manually align multiword and phrasal units within monolingual parallel sentences of EP and BP texts. This alignment process involved identifying, analyzing, and registering corresponding phrasal equivalents. The methodology prioritizes capturing both lexical and semantico-syntactic phenomena that can be challenging for paraphrasing systems. They also developed "CLUE4Paraphrasing Alignment Guidelines" to standardize this manual alignment process, using syntactic and semantic properties of phrases.
Datasets
The primary dataset used and introduced in this paper is the e-PACT corpus (eSPERTO Paraphrase Aligned Corpus of EN-EP/BP Translations). This corpus is a parallel corpus of aligned paraphrases derived from English-European Portuguese and English-Brazilian Portuguese translations of David Lodge's novels, "Therapy" and "Changing Places." The authors manually annotated 40% of the aligned sentences from these novels, resulting in approximately 1,277 parallel sentences and thousands of collected paraphrastic alignments. This dataset is intended to be a valuable resource for the scientific community for research purposes.
Key Insights and Ideas
- Variety Adaptation is Crucial: The paper strongly emphasizes that understanding and leveraging paraphrastic differences between EP and BP is essential for creating truly adaptable language technologies.
- Multiwords and Phrasal Units are Key: The focus on multiword expressions and phrasal units, rather than just single words, is critical for capturing the subtle but significant differences in how ideas are expressed across varieties.
- Systematic Alignment is Necessary: Moving beyond ad-hoc methods, the CLUE-Aligner and the developed guidelines provide a structured approach to building a reliable paraphrastic database.
- Lexicon-Grammar as a Foundation: The integration of lexicon-grammar tables, particularly for EP, demonstrates a promising avenue for formalizing linguistic knowledge and enabling more sophisticated paraphrasing.
- Stylistic vs. Contrastive Variance: A key distinction is made between paraphrases that are stylistic choices available in both varieties and those that represent genuine contrastive variance, exclusive to one variety.
Key Results
The paper successfully developed a methodology and a new dataset (e-PACT corpus) for extracting EP-BP paraphrastic variants. They illustrated several types of paraphrastic phenomena, including:
- Support Verb Constructions: Differences in the use of support verbs and their associated structures (e.g., EP's
subiram em espiralvs. BP'sespiralando). - Progressive Constructions: The contrast between EP's
estar a + V-Inf(e.g.,estive a observar) and BP'sficar + V-Ger(e.g.,fiquei observando). - Clitic Pronoun Placement: Differences in the position of clitic pronouns (e.g., EP's
puseram-mevs. BP'sme puseram). - Determiner Usage: The higher frequency of zero determiners in BP compared to EP.
- Forms of Address: Variations in the use of second and third-person pronouns for politeness and formality.
While quantitative results like accuracy scores are not the primary focus, the paper demonstrates the feasibility of creating a structured paraphrastic database and highlights the rich linguistic phenomena that can be captured.
Limitations and Future Work
The authors acknowledge several limitations and outline future directions:
- Corpus Bias: The use of translated novels means the e-PACT corpus might carry a syntactic and lexical bias from the English source. Future work could explore non-translation corpora, though these are harder to find and prepare.
- Incomplete Coverage: Due to space constraints, only a few examples of paraphrasing phenomena are discussed, and a large amount of collected data remains undiscussed.
- Manual Alignment Effort: The current alignment process is manual, which is time-consuming.
- BP Lexicon-Grammar: While lexicon-grammar tables exist for EP, similar formalization for BP is needed.
- Automated Alignment: The ultimate goal is to develop an automated alignment tool, moving beyond manual annotation.
- Grammar Generation: The authors plan to create grammars from the identified contrasts to enable more generalized paraphrasing.
- Testing in Applications: The lexicon-grammar resources have not yet been fully tested in applications like question-answering or summarization, which is a planned future step, particularly for e-learning environments.
- Validation: Further validation of the collected contrasts by expert linguists is necessary.
Summary updated