TL;DR
This paper introduces the first general-purpose Abstract Meaning Representation (AMR) corpus for Brazilian Portuguese, built by adapting English AMR guidelines and leveraging existing resources. The work addresses the scarcity of AMR data for Portuguese, crucial for advancing Natural Language Processing (NLP) tasks in the language.
What Problem Does This Paper Address?
Natural Language Processing (NLP) has seen a resurgence in interest in language understanding and dialogue systems. A key component of this is representing the meaning of sentences in a structured, machine-readable format. Abstract Meaning Representation (AMR) is a prominent semantic formalism that encodes sentence meaning as a directed, rooted graph, capturing semantic roles, entities, and more. While a substantial AMR corpus exists for English, there's a significant gap for other languages, including Brazilian Portuguese. The existing Portuguese AMR corpus is limited to a single book ("The Little Prince"), which has a unique style and restricted vocabulary, hindering the development of general-purpose NLP tools. This paper aims to bridge this gap by creating a more comprehensive AMR corpus for Brazilian Portuguese, covering diverse domains and text genres.
Methods
The authors adapted the existing AMR guidelines for English to Brazilian Portuguese. The core methodology involved:
- Guideline Adaptation: Translating and adapting the English AMR guidelines to account for specific linguistic phenomena in Portuguese. This included defining how to handle modal verbs, indeterminate subjects, and multi-word expressions.
- Lexical Resource Utilization: Leveraging Verbo-Brasil, a Portuguese lexical resource that provides semantic role annotation, to aid in verb sense disambiguation and argument structure identification.
- Corpus Construction:
- News Domain: Collecting news articles from Folha de São Paulo, a major Brazilian news agency, across various sections (daily news, world news, education, etc.).
- PropBank.Br Integration: Incorporating sentences from PropBank.Br, which already have semantic role annotations, to further enrich the corpus.
- Annotation Process:
- Two-Step Annotation: Annotators first independently analyzed sentence structure and identified concepts, followed by using the AMR Editor tool to produce the final AMR annotation in PENMAN format.
- Sentence Selection: Initially focusing on shorter sentences (below the average token count) to manage complexity and gradually increasing sentence length as annotators gained experience.
- Handling Unannotated Verbs: Verbs not found in Verbo-Brasil were collected for future repository enrichment.
Datasets
- Existing AMR Corpus: The paper references the AMR-annotated corpus for English (approximately 39,000 sentences) and the existing, but limited, AMR corpus for Brazilian Portuguese based on "The Little Prince" (1,527 sentences).
- Newly Introduced Corpus: The authors created a new general-purpose AMR-annotated corpus for Brazilian Portuguese. This corpus consists of:
- News Sentences: 4,563 sentences selected from news articles, covering diverse domains.
- PropBank.Br Sentences: 3,012 sentences from PropBank.Br.
- Total: The current corpus comprises 299 AMR-annotated sentences (including those used for inter-annotator agreement).
Key Insights and Ideas
- Feasibility of Adaptation: Adapting English AMR guidelines to Portuguese is feasible, but requires careful consideration of language-specific phenomena.
- Importance of Lexical Resources: Verbo-Brasil proved invaluable for Portuguese AMR annotation, but its coverage needs expansion.
- Challenges in Modal Verbs and MWEs: Modal verbs and multi-word expressions presented significant annotation challenges due to their varied translations and the reliance on English framesets.
- News vs. Literary Text: The statistical differences in relation frequencies between the news corpus and the "The Little Prince" corpus highlight how text genre influences semantic representation (e.g., news favoring "quant" relations for quantities, while tales might use more "degree" and "poss" relations).
- Annotator Training is Crucial: The annotation process required extensive training and iterative refinement to achieve reasonable inter-annotator agreement.
Key Results
- Corpus Size: The current corpus contains 299 AMR-annotated sentences, with 907 concepts and 711 relations (excluding instance, name, and op relations).
- Inter-Annotator Agreement: The Smatch score for inter-annotator agreement during the training stage was generally between 0.70 and 0.77, which is comparable to results reported for English AMR annotation.
- Identified Challenges: The paper details several specific annotation challenges, including handling modal verbs, auxiliary verbs, indeterminate subjects, and specific Portuguese characters.
Limitations and Future Work
- Lexical Resource Coverage: The Verbo-Brasil resource needs to be expanded to cover more verbs and verb senses to improve annotation efficiency and coverage.
- Annotation Scope: The current corpus is relatively small (299 sentences) and the focus was initially on shorter sentences. Future work will involve increasing the corpus size and coverage.
- Back-Translation: The authors plan to explore back-translation strategies to accelerate the annotation process.
- OPINANDO Project: Further details about the corpus and ongoing work are available through the OPINANDO project.
Summary updated