🤖 AI Summary
Constructing structured knowledge from multilingual (English/French/Spanish) social media texts in tourism faces challenges of linguistic diversity and prohibitively high annotation costs. Method: This paper proposes a multitask NLP framework tailored for few-shot learning, introducing the first fine-grained, multilingual, domain-specific dataset for tourism. It systematically evaluates and integrates few-shot learning, pattern-based prompting, and parameter-efficient fine-tuning, jointly leveraging sequence labeling and semantic resource alignment. Contribution/Results: The framework achieves performance on par with fully supervised baselines using only 15 samples for sentiment analysis, 160 for location recognition, and 200 for topic concept extraction (315 classes), effectively overcoming low-resource bottlenecks. It establishes a scalable, low-dependency automation paradigm for cross-lingual tourism sentiment analysis and knowledge graph construction.
📝 Abstract
The rising influence of social media platforms in various domains, including tourism, has highlighted the growing need for efficient and automated natural language processing (NLP) approaches to take advantage of this valuable resource. However, the transformation of multilingual, unstructured, and informal texts into structured knowledge often poses significant challenges. In this work, we evaluate and compare few-shot, pattern-exploiting and fine-tuning machine learning techniques on large multilingual language models (LLMs) to establish the best strategy to address the lack of annotated data for 3 common NLP tasks in the tourism domain: (1) Sentiment Analysis, (2) Named Entity Recognition, and (3) Fine-grained Thematic Concept Extraction (linked to a semantic resource). Furthermore, we aim to ascertain the quantity of annotated examples required to achieve good performance in those 3 tasks, addressing a common challenge encountered by NLP researchers in the construction of domain-specific datasets. Extensive experimentation on a newly collected and annotated multilingual (French, English, and Spanish) dataset composed of tourism-related tweets shows that current few-shot learning techniques allow us to obtain competitive results for all three tasks with very little annotation data: 5 tweets per label (15 in total) for Sentiment Analysis, 10% of the tweets for location detection (around 160) and 13% (200 approx.) of the tweets annotated with thematic concepts, a highly fine-grained sequence labeling task based on an inventory of 315 classes. This comparative analysis, grounded in a novel dataset, paves the way for applying NLP to new domain-specific applications, reducing the need for manual annotations and circumventing the complexities of rule-based, ad hoc solutions.
