TL;DR
Researchers have developed SISA, the first multilingual, syntax-based sentiment analysis system designed specifically for the five official languages of the Iberian Peninsula (Basque, Catalan, Galician, Portuguese, and Spanish). By leveraging Universal Dependencies and a shared set of syntactic rules, the system achieves robust performance across different language families without requiring language-specific rule engineering.
What Problem Does This Paper Address?
Sentiment analysis (SA) isn't just about counting "good" and "bad" words. To truly understand a sentence, a system needs to handle linguistic nuances like negation ("It is not good"), intensification ("It is very good"), and adversative clauses ("It is good but expensive").
Traditionally, these nuances are captured through syntactic rules tied to specific languages and specific parser outputs. This creates a massive bottleneck: if you want to support five different languages, you often have to write and maintain five different sets of rules. This paper tackles the challenge of creating a unified symbolic system where a single set of syntactic rules can be shared across multiple languages, even those from different families like Basque (an isolate) and Spanish (Romance).
Methods
The authors built a pipeline called SISA (Syntactic Iberian SA) based on three core pillars:
- Multilingual Lexica: They combined the Spanish SFU lexicon with the ML-Senticon resource, translating them into the other four languages using Apertium. To handle "false friends" (words that look the same but have different meanings/polarities across languages), they averaged the polarities in the final unified Iberian lexicon.
- Universal Parsing: Instead of five separate parsers, they trained a single multilingual PoS-tagger and dependency parser. By adhering to Universal Dependencies (UD) guidelines, the parser produces a standardized structural representation regardless of the input language.
- Compositional Operations: The "secret sauce" is a set of four universal syntactic rules:
- Intensification: Adjusts polarity based on modifiers like "very."
- Adversative Clauses: Handles the polarity shift in "but" constructions.
- Negation: Shifts the polarity of the head structure.
- 'If' Irrealis: Accounts for the uncertainty in conditional statements.
Because these rules operate on the standardized UD trees, the exact same rule for negation works for both Portuguese and Basque.
Datasets
The researchers evaluated SISA using a mix of established and synthetic corpora:
- Spanish: SFU reviews (400 long reviews).
- Portuguese: SentiCorpus-PT 0.1 (2,086 sentences).
- Basque: Basque Opinion Dataset (224 reviews).
- Catalan & Galician: Due to a lack of native sentence-level corpora, the authors created synthetic datasets by automatically translating the Spanish SFU corpus. Notably, this paper provides the first published SA results for the Galician language.
The scale of the final combined lexica used is detailed in the table below:
Key Insights and Ideas
- The Power of Universality: The study proves that syntactic rules for sentiment are largely language-agnostic if the underlying syntactic representation (the dependency tree) is harmonized.
- Lexical Fusion vs. False Friends: Merging lexica into a single "Iberian" resource generally improves performance by increasing coverage. However, it can be a double-edged sword for languages like Portuguese, where specific "false friends" (like espantoso, which is negative in Spanish but can be positive in Portuguese) can lead to slight drops in accuracy if not handled carefully.
- Synergy is Decisive: The most significant performance gains occurred when combining the multilingual lexicon with syntactic rules. In Catalan, for instance, this combination was responsible for a ~15-point jump in accuracy.
Key Results
SISA was compared against LinguaKit (LKit), a supervised sentiment analyzer. The results showed that the symbolic, rule-based approach is highly competitive and often superior to supervised models in these multilingual settings.
As seen in Figure 5 (Table 4), the ML+O setting (Multilingual lexica + universal Operations) achieved the best results for Spanish (76.50%), Catalan (73.00%), and Basque (72.32%). The impact of the syntactic operations alone was substantial, providing a boost of over 15 points in Spanish.
Limitations and Future Work
The authors noted a few hurdles:
- Basque Parsing: Performance in Basque was lower than other languages (62.01 LAS). The authors attribute this to the high frequency of non-projective syntactic structures in Basque, which their current parser wasn't optimized to handle.
- Synthetic Data: While useful for initial evaluation, the synthetic Catalan and Galician datasets may not capture the full linguistic variety of native speakers.
- Social Media: The current system was tested on reviews and news. The authors are already working on a manually labeled Galician tweet corpus to test SISA on the "noisier" language found on social media.
By demonstrating that a single engine can power sentiment analysis for an entire peninsula, this work paves the way for more efficient, cross-lingual NLP tools in other multi-language regions.
Summary updated