[Paper Review] Zero-Shot Cross-Lingual Opinion Target Extraction

Soufian Jebbara, Philipp Cimiano|arXiv (Cornell University)|Apr 19, 2019

Sentiment Analysis and Opinion Mining17 references7 citations

TL;DR

This paper proposes a zero-shot cross-lingual approach for opinion target extraction using multilingual word embeddings within a convolutional neural network, enabling accurate OTE prediction in low-resource languages without any annotated target-language data. The method achieves up to 87% of the performance of a monolingual model trained on target-language data when leveraging multiple source languages.

ABSTRACT

Aspect-based sentiment analysis involves the recognition of so called opinion target expressions (OTEs). To automatically extract OTEs, supervised learning algorithms are usually employed which are trained on manually annotated corpora. The creation of these corpora is labor-intensive and sufficiently large datasets are therefore usually only available for a very narrow selection of languages and domains. In this work, we address the lack of available annotated data for specific languages by proposing a zero-shot cross-lingual approach for the extraction of opinion target expressions. We leverage multilingual word embeddings that share a common vector space across various languages and incorporate these into a convolutional neural network architecture for OTE extraction. Our experiments with 5 languages give promising results: We can successfully train a model on annotated data of a source language and perform accurate prediction on a target language without ever using any annotated samples in that target language. Depending on the source and target language pairs, we reach performances in a zero-shot regime of up to 77% of a model trained on target language data. Furthermore, we can increase this performance up to 87% of a baseline model trained on target language data by performing cross-lingual learning from multiple source languages.

Motivation & Objective

• To address the scarcity of annotated data for opinion target extraction in low-resource languages.
• To enable zero-shot transfer of OTE models from a source language to a target language without any annotated examples in the target language.
• To investigate the impact of using multiple source languages on cross-lingual transfer performance.
• To evaluate the effectiveness of multilingual word embeddings in a zero-shot setting for sequence labeling tasks.
• To reduce the annotation cost for low-resource languages by leveraging cross-lingual transfer.

Proposed method

• The task is framed as a sequence tagging problem using the IOB scheme to label opinion target expressions.
• A multi-layer convolutional neural network (CNN) is used as the sequence tagging model, with input features limited to word embeddings to ensure simplicity and broad applicability.
• Multilingual word embeddings from Smith et al. (2017) and Lample et al. (2018) are used to align word representations across languages into a shared vector space.
• The model is trained on annotated data from a source language and applied directly to a target language without fine-tuning on target-language data.
• Performance is improved by training on multiple source languages, which enhances generalization and cross-lingual transfer.
• The approach is evaluated using F1-score on five languages, including English, Spanish, Dutch, French, and Turkish.

Experimental results

Research questions

• RQ1 Can a model trained on a source language achieve high performance in extracting opinion targets in a target language without any annotated training data in the target language?
• RQ2 How does using multiple source languages affect the zero-shot performance of cross-lingual opinion target extraction?
• RQ3 To what extent can multilingual word embeddings enable effective zero-shot transfer for sequence labeling in low-resource settings?
• RQ4 How does the performance of the cross-lingual model compare to a monolingual baseline trained on target-language data?
• RQ5 Can the inclusion of a small number of annotated examples from the target language significantly reduce the required annotation effort?

Key findings

• The zero-shot model achieved an F1-score of 0.50 for Spanish and 0.46 for Dutch when using English as the source language, corresponding to 74% and 77% of the performance of a monolingual baseline, respectively.
• By leveraging multiple source languages, the model improved to F1-scores of 0.58 and 0.53 for Spanish and Dutch, achieving 85% and 87% of the monolingual baseline performance.
• The model reduced the required number of annotated target-language examples by several hundred when combined with a small amount of target-language data.
• Turkish, an agglutinative language, showed the least improvement from cross-lingual learning, suggesting linguistic differences may affect transferability.
• The two multilingual embedding methods evaluated produced comparable results overall, though performance varied across language pairs.
• The model achieved state-of-the-art or competitive performance for Russian and Turkish, outperforming existing systems in these languages.