Untitled

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.

Cross-lingual Aspect-based Sentiment Analysis with Aspect Term Code-Switching

“…We treat English as the source language and other languages as targets. Following existing works to simulate true unsupervised setting (Jebbara and Cimiano, 2019;Hu et al, 2020), we use the English validation set in all experiments for the model selection. The original workshop also provides training data for each target language as well 6 , we thus discard the label of the training set in each target language and use the raw sentences as the unlabeled data, similar with previous studies (Wang and Pan, 2018;Wu et al, 2020).…”

“…Another line of work uses the cross-lingual word embeddings trained on large parallel bilingual corpus (Ruder et al, 2019). By switching the word embeddings between different languages, the model can be used in a language-agnostic manner (Barnes et al, 2016;Akhtar et al, 2018;Wang and Pan, 2018;Jebbara and Cimiano, 2019). Wang and Pan (2018) Table 6: Example of different label projection methods with French as the target language.…”

“…Then, word alignment algorithms (Dyer et al, 2013) are applied to project the labels (i.e., the position of the aspect term) to obtain labeled target language data (Zhou et al, 2015;Klinger and Cimiano, 2015). Later methods make use of the cross-lingual word embeddings (Ruder et al, 2019) trained on large parallel corpus to allow the model to be used in a language-independent manner, by simply switching the word embedding layer while keeping the model unchanged (Barnes et al, 2016;Akhtar et al, 2018;Jebbara and Cimiano, 2019) when adopted for different languages.…”

Improving Cross-lingual Text Classification with Zero-shot Instance-Weighting

“…For the sentiment and document classification task, we train a fully-connected layer on top of the output of the [CLS] token, which is considered to be the representation of the input sequence. For the opinion target extraction task, we formulate it as sequence labeling task (Agerri and Rigau, 2019;Jebbara and Cimiano, 2019). To extract such opinion target tokens is to classify each token into one of the following: Beginning, Inside and Outside of an aspect.…”

Analysis of Zero-Shot Crosslingual Learning between English and Korean for Named Entity Recognition

“…One important application of crosslingual embeddings has been found for transferring models trained on a high-resource language to a low-resource one (Lin et al, 2019;Schuster et al, 2019a;Artetxe and Schwenk, 2019). The latest multilingual transformer encoders such as BERT (Devlin et al, 2019) and XLM (Conneau et al, 2020) have made it possible to develop robust crosslingual models through zero-shot learning that requires no labeled training data on the target side (Jebbara and Cimiano, 2019;Chidambaram et al, 2019;Chi et al, 2020). However, these approaches tend not to work as well for languages whose words cannot be easily aligned.…”