Multilinguals at SemEval-2022 Task 11: Complex NER in Semantically Ambiguous Settings for Low Resource Languages

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1469 - 1476 July 14-15, 2022 ©2022 Association for Computational Linguistics Multilinguals at SemEval-2022 Task 11: Complex NER in Semantically Ambiguous Settings for Low Resource Languages Amit Pandey*, Swayatta Daw*, Narendra Babu Unnam and Vikram Pudi International Institute of Information Technology, Hyderabad, India {amit.pandey, swayatta.daw}@research.iiit.ac.in narendra.babu.unnam@research.iiit.ac.in vikram@iiit.ac.in Abstract We leverage pre-trained language models to solve the task of complex NER for two low- resource languages: Chinese and Spanish. We use the technique of Whole Word Masking (WWM) to boost the performance of masked language modeling objective on large and un- supervised corpora. We experiment with multi- ple neural network architectures, incorporating CRF, BiLSTMs, and Linear Classifiers on top of a fine-tuned BERT layer. All our models outperform the baseline by a significant margin and our best performing model obtains a com- petitive position on the evaluation leaderboard for the blind test set. 1 Introduction The Named Entity Recognition (NER) task aims to identify the named entities in an input sequence and categorize them into certain predefined cate- gories or class labels. The task of NER can be further broken down into two subtasks: 1) identifi- cation of the entity span and 2) classification of the identified entity span into predefined class labels. For example, in the sentence: New York City is the most densely populated major city in the United States., New York City is a named entity of type LOCATION with an entity span of 3 tokens. The most popular NER task in the English lan- guage is CoNLL (Baevski et al., 2019), which is widely used as a benchmark for most NER mod- els. Multiple models have been able to obtain sufficiently high performances in this task setting (Wang et al., 2021a; Zhou and Chen, 2021; Lu- oma and Pyysalo, 2020; Schweter and Akbik, 2020; Ye et al., 2021; Yamada et al., 2020; Wang et al., 2021b). The CoNLL training set consists of 14,987 train sentences, which comprise 203,621 tokens in total for English data. The entity space con- sists of 4 different types of entity type labels (loca- tions, persons, organizations, and miscellaneous) *Equal Contribution to classify each named token. The English data was taken from the Reuters Corpus, which com- prises of Reuters News Stories for one year. The training data source, and by extension the labeled named entities, comprises of majorly popular en- tities found in the general English textual content prevalent in the media. Hence, these entities were easier to classify into the correct classes due to the large prevalence of training data. With the use of pre-trained transformer-based language models, which are already trained on a large unlabelled corpus of English text, this task became even less challenging, as the nature of textual structure in these corpora largely overlap with that of CoNLL. However, this task becomes challenging in prac- tical settings where a multitude of varieties of named entities is possible. Many of these entities, like Creative Works (CW) and Products …

Content truncated to 500 words