Papers by Ulli Waltinger
E-BERT: Efficient-Yet-Effective Entity Embeddings for BERT (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods to enhance BERT with factual knowledge about entities require no additional pretraining and no changes to the encoder itself. |
| Approach: | They propose a way to inject factual knowledge into the pretrained BERT model by aligning Wikipedia2Vec entity vectors with BERT's native wordpiece vector space and feeding the aligned entity vector into BERT as if they were wordpieces. |
| Outcome: | The proposed version outperforms baseline models on unsupervised question answering, supervised relation classification and entity linking tasks. |
News Article Teaser Tweets and How to Generate Them (N19-1)
Copied to clipboard
| Challenge: | A teaser is a short reading suggestion for an article that is illustrative and includes curiosity-arousing elements to entice potential readers to read particular news items. |
| Approach: | They propose a benchmark and baseline system for the process of generating teasers. |
| Outcome: | The proposed system is best performing with a pointer network. |
Sentence Meta-Embeddings for Unsupervised Semantic Textual Similarity (2020.acl-main)
Copied to clipboard
| Challenge: | Existing word embeddings combine complementary strengths of their components to achieve unsupervised semantic similarity (STS). |
| Approach: | They propose to ensemble pre-trained sentence encoders into sentence meta-embeddings to achieve unsupervised Semantic Textual Similarity (STS) they adapt dimensionality reduction, generalized Canonical Correlation Analysis and cross-view auto-encoders to their work. |
| Outcome: | The proposed method achieves 3.7% to 6.4% Pearson’s r over single-source word embeddings on the STS Benchmark and on the StS12-STS16 datasets. |
Inexpensive Domain Adaptation of Pretrained Language Models: Case Studies on Biomedical NER and Covid-19 QA (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Pretrained Language Models (PTLMs) are typically pretraining on target-domain text, which is expensive in terms of hardware, runtime and CO 2 emissions. |
| Approach: | They propose a faster, CPU-only domainadaptation method that trains Word2Vec on target-domain text and aligns the resulting word vectors with the wordpiece vectors of a general-domain PTLM. |
| Outcome: | The proposed method covers 60% of the BioBERT - BERT F1 delta, 5% of BioBERTS’s CO2 footprint and 2% of its cloud compute cost. |