Robust Named Entity Recognition with Truecasing Pretraining
Authors: Stephen Mayhew, Gupta Nitish, Dan Roth8480-8487
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In experiments over several datasets of varying domain and casing quality, we show that our new model improves performance in uncased text, even adding value to uncased BERT embeddings. Our method achieves a new state of the art on the WNUT17 shared task dataset. |
| Researcher Affiliation | Academia | University of Pennsylvania Philadelphia, PA, 19104 {mayhew, nitishg, danroth}@seas.upenn.edu |
| Pseudocode | No | The paper includes model diagrams (Figure 2, Figure 3) but does not provide any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper references external tools and libraries used for experiments (e.g., 'scripts from the moses package', 'Allen NLP', 'bert-base-uncased as provided by Hugging Face', 'Glo Ve', 'Fast Text') but does not state that the authors' own implementation code for the described methodology is open-source or provide a link. |
| Open Datasets | Yes | We experiment on 3 English datasets: Co NLL 2003 (Sang and Meulder 2003), Ontonotes v5 (Hovy et al. 2006), WNUT17 (Derczynski et al. 2017). ... We train on two different datasets: the Wikipedia dataset (Wiki) introduced in Coster and Kauchak (2011) and used in Susanto, Chieu, and Lu (2016), and a specially preprocessed large dataset from English Common Crawl (CC).1 commoncrawl.org |
| Dataset Splits | Yes | Data statistics are shown in Table 3. ... Dataset Train Dev Test Co NLL2003 203,621 51,362 46,435 ... Ontonotes 1,088,503 147,724 152,728 ... WNUT17 62,730 15,733 23,394 |
| Hardware Specification | No | The paper states 'All experiments used Allen NLP' but does not provide any specific hardware details such as CPU/GPU models or memory specifications used for running the experiments. |
| Software Dependencies | No | The paper mentions using 'Allen NLP (Gardner et al. 2017)', 'Glo Ve', 'Fast Text', and 'bert-base-uncased as provided by Hugging Face', but does not specify version numbers for these or other software dependencies. |
| Experiment Setup | Yes | For the truecaser, with character embedding of size 50, hidden dimension size 100, and dropout of 0.25. One trick during training was to leave 20% of the sentences in their original case... For the NER model, we used character embeddings of size 16, hidden dimension size 256, and Glo Ve 100-dimensional uncased embeddings |