Automated Concatenation of Embeddings for Structured Prediction

Pretrained contextualized embeddings are powerful word representations for structured prediction tasks. Recent work found that better word representations can be obtained by concatenating different types of embeddings... However, the selection of embeddings to form the best concatenated representation usually varies depending on the task and the collection of candidate embeddings, and the ever-increasing number of embedding types makes it a more difficult problem. In this paper, we propose Automated Concatenation of Embeddings (ACE) to automate the process of finding better concatenations of embeddings for structured prediction tasks, based on a formulation inspired by recent progress on neural architecture search. Specifically, a controller alternately samples a concatenation of embeddings, according to its current belief of the effectiveness of individual embedding types in consideration for a task, and updates the belief based on a reward. We follow strategies in reinforcement learning to optimize the parameters of the controller and compute the reward based on the accuracy of a task model, which is fed with the sampled concatenation as input and trained on a task dataset. Empirical results on 6 tasks and 21 datasets show that our approach outperforms strong baselines and achieves state-of-the-art performance with fine-tuned embeddings in all the evaluations. read more

PDF Abstract ACL 2021 PDF ACL 2021 Abstract

Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Benchmark
Part-Of-Speech Tagging ARK ACE Acc 94.4 # 1
Chunking CoNLL 2000 ACE Exact Span F1 97.3 # 1
Named Entity Recognition CoNLL 2002 (Dutch) ACE + document-context F1 95.7 # 1
Named Entity Recognition CoNLL 2002 (Dutch) ACE F1 94.6 # 3
Named Entity Recognition CoNLL 2002 (Spanish) ACE + document-context F1 95.9 # 1
Named Entity Recognition CoNLL 2002 (Spanish) ACE F1 91.7 # 2
Chunking CoNLL 2003 (English) ACE F1 92.5 # 1
Named Entity Recognition CoNLL 2003 (English) ACE + document-context F1 94.6 # 1
Named Entity Recognition CoNLL 2003 (English) ACE F1 93.64 # 9
Chunking CoNLL 2003 (German) ACE F1 95.0 # 1
Named Entity Recognition CoNLL 2003 (German) ACE F1 87.0 # 4
Named Entity Recognition CoNLL 2003 (German) ACE + document-context F1 88.38 # 1
Named Entity Recognition CoNLL 2003 (German) Revised ACE F1 90.5 # 3
Named Entity Recognition CoNLL 2003 (German) Revised ACE + document-context F1 91.7 # 2
Semantic Dependency Parsing DM ACE In-domain 95.6 # 1
Out-of-domain 92.6 # 1
Semantic Dependency Parsing PAS ACE In-domain 95.8 # 1
Out-of-domain 94.6 # 1
Dependency Parsing Penn Treebank ACE UAS 97.2 # 3
LAS 95.8 # 2
Chunking Penn Treebank ACE F1 score 97.3 # 1
Semantic Dependency Parsing PSD ACE In-domain 83.8 # 1
Out-of-domain 83.4 # 1
Part-Of-Speech Tagging Ritter ACE Acc 93.4 # 1
Aspect Extraction SemEval 2014 Task 4 Sub Task 2 ACE Laptop (F1) 87.4 # 1
Restaurant (F1) 92.0 # 1
Aspect Extraction SemEval 2015 Task 12 ACE Restaurant (F1) 80.3 # 1
Aspect Extraction SemEval-2016 Task 5 Subtask 1 ACE F1 81.3 # 1
Aspect Extraction SemEval-2016 Task 5 Subtask 1 (Dutch) ACE F1 80.5 # 1
Aspect Extraction SemEval-2016 Task 5 Subtask 1 (Russian) ACE F1 79.4 # 1
Aspect Extraction SemEval-2016 Task 5 Subtask 1 (Spanish) ACE F1 79.9 # 1
Aspect Extraction SemEval-2016 Task 5 Subtask 1 (Turkish) ACE F1 81.9 # 1
Part-Of-Speech Tagging Tweebank ACE Acc 95.8 # 1

Methods