UCPhrase: Unsupervised Context-aware Quality Phrase Tagging
This project provides an unsupervised framework for mining and tagging quality phrases on text corpora. In this work, we recognize the power of pretrained language models in identifying the structure of a sentence. The attention matrices generated by a Transformer model are informative to distinguish quality phrases from ordinary spans, as illustrated in the following example.
With a lightweight CNN model to capture inter-word relationships from various ranges, we can effectively tackle the task of phrase tagging as a multi-channel image classifiaction problem.
For model training, we seek to alleviate the need for human annotation and external knowledge bases. Instead, we show that sufficient supervision can be directly mined from large-scale unlabeled corpus. Specifically, we mine frequent max patterns with each document as context, since by definition, high-quality phrases are sequences that are consistently used in context. Compared with labels generated by distant supervision, silver labels mined from the corpus itself preserve better diversity, coverage, and contextual completeness. The superiority is supported by comparison on two public datasets.
We compare our method with existing ones on the KP20k dataset (publication data from CS domain) and the KPTimes dataset (news articles). UCPhrase significantly outperforms prior arts without supervision. Compared with off-the-shelf phrase tagging tools, UCPhrase also shows unique advantages, especially in its ability to generalize to specific domains without reliance on manually curated labels or KBs. We provide comprehensive case studies to demonstrate the comparison among different tagging methods. We also have some interesting findings in the discussion sections.
We aim to build UCPhrase as a practical tool for phrase tagging, though it is certainly far from perfect. Please feel free to try on your own corpus and give us feedbacks if you have any ideas that can help build better phrase tagging tools!
Facts: UCPhrase is a joint work by researchers from UI at Urbana Champaign, and University of California San Diago.
Quick Start
Step 1: Download and unzip the data folder
wget https://www.dropbox.com/s/1bv7dnjawykjsji/data.zip?dl=0 -O data.zip
unzip -n data.zip
Step 2: Install and compile dependencies
bash build.sh
Step 3: Run experiments
cd src
python exp.py --gpu 0 --dir_data ../data/devdata
Model checkpoint and output files will be stored under the generated "experiments" folder.
Citation
If you find the implementation useful, please consider citing the following paper:
Xiaotao Gu*, Zihan Wang*, Zhenyu Bi, Yu Meng, Liyuan Liu, Jiawei Han, Jingbo Shang, "UCPhrase: Unsupervised Context-aware Quality Phrase Tagging", in Proc. of 2021 ACM SIGKDD Int. Conf. on Knowledge Discovery and Data Mining (KDD'21), Aug. 2021
