基于短语及依存的标注规则和短语识别算法研究

Summary

Most Chinese dependency parsing runs on word-segmented text and relies on end-to-end supervised models. The authors identify two weaknesses in this approach. First, annotation schemes such as LTP and Baidu's use dozens of POS and relation labels, which makes annotation slow and error-prone for non-linguists. Second, sequence-labelling models assign one tag per span and so cannot represent nested constituents.

To address this, the authors take the phrase rather than the word as the basic unit and define a simplified scheme with seven nestable phrase types, plus dependency links between phrases. Using this scheme they built the Chinese Phrase Window Dataset (CPWD).

They also propose the Syntax Window Model (SWM), which borrows its design from Faster R-CNN object detection. SWM enumerates every candidate span, predicts whether each span is a phrase, adjusts its start and end boundaries with small offsets, and then classifies the phrase type. This lets it recover overlapping phrases and their dependencies jointly.

On CPWD, SWM beats a conventional end-to-end baseline by over 1 F1 point. The same method also won first place in the CCL2018 Chinese metaphor sentiment analysis task. The work offers an easier-to-annotate, nesting-aware alternative to fine-grained Chinese parsing schemes.

Key points

• Seven phrase categories, each with its own bracket notation: noun, verb, numeral-classifier, prepositional, conjunction, modal-particle, and clause. Pivotal, serial-verb, and subject/object clauses are all marked as clauses.
• CPWD corpus:
  • 45,000 mostly modern Chinese sentences drawn from dialogue, news, law, policy texts, and fiction.
  • Sentences are capped at 50 characters.
  • The data is split into 35,000 training, 5,000 validation, and 5,000 test sentences.
• SWM architecture:
  • A feature encoder (LSTM or BERT; experiments mention a 2-layer BiLSTM) feeds a window-localisation network.
  • This network scores all N(N+1)/2 spans and regresses start/end offsets of at most one character.
  • Training combines a cross-entropy loss for phrase detection with an RMSE loss for the offsets.
• Results: SWM improves F1 by more than 1 point over a traditional end-to-end model on CPWD. In CCL2018 metaphor sentiment analysis, the method added more than 1 F1 point and ranked first.