汉语词语离合现象识别研究

Summary

Chinese separable words (离合词) are words such as 帮忙 whose two morphemes can be split by inserted material (帮了我一个忙) while still expressing a single meaning. They complicate word segmentation and parsing. Earlier detectors used hand-written rules or feature templates over segmented, POS-tagged text, so tagging errors carried over into detection.

The authors treat detecting split two-character verbs as character-level sequence labeling with B/I/E/O tags, which places no limit on the inserted span. They re-annotated an existing news-based corpus (from Zhao Yuxi's 2019 thesis) so that one sentence can contain several split instances.

Their model, BERT_MASK + 2BiLSTMs + CRF, uses two encoders during training:

• one reads the original sentence;
• the other reads a copy in which the first and second morphemes are replaced by two different placeholder tokens. This pushes the model to learn the structure of the insertion rather than memorize frequent words.

At inference only the original-sentence encoder is used. The model reaches 98.76 F1, 2.85 points above the previous best cascade system. It also handles rare and unseen separable words well, which could help downstream segmentation, dependency parsing and CAMR annotation.

Key points

• Main results: P/R/F1 of 98.48/99.04/98.76%, against 94.59/97.26/95.91% for the reproduced KNN/SVM/CNN+LSTM+Attention cascade (Zhao Yuxi).
• Setup:
  • Encoder: Chinese BERT (Chinese_L-12_H-768_A-12), fine-tuned during training.
  • Masking: the two morphemes are replaced by [unused1] and [unused2].
  • Training: Adam, learning rate 2e-5 for BERT and 1e-4 for the other layers, decay of 0.7 per epoch, 40 epochs.
• Rare and unseen words:
  • Of 68 test-set words seen only once in training, 65 were detected (95.59%).
  • All 45 words absent from training were detected.
  • Two unannotated new cases (拄拐, 失球) were also found and judged correct on manual checking.
• Multiple instances per sentence: in 10 test sentences predicted to contain several split instances, 16 of the 20 instances were correct (80%).
• Comparison and ablation:
  • On 3,824 positive test sentences, the model linked the split morphemes correctly in 98.30% of cases, versus 95.69% for DDParser.
  • Removing MASK lowers F1 to 98.56%; with masking, precision drops slightly (−0.07) but recall rises (+0.46).