MFE-NER: Multi-feature Fusion Embedding for Chinese Named Entity Recognition

TL;DR

This work addresses the pervasive character substitution issue in Chinese Named Entity Recognition by introducing MFE-NER, a lightweight fusion of semantic, glyph, and phonetic features. Glyph information is captured via Five-Strokes, a structure-based encoding, while phonetic similarity is modeled with Trans-pinyin, which blends Pinyin with international phonetics. The approach attaches additional glyph and phonetic embeddings to the standard semantic embedding through simple input-layer concatenation, incurring modest overhead. Empirical results on substitution and standard NER datasets show enhanced robustness to substitutions and modest overall gains, particularly in social media contexts, indicating practical value for real-world Chinese NLP systems.

Abstract

In Chinese Named Entity Recognition, character substitution is a complicated linguistic phenomenon. Some Chinese characters are quite similar as they share the same components or have similar pronunciations. People replace characters in a named entity with similar characters to generate a new collocation but referring to the same object. As a result, it always leads to unrecognizable or mislabeling errors in the NER task. In this paper, we propose a lightweight method, MFE-NER, which fuses glyph and phonetic features, to help pre-trained language models handle the character substitution problem in the NER task with limited extra cost. Basically, in the glyph domain, we disassemble Chinese characters into Five-Stroke components to represent structure features. In the phonetic domain, an improved phonetic system is proposed in our work, making it reasonable to describe phonetic similarity among Chinese characters. Experiments demonstrate that our method performs especially well in detecting character substitutions while slightly improving the overall performance of Chinese NER.

Figures (5)

• Figure 1: The substitution example of a Chinese word. On the left is a famous place in Shanghai. On the right is a new word after character substitution, which is more like a person or a brand.
• Figure 2: The structure of MFE-NER.
• Figure 3: The 'Five-Strokes' representation of the Chinese characters, 'pu3' (somewhere close to a river). 'Five-Strokes' divides the character into four character roots ordered by writing custom so that the structure similarity can be denoted.
• Figure 4: The 'Pinyin' form and standard form of two Chinese characters, 'cao3' (grass) on the left and 'zao3' (early) on the right.
• Figure 5: An example is drawn from our dataset, 'he wants to go to Dapuqiao'. The sentence at the top of the figure is the original sentence, while the sentence at the bottom is after character substitution. The model using MFE-NER gives the correct prediction.