Consider the following table, generated from the 10M Character ROCLING corpus, a different corpus from the Academia Sinica corpus used by Huang and colleagues. As in Huang et al's table, it shows, for each component of the two-character county name, the mutual information between that character and the character 縣 `county'.

This was produced by a script that eliminates from consideration any examples where one of the two characters in a county name occurs directly before the character 縣. This has the desirable effect of eliminating the actual suoxie examples, as Huang et al did; but it is likely to err somewhat on the side eliminating too many examples. Is the resulting table a reasonable rendition of what's in the Huang et al papers, and if not what are the important differences?

台 東	台	0.285	東	1.402
台 北	台	0.285	北	1.173
花 蓮	花	0.623	蓮	1.242
彰 化	彰	1.483	化	1.129
苗 栗	苗	2.307	栗	2.985
台 中	台	0.285	中	1.004
澎 湖	澎	1.490	湖	0.991
雲 林	雲	1.718	林	1.608
台 南	台	0.285	南	1.573
桃 園	桃	0.986	園	0.911
高 雄	高	0.464	雄	1.018
南 投	南	1.573	投	-0.127
屏 東	屏	3.646	東	1.402
嘉 義	嘉	2.194	義	0.926
新 竹	新	0.405	竹	0.765
宜 蘭	宜	0.673	蘭	1.469

References

Huang, Chu-Ren, Kathleen Ahrens, and Keh-Jiann Chen. 1994. "A data-driven approach to psychological reality of the mental lexicon: Two studies on chinese corpus linguistics." In Language and its Psychobiological Bases, Taipei.

Huang, Chu-Ren, Wei-Mei Hong, and Keh-Jiann Chen. 1994. Suoxie: An information based lexical rule of abbreviation. In Proceedings of the Second Pacific Asia Conference on Formal and Computational Linguistics II, pages 49-52, Japan.