I strongly recommend a project: Chinese NLP, which is created and maintained by the natural language processing team of Didi Artificial Intelligence Laboratory . This project has carefully sorted out Chinese natural language processing related tasks, data sets and current best results. Quite complete.
Project homepage: https://chinesenlp.xyz
Github: https://github.com/didi/ChineseNLP
This project currently contains 18 Chinese natural language processing tasks, as well as one other category:
Below each subtask, the related task background, examples, evaluation indicators, related data sets and current best results will be introduced in detail. Taking Chinese word segmentation as an example, in addition to the backoff2005 data set we are familiar with, there are some other data sources:
Let's look at the machine translation task again. About the evaluation index, the description is quite detailed:
Direct evaluation (manual evaluation). An annotator on Amazon Mechnical Turk will see a system-generated translation and a human translation, and then answer the question: "How accurately does the system translation express the meaning of the human translation? "
Blue score ( Papineni et al 02 ).
Case sensitive vs. case insensitive
Brevity penalty trigger condition: When the machine translation result is shorter than the shortest reference translation (reference) or shorter than the closest reference translation (reference).
brevity penalty: A coefficient used to penalize machine translation results that are shorter than the reference translation.
The standard Bleu calculation process will first tokenize the reference translation and machine translation results.
If Chinese is the target language, use character level {1,2,3,4}-gram matching.
Use Bleu-n4r1 to evaluate when there is only one human reference translation translation.
Bleu-n4r4: Word level {1,2,3,4}-gram matching, compared with 4 human reference translations
There are many important variants of the standard Bleu:
NIST. A variant of Bleu that gives higher weight to rare n-grams.
TER (Translation Edit Rate). Edit distance between computer translation and human reference translation.
BLEU-SBP ((Chiang et al 08)[ http://aclweb.org/anthology/D08-1064] ). Solved the decomposability of Bleu, and achieved a compromise between Bleu and word error rate.
HTER. The number of edits required for a good translation (the number of edits).
In terms of corpus resources related to machine translation, it also includes the familiar United Nations corpus and AI Challenger:
Students who are interested in other related tasks can refer to it by themselves. This is a pretty good reference point for understanding current Chinese NLP related tasks. Thanks to the students who built and maintained this project.
Click to read the original text and go directly to the project homepage.
