内容简介：NUBIA is a SoTA evaluation metric for text generation. It stands for NeUral Based Interchangeability Assessor. In addition to returning an interchangeability score, NUBIA also returns scores for semantic relation, contradiction, irrelevancy, logical agreem
What is NUBIA?
NUBIA is a SoTA evaluation metric for text generation. It stands for NeUral Based Interchangeability Assessor. In addition to returning an interchangeability score, NUBIA also returns scores for semantic relation, contradiction, irrelevancy, logical agreement, and grammaticality.
View Paper | Repo | Notebook |Nubia Backstory | FAQ
How does it work?
Nubia is composed of three modules. The first is neural feature extraction. The three main neural features that power the metric are semantic similarity, logical inference, and sentence legibility. These are extracted by exposing layers from powerful (pretrained) language models: RoBERTa STS for semantic similarity, RoBERTa MNLI for logical inference, and GPT-2 for sentence legibility.
The second module is the aggregator. This module is trained to approximate a function mapping input neural features to a quality score that reflects how interchangeable the sentences are. The objective is to come as close as possible to human evaluation.
The final module is calibration. This is necessary because the aggregator is not bound between 0 and 1, nor does a regressed score comparing a reference sentence with itself always ouput 1. So to calibrate, the output is normalized against the score of the reference sentence compared with itself, and bound between 0 and 1.
How does it perform?
We have been able to achieve SoTA performance in assessing the quality of machine translation and image captioning. Results below are the Absolute Pearson correlations with segment-level human judgments on WMT17 to-English translations.
Results below are the Kendall’s Tau correlations with human judgment on the Flickr 8K dataset.
Nubia is especially powerful against BLEU and ROUGE. The following are the score and label graphs of Nubia, ROUGE-L and BLEU for the entire WMT-2017 segment level set.
Can you show me an example?
The following two sentences mean the same thing:
s1, s2 = "He agreed to a proposal of mine.","He gave his agreement to my proposal."
but the BLEU, ROUGE and BERTScore are low (0.43, 0.28, 0.63) whereas the Nubia score is 0.95
Rouge_1 is: 0.4285714235714286 Rouge_2 is: 0.0 BlEU is: 0.2802636449867835 BERTScore is: (0.5967282652854919, 0.6308560967445374, 0.6143679618835449) Nubia Score: 0.9504227034094436 Semantic relation: 4.672990322113037/5.0 Percent chance of contradiction: 0.26220036670565605% Percent chance of irrelevancy or new information: 0.5306123290210962% Percent chance of logical agreement: 99.20719265937805%
Another example shows how the metrics can keep up despite differences in sentence structure:
When "Impossible!" was changed to "Totally feasible!" , the contradiction score went from 0.073% to 99.5% , even though sentence s1 is structured very differently.
For more examples, check out the repo and colab notebook . To better understand the methodology and motivation check out the paper andNubia’s backstory.
以上所述就是小编给大家介绍的《Nubia: A SoTA evaluation metric for text generation》，希望对大家有所帮助，如果大家有任何疑问请给我留言，小编会及时回复大家的。在此也非常感谢大家对 码农网 的支持！
猜你喜欢:
本站部分资源来源于网络，本站转载出于传递更多信息之目的，版权归原作者或者来源机构所有，如转载稿涉及版权问题，请联系我们。
数据结构与算法分析（Java版）（英文原版）
(美)Clifford A.Shaffer / 电子工业出版社 / 2002-5 / 39.00元
《数据结构与算法分析(C++版)(第2版)》采用程序员最爱用的面向对象C++语言来描述数据结构和算法，并把数据结构原理和算法分析技术有机地结合在一起，系统介绍了各种类型的数据结构和排序、检索的各种方法。作者非常注意对每一种数据结构的不同存储方法及有关算法进行分析比较。书中还引入了一些比较高级的数据结构与先进的算法分析技术，并介绍了可计算性理论的一般知识。本版的重要改进在于引入了参数化的模板，从而提......一起来看看 《数据结构与算法分析（Java版）（英文原版）》 这本书的介绍吧!
