镜像图灵测试:古诗的机器识别

Mirror Turing Test:Machine Recognition of Ancient Chinese Poetry

古诗伴随着中华文化的历史进程不断发展,有着数千年的灿烂历史,古诗将丰富的情感、有内涵的灵魂和生动的形式完美结合,表现出了中华民族语言的力量.“自然语言处理是人工智能皇冠上的明珠”,用机器生成语言是机器智慧的核心体现,对机器的语言进行测试是图灵测试的重要内容,用机器生成的中国古代诗词已经可以初步通过图灵测试,在普通人面前得以瞒天过海.本文提出了“镜像图灵测试”框架,其主要设计思想是将图灵测试中的测试者由人更换为计算机,要求测试者在图灵测试的同等条件下对被测试的人和计算机进行识别,若测试计算机不能完成对被测试者的识别,则认为被测试的机器通过了镜像图灵测试.本文以机器生成的古诗和诗人创作的古诗为测试对象,以经过LDA主题模型调节的融合自注意力机制和切片LSTM网络的模型为测试机,设计了镜像图灵测试实验.实验将古诗分为写景、抒情以及爱国诗三类,为每类诗歌构建了8组数据集,共8万句古诗,采用了4种模型对24组数据集进行测试,利用测试机判别诗歌来自诗人还是机器,识别结果可达80%左右,实验结果显示,镜像图灵测试机可以对机器生成的诗歌进行识别,即机器生成的通过了图灵测试的诗歌并没有通过镜像图灵测试,说明了诗歌作为人类语言文明的结晶,是人脑情感最突出的反应,是诗人全身心的投入后的灵魂映射,在一定意义上是图灵可测的,即如果存在图灵可测的不完备性,那么诗歌这个人类语言的精华所在,就是突破这个图灵不完备性的关隘.本文提出的镜像图灵测试框架为后续图灵测试的研究提供了新的思路与方向.

With the continuous development of Chinese culture and thousands of years of splendid history,ancient poetry combines rich emotions,connotative souls and vivid forms perfectly,showing the power of Chinese language.Language is also an important research direction in the field of artificial intelligence.Using machine-generated language is the core embodiment of machine intelligence.Testing machine language is an important content of Turing test.The ancient Chinese poetry generated by machine has passed Turing test preliminarily and can deceive ordinary people.This paper puts forward the thought model of“Mirror Turing test”.Its main design idea is to replace the tester in Turing test with computer,and require the tester to identify the tested person and computer under the same conditions of Turing Test.If the test computer cannot complete the identification of the tested machine,it is considered that the tested machine has passed the Mirror Turing test.Considering that in the field of recognition,the ability of computer has surpassed that of human beings,it is more difficult for the poetry generator to pass the Mirror Turing test.In this paper,the machine-generated poetry and the poetry created by poets are taken as the test objects,and the model of integrating self-attention mechanism and slice LSTM network modified by LDA theme model is taken as the test machine,and the Mirror Turing test experiment is designed.In the experiment,ancient poetry is divided into three categories:landscape,lyric and patriotic poetry.Eight data sets are constructed for each category of poetry.Four models are used to test 24 groups of data sets.The test machine is used to determine whether poetry comes from a poet or a machine,and the recognition results can reach about 80%.The recognition accuracy of lyric poetry is relatively low,which shows that when the poetry contains emotion and has more emotion,it will bring more difficulties to the test of the machine.Behind the poetry is the poet’s soul,which can’t be imitated by the current machine in any case.The research of machine emotion is the most far-reaching in front of artificial intelligence research.The experimental results show that the Mirror Turing test machine can identify the poetry generated by the machine,that is,the poetry generated by the machine that has passed the test The poetry of Turing test cannot pass the Mirror Turing test,which shows that poetry,as the crystallization of human language civilization,is still difficult to be surpassed by machines so far.In the field of poetry generation,Turing test may be incomplete.The problem of Mirror Turing test is the game between the testing machine and the tested machine.Human beings can improve the requirements of Mirror Turing test by constantly improving the testing machine.Considering that the recognition level of the machine can surpass that of human beings,the computer that has passed the Mirror Turing test is far more intelligent than human beings.The Mirror Turing test framework proposed in this paper provides new ideas and directions for the subsequent research on Turing testing.