|
|
自从ChatGPT火了之后，总是冒出「国内版ChatGPT」这种话题，似乎有碰瓷之嫌，但是稍微深入了解一下，就会知道国内的AI研究并不弱，比如下图是斯坦福大学统计的AI专利申请量（文章链接），可见中国似乎比美国还更厉害。
AI当然是一个大话题，可以与众多行业结合。从英语这个角度来说，AI的一大用途就是做语法分析、文章润色等，毕竟国内有应试英语的需要，即便是母语人士也有文章润色的需要，不然grammarly也不会火。
https://www.zhihu.com/video/1615299493481631744
我能想到的国内实际用到AI语法分析能力的场景大概有学习笔、作文批改等。毕竟学生在学英语的时候，时常需要分析某个句子的语法结构，每一句都去问老师也不现实，所以学习笔如果能够自动检测句子的语法结构，实在是有市场的。刚好之前测评过几款学习笔，其中就涉及语法部分，结论是只有网易有道词典笔X5的「有道神经网络」在语法分析方面表现突出，本文就来详细测评其语法分析能力。
之前是从「基本句型识别」「长难句分析」两个方面去测评ChatGPT的语法分析能力的，那么这次测评也一样操作，用同样的句子来测试，以结果说话。
基本句型识别
首先，我们常说英语中有5大基本句型（参考《英语语法体系》），那首先每种基本句型选一个最简单的句子来测试：
主谓：Birds can fly.
主谓宾：I love you.
主系表：I am happy.
主谓双宾：I gave you an apple.
主谓宾宾补：He made me angry. 网易有道词典笔X5的语法分析如下：
可见，「主谓 Birds can fly.」「主谓宾 I love you.」「主系表 I am happy.」的分析都完全正确✅。但是「主谓双宾 I gave you an anpple.」没有区分「direct object直接宾语」和「indirect object间接宾语」，而「主谓宾宾补 He made me angry.」则直接无法区分「宾语」和「宾补」，只是笼统将me angry视为「宾语」。
为了测试网易有道词典笔X5是不是真的不区分「宾语」和「宾补」，我又测试了更多「主谓宾宾补」句型的句子：
You can leave the door open.
They named the baby Andrew.
I have kept her waiting too long.
He ordered the anchor dropped.
由上图可见，网易有道词典笔X5将这些「主谓宾宾补」句型的句子都视为「主谓宾」，其中的「宾语 + 宾补」整体视为「宾语」。
小结一下：网易有道词典笔X5能够准确识别「主谓」「主谓宾」「主系表」等基本句型，但是不区分「主谓双宾」中的「直接宾语」与「间接宾语」，不区分「主谓宾宾补」中的「宾语」和「宾补」，都笼统将其视为「宾语」——这样做如果自成体系似乎也没问题，但是不太符合国内的英语语法教学实践。
请不要划走！虽然看到这里，你可能对网易有道词典笔X5失去兴趣，毕竟连基本句型都无法全部识别的AI有什么用呢，但是接下来的长难句分析，网易有道词典笔X5将扳回一局。
长难句分析
大多数同学对于AI语法分析的真正需求在于长难句分析，毕竟简单的句子自己已经能看懂了。上次测评ChatGPT的语法分析能力采用的是《英语语法长难句100》（可在EnglishBuddy公众号回复「长难句100」获取文档）中每个阶段的第一个句子（按原编号列出），共11句，涵盖高考、四级、六级、雅思、托福、SAT、专四、专八、GMAT、GRE等11门常见英语考试，这次测评网易有道词典笔X5同样采用这些句子。
1. All customers travelling on TransLink services must be in possession of a valid ticket before boarding. (2020年高考)
11. Our view is that if many people are living for longer, and are healthier for longer, then this will result in an inevitable redesign of work and life. (2019年6月四级一卷)
21. The project, in collaboration with a number of specialized IT and agricultural partners, had one goal in mind – to take a variety of academic and physical data sources from an agricultural environment and turn these into automatic predictive solutions for farmers that would assist them in making real-time decisions in the field. (2020年12月六级一卷)
31. It has become a grimly reliable annual ritual: every January the cost of travelling by train rises, imposing a significant extra burden on those who have no option but to use the rail network to get to work or otherwise. (2021年考研)
41. The tree is thickly branched with dense foliage of tough, dark green oval leaves, and produces small, yellow, bell-shaped flowers and pale yellow pear-shaped fruits. (剑桥雅思15)
51. However, a blindfolded snake strikes correctly at a dead rat that is pulled across the cage, provided the rat is warmer than the surroundings. (托福TPO65)
61. For the imitation to succeed in fooling us, it must resemble one or more works that we have been led to believe are undoctored originals. (SAT)
66. In Belgium, immigrants who lose jobs can receive almost two-thirds of their most recent wage in state benefits, which must make the hunt for a new job less urgent. (2019年专四)
76. Under the changes of weather it may look like marble or like sea water, black as slate in the fog, white as tufa in sunlight. (2019年专八)
86. First, minority firms risk expanding too fast and overextending themselves financially, since most are small concerns and, unlike large businesses, they often need to make substantial investments in new plants, staff, equipment, and the like in order to perform work subcontracted to them. (GMAT)
91. Black people were most influenced by those Euro-American tales whose functional meaning and aesthetic appeal had the greatest similarity to the tales with deep roots in their ancestral homeland. (GRE) 与上次一样，这11个句子未必都需要用到，只要能够比较准确地判断网易有道词典笔X5的语法分析水平即可。
<hr/>1. All customers travelling on TransLink services must be in possession of a valid ticket before boarding. (2020年高考)
完全正确✅。这里要说明的是这句话的谓语部分有两种分析：
① ...谓语must be in possession of + 宾语a valid ticket
② ...谓语must be + 表语in possession of a valid ticket 这两种分析都是对的，只是视角差异，并无对错与优劣之分。前者是ChatGPT采用的整体法，后者是网易有道词典笔X5的细致分析。
<hr/>
11. Our view is that if many people are living for longer, and are healthier for longer, then this will result in an inevitable redesign of work and life. (2019年6月四级一卷)
完全正确✅。注意will result in an inevitable redesign一处可以有两种分析：
① 整体法：...谓语will result in + 宾语an inevitable redesign
② 细分法：...谓语will result + 状语in an inevitable redesign 同上，这两种分析方法本身并无对错与优劣之分。从实用的角度来说，整体法更简洁。另外，细分法中的in an inevitable redesign到底是不是状语尚可争议。
<hr/>51. However, a blindfolded snake strikes correctly at a dead rat that is pulled across the cage, provided the rat is warmer than the surroundings. (托福TPO65)
完全正确✅。只要provided...处需要说明，这个结构可以总结为「主句，provided + 简单句」，比如：
A picnic will be very pleasant, provided (that) we get good weather. (Oxford)
In fact, going beyond compliance saves cost at the same time that it generates cash, provided that management adopts the new lean and green model. (2014 年 12 月六级三卷)
Provided that these surfaces are strategically located, the consequences of this combination can significantly increase the pollen-capture efficiency of a female reproductive organ. (GRE) 这个结构有两种分析：
① 整体法：将provided视为引导词，将「provided + 简单句」视为provided引导的状语从句
② 细分法：将provided视为非谓语动词做状语，将「(that) + 简单句」视为provided的宾语从句 参考《英语语法体系》2.3 状语从句的解释：
可见，网易有道词典笔X5采用的是细分法，虽然没有整体法简洁，但也不错。
<hr/>61. For the imitation to succeed in fooling us, it must resemble one or more works that we have been led to believe are undoctored originals. (SAT)
句子总体结构分析正确✅，但是定语从句that we have been led to believe are undoctored originals分析错误❌。这个定语从句理顺一下结构是：
We have been led to believe (宾语从句引导词that省略) one or more works are undoctored originals.
We have been led to believe one or more works to be undoctored originals. 现在one or more works在主句中，所以现在的定语从句的that=one or more works，既做believe的宾语，又做are undoctored originals的主语。可见，这个句子是一个杂糅的结构（并不错），一般语法分析为：
引导词that [插入语we have been led to believe] 谓语动词are 表语undoctored originals. <hr/>76. Under the changes of weather it may look like marble or like sea water, black as slate in the fog, white as tufa in sunlight. (2019年专八)
总体句子结构分析正确✅，但是有1处错误：
black as slate in the fog, white as tufa in sunlight 因为语义上来说it与marble or sea water在本句中是统一的，所以从语义上无法区分这个black...和white...是修饰主语it还是宾语marble or sea water，所以可以视为主语it的状语（也有视为定语的），或者视为marble or sea water的定语。总之不是网易有道词典笔以为的「宾语」。
<hr/>那其他句子呢？
21. The project, in collaboration with a number of specialized IT and agricultural partners, had one goal in mind – to take a variety of academic and physical data sources from an agricultural environment and turn these into automatic predictive solutions for farmers that would assist them in making real-time decisions in the field. (2020年12月六级一卷)
31. It has become a grimly reliable annual ritual: every January the cost of travelling by train rises, imposing a significant extra burden on those who have no option but to use the rail network to get to work or otherwise. (2021年考研)
41. The tree is thickly branched with dense foliage of tough, dark green oval leaves, and produces small, yellow, bell-shaped flowers and pale yellow pear-shaped fruits. (剑桥雅思15)
66. In Belgium, immigrants who lose jobs can receive almost two-thirds of their most recent wage in state benefits, which must make the hunt for a new job less urgent. (2019年专四)
86. First, minority firms risk expanding too fast and overextending themselves financially, since most are small concerns and, unlike large businesses, they often need to make substantial investments in new plants, staff, equipment, and the like in order to perform work subcontracted to them. (GMAT)
91. Black people were most influenced by those Euro-American tales whose functional meaning and aesthetic appeal had the greatest similarity to the tales with deep roots in their ancestral homeland. (GRE)
原来网易有道词典笔X5的AI语法分析功能有字符数限制，只能分析150个字符以内的句子——另外它也有最低单词数限制，最少要3个单词，所以上面的「主谓」句型没有采用上次测试ChatGPT时用的Birds fly. 而改为了Birds can fly.
总结
测试了以上句子，已经基本可以判断网易有道词典笔X5的语法分析水平了。
⚈ 可以比较准确识别基本句型，但是不区分「主谓双宾」中的「直接宾语」和「间接宾语」，不区分「主谓宾宾补」中的「宾语」和「宾补」，而是将其整体视为「宾语」，这种分析方法不符合一般英语语法教学实践，不过影响不算大，毕竟这么简单的句子一般也不需要去做AI语法分析。
⚈ 对于长难句，大部分句子都分析正确，往往采用细分法而非整体法。少数句子整体结构分析正确，但是有细节处分析错误，不过出错比ChatGPT少一些。
如果对比ChatGPT的语法分析能力，可以说：
⚈ 长难句分析能力网易有道词典笔X5总体略优于ChatGPT（不是比整个ChatGPT，只是比ChatGPT的语法分析能力），毕竟出错较少。
⚈ ChatGPT的扩展能力更强，毕竟可以重新设置提示词（prompt）追问。
⚈ 两者都可以作为语法分析的有益参考，但是都尚未达到独当一面、替代专业老师的程度。 |
|
发表于 2023-4-10 16:37:14