219mze 发表于 2024-8-18 23:40:21

程羽黑:​人工智能诗歌论


    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://mmbiz.qpic.cn/mmbiz_png/ffC4nP2WfckTtoA5f7rB3BsMQYJWUtXlN9xLiahUmFuNNtDKNNyElLE6Hk2bU9xEYyhsz72pJGTfBu5O9Ds3Xgg/640?wx_fmt=png&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1" style="width: 50%; margin-bottom: 20px;"></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;">人工智能诗歌论</span></strong></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">2017<span style="color: black;">年</span>5<span style="color: black;">月,微软(亚洲)互联网工程院的</span>AI“<span style="color: black;">小冰</span>”<span style="color: black;">推出了世界上<span style="color: black;">第1</span>部完全由人工智能创作的诗集《阳光失了玻璃窗》。小冰学习了</span>519<span style="color: black;">位汉语新诗诗人的作品,经过上万次迭代(</span>iteration<span style="color: black;">),在</span>2760<span style="color: black;">小时内作诗万首,其中</span>139<span style="color: black;">首被选入该诗集。据微软小冰的首席<span style="color: black;">专家</span>宋睿华介绍,小冰作诗的原理是</span>“<span style="color: black;">触景生情</span>”<span style="color: black;">:<span style="color: black;">经过</span>图像识别技术,找到与<span style="color: black;">照片</span>对应的词语,<span style="color: black;">从而</span>获取<span style="color: black;">相关</span>性较高的词组,采用循环神经网络(</span>Recurrent NeuralNetworks<span style="color: black;">),拓展出<span style="color: black;">更加多</span>表达</span><span style="color: black;"></span><span style="color: black;">。虽然总体来看,小冰创作的诗歌语句颇为粗糙,但仍有不少可圈可点之处,有<span style="color: black;">有些</span>比喻句相当巧妙。<span style="color: black;">据说</span>,小冰还在天涯、豆瓣、简书等平台上,以</span>“<span style="color: black;">骆梦</span>”<span style="color: black;">、</span>“<span style="color: black;">风的指尖</span>”<span style="color: black;">等笔名</span>“<span style="color: black;">伪装</span>”<span style="color: black;">成人类<span style="color: black;">发布</span>了<span style="color: black;">海量</span>诗歌</span><span style="color: black;"></span><span style="color: black;">。</span>2018<span style="color: black;">年</span>8<span style="color: black;">月,科技博客</span>BoyGenius Report<span style="color: black;"><span style="color: black;">报告</span>,</span>IBM<span style="color: black;"><span style="color: black;">机构</span>在澳大利亚的<span style="color: black;">开发</span>团队与多伦多大学、墨尔本大学合作,<span style="color: black;">研发</span>出了一种具备写诗能力的“算法”。该</span>AI<span style="color: black;">接受了近</span>3000<span style="color: black;">首十四行诗(</span>sonnet<span style="color: black;">)的训练,学习词语的搭配,<span style="color: black;">而后</span><span style="color: black;">按照</span>格律创作诗歌。<span style="color: black;">科研</span>人员将其作品与人类的诗歌混合在<span style="color: black;">一块</span>,让志愿者寻找,结果正确率<span style="color: black;">仅有</span>一半,可见它的诗歌在<span style="color: black;">必定</span>程度上已能以假乱真</span><span style="color: black;"></span><span style="color: black;">。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">人工智能诗歌创作方兴未艾,<span style="color: black;">经过</span>深度学习训练的人工神经网络(</span>Artificial NeuralNetwork<span style="color: black;">)使</span>AI<span style="color: black;">“诗人”<span style="color: black;">得到</span>了初步的智力,相比以往<span style="color: black;">按照</span>格律和韵脚随机填充词汇的<span style="color: black;">所说</span>“作诗机”</span>——<span style="color: black;"><span style="color: black;">咱们</span>应<span style="color: black;">叫作</span>之为“伪人工智能”(</span>Pseudo AI<span style="color: black;">)</span>——<span style="color: black;">有了本质的飞跃。<span style="color: black;">那样</span>,<span style="color: black;">此刻</span>的人工智能在诗歌创作<span style="color: black;">行业</span><span style="color: black;">是不是</span>已走上了通畅无碍的阳关大道?<span style="color: black;">最少</span>有四个问题仍待<span style="color: black;">处理</span>。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;"><span style="color: black;">1、</span>人工智能创作诗歌的<span style="color: black;">道理</span>问题</span></strong></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">人工智能已能写出符合语法的句子,然而符合语法只是具备<span style="color: black;">道理</span>的必要<span style="color: black;">要求</span>而非充分<span style="color: black;">要求</span>。乔姆斯基指出,“无色的绿色想法在狂热地<span style="color: black;">睡觉</span>”(</span>The colorless greenideas sleep furiously.<span style="color: black;">)。<span style="color: black;">这般</span>的句子虽然“符合语法”(</span>grammatical<span style="color: black;">),却<span style="color: black;">不可</span>说是“有<span style="color: black;">道理</span>”(</span>meaningful<span style="color: black;">)或“有<span style="color: black;">寓意</span>”(</span>significant<span style="color: black;">)的</span><span style="color: black;"></span><span style="color: black;">。诗歌固然不必追循<span style="color: black;">平常</span>语言的“<span style="color: black;">道理</span>”或“<span style="color: black;">寓意</span>”,<span style="color: black;">乃至</span>有<span style="color: black;">所说</span>“无理而妙”者,然而其中的“妙”仍须诉诸人类微妙的感觉和经验。<span style="color: black;">咱们</span><span style="color: black;">能够</span>参考苏轼的话:“诗以奇趣为宗,反常合道为趣。”</span><span style="color: black;"></span><span style="color: black;">这一“趣”(“妙”)虽然“反常”,却须“合道”,正如大自然中随机形成的奇木异石,<span style="color: black;">仅有</span>极少部分能被人类视为<span style="color: black;">宝贝</span>。有一个著名的“无限猴子定理”(</span>infinite monkey&nbsp;theorem<span style="color: black;">):给猴子一台打字机,让它随机敲击键盘,只要给它无限的时间,它迟早能打出任何文本,<span style="color: black;">包含</span>莎士比亚的全集。这<span style="color: black;">必定</span>理<span style="color: black;">本来</span>只能被视为纯粹的假设,但在电脑<span style="color: black;">显现</span>后,它有了实现的可能。早在</span>1976<span style="color: black;">年,数学家就<span style="color: black;">经过</span>计算机的高速运算能力,用“粗暴”的穷举法<span style="color: black;">处理</span>了“四色猜想”(</span>four-color conjecture<span style="color: black;">)。将这种穷举法用于诗歌创作并不更加困难。前举的“伪人工智能”作诗机曾在网络上风靡一时,其原理<span style="color: black;">亦</span>是在<span style="color: black;">按照</span>设定的规则随机生成文本后由人类采择。即便具备了初步智能的微软“小冰”,<span style="color: black;">亦</span>仍需人类从它创作的万首诗中<span style="color: black;">选择</span></span>139<span style="color: black;">首。<span style="color: black;">倘若</span>人工智能创作的诗句<span style="color: black;">必须</span>人类筛选,<span style="color: black;">那样</span>它还<span style="color: black;">不可</span>算真正<span style="color: black;">道理</span>上的作者,而只是辅助人类写诗的语料库,其<span style="color: black;">功效</span>一如千年前就已<span style="color: black;">显现</span>的类书。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">问题的关键在哪里?一个好的诗人<span style="color: black;">必要</span>具备分辨好坏的能力,<span style="color: black;">亦</span><span style="color: black;">便是</span>审美的能力,而这不仅关乎诗歌文本的特性,更关乎文本之外的体验。换言之,要理解诗的好处,<span style="color: black;">不可</span>仅<span style="color: black;">经过</span>文本本身</span>——<span style="color: black;"><span style="color: black;">咱们</span>不妨参考苏轼在《书黄子思诗集后》中引用的司空图语:“梅止于酸,盐止于咸,<span style="color: black;">膳食</span>不可无盐梅,而其美常在咸酸之外。”</span><span style="color: black;"></span><span style="color: black;"><span style="color: black;">倘若</span>料理人自己<span style="color: black;">无</span><span style="color: black;">敏锐</span>的舌头,<span style="color: black;">那样</span><span style="color: black;">咱们</span>很难设想他<span style="color: black;">到底</span><span style="color: black;">怎样</span>理解“咸酸之外”的美味,更难设想他<span style="color: black;">怎样</span>做出<span style="color: black;">这般</span>的美味。前举</span>IBM<span style="color: black;">推出的算法所创作的十四行诗几乎能以假乱真,这固然是一个可喜的进步,但<span style="color: black;">咱们</span><span style="color: black;">亦</span>有理由假设,<span style="color: black;">倘若</span>应试者是优秀的诗人,理当辨千剑而识器</span>——<span style="color: black;">就像好的古董商能凭微妙的感觉鉴别<span style="color: black;">一般人</span>眼中的<span style="color: black;">水货</span><span style="color: black;">同样</span>。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">当然,<span style="color: black;">咱们</span>应对前景保持<span style="color: black;">阳光</span>,随着技术发展,人工神经网络将能精确模仿人类神经元的刺激——反应(</span>stimulus–response<span style="color: black;">)</span>——<span style="color: black;">让</span>AI<span style="color: black;">具备人类的情感,并非天方夜谭。但或许仍有新的问题<span style="color: black;">必须</span><span style="color: black;">处理</span>:人类始终在理解他人的意图,诗歌的文本<span style="color: black;">必须</span>潜在的读者,</span>AI<span style="color: black;"><span style="color: black;">是不是</span>具备产生这种“主体间性”(</span>intersubjectivity<span style="color: black;">)的可能?这关乎两个疑问:一,</span>AI<span style="color: black;">之间能否<span style="color: black;">做为</span>对等的主体进行交流?二,</span>AI<span style="color: black;"><span style="color: black;">是不是</span>可能<span style="color: black;">拥有</span>刺激理解意图产生的必要遮障?印度的吠檀多学派主张去除无明遮障,使个体小我重归于梵,从某种<span style="color: black;">道理</span>上说,</span>AI<span style="color: black;">一<span style="color: black;">起始</span>就达到了“我即梵”的境界,并无“遮障”和“小我”为患,而</span>AI<span style="color: black;">要创作诗歌,反而要学习人类这种“缺陷”</span>——<span style="color: black;">套用佛教的话,这是一种有意的“退转”(</span>vaivartika<span style="color: black;">)</span>——<span style="color: black;">或许诗歌正是对人类这一“缺陷”的<span style="color: black;">弥补</span>。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;"><span style="color: black;">2、</span>人工智能创作诗歌的可理解性问题</span></strong></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">Herbert Paul Grice<span style="color: black;">的“会话隐含理论”(</span>conversationalimplicature theory<span style="color: black;">)认为,会话<span style="color: black;">背面</span>必有不随形式改变的<span style="color: black;">知道</span>含义。诗歌<span style="color: black;">一样</span>遵循类似的规则。诗人将某种意图埋藏于诗中,读者在阅读的过程中追索诗句<span style="color: black;">背面</span>的所指。“丸之走盘,横斜曲直,计于临时,不可尽知。其必可知者,是丸之<span style="color: black;">不可</span>出于盘<span style="color: black;">亦</span>。”</span><span style="color: black;"></span><span style="color: black;">诗人的意图<span style="color: black;">怎样</span>,“不可尽知”</span>——<span style="color: black;">或许永远不会有读者能够准确把握</span>——<span style="color: black;">但诗人有意图却是“必可知”的事实。威廉</span>·<span style="color: black;">燕卜荪在《朦胧的七种类型》中写道:“诗人将两个<span style="color: black;">描述</span>放在<span style="color: black;">一块</span>,似乎二者是相互联系的,而读者则被迫去<span style="color: black;">思虑</span><span style="color: black;">她们</span>的联系,还得自己去设想诗文<span style="color: black;">为何</span><span style="color: black;">选取</span>了这些事实。他会设想出<span style="color: black;">各样</span><span style="color: black;">原由</span>,并在脑中将它们<span style="color: black;">摆列</span>起来。这是诗歌语言在运用方面的基本事实。”</span><span style="color: black;"></span><span style="color: black;">一旦取消<span style="color: black;">描述</span>之间相互联系的预设,读者理解的企图便会落空。换言之,<span style="color: black;">无</span>作者意图的诗歌本质上是不可理解的,哪怕它们在字面上<span style="color: black;">能够</span>理解。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">早在人类进入信息时代前,便有一个绝佳的例子说明这一点。二十世纪初,法国曾流行一种名为“精致的尸体”(</span>exquisite corps<span style="color: black;">)的游戏</span>——<span style="color: black;">大致的流程是:参与者们在互不知晓的<span style="color: black;">状况</span>下,依次写下词语,<span style="color: black;">而后</span>组合成句。<span style="color: black;">据述</span>在某次游戏中,得到了“精致的尸体会喝下新的葡萄酒”(</span>Le cadavre exquis boirale vin nouveau.<span style="color: black;">)的“诗句”</span>——<span style="color: black;">这句“诗”确实意象诡丽,饶有波德莱尔的风味,“</span>cadaver<span style="color: black;">”一词在法语<span style="color: black;">白话</span>中又有“喝空的酒瓶”之义,一语双关,颇富妙趣,然而它是随机偶合的产物,而非某个诗人精心结撰的成果。<span style="color: black;">咱们</span>固然<span style="color: black;">能够</span>说,诗的好坏在其本身,但这一问题仍牵涉到诗的经典定义</span>——<span style="color: black;">中文“诗”义在传世文献中最早见于《毛诗序》:“诗者,志之所之<span style="color: black;">亦</span>。在心为志,发言为诗。”</span><span style="color: black;"></span><span style="color: black;">而英语的“</span>poem<span style="color: black;">”一词源于希腊语的</span>poema<span style="color: black;">,意指“所造之物”</span>——<span style="color: black;">无论中国的“志之所之”还是希腊的“所造之物”,都强调了诗歌是某种意图的产物而非任意的组合。《孟子</span>·<span style="color: black;">万章上》:“故说诗者,不以文害辞,不以辞害志;以意逆志,是为得之。”</span><span style="color: black;"></span><span style="color: black;"><span style="color: black;">倘若</span>诗是任意的组合,那就无志可逆了。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">日前</span>的人工智能尚不具备自主<span style="color: black;">认识</span>,写出的“诗”仅仅是符号的集合,<span style="color: black;">无</span>所指,<span style="color: black;">无</span>隐喻,读者的一切追索将成<span style="color: black;">白费</span>,就像人工合成的语音,纵然模仿得惟妙惟肖,<span style="color: black;">背面</span>却无真情实感</span>——<span style="color: black;"><span style="color: black;">咱们</span>在<span style="color: black;">晓得</span>真相的<span style="color: black;">状况</span>下,恐怕很难为之感动。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">人们有理由预期具备自我<span style="color: black;">认识</span>的“强人工智能”(</span>strong AI<span style="color: black;">)<span style="color: black;">显现</span>,届时<span style="color: black;">这般</span>的</span>AI<span style="color: black;">或许能创作出有意图的、真正的诗。它们<span style="color: black;">到底</span>会是<span style="color: black;">怎么样</span>的呢?<span style="color: black;">此刻</span>预言为时尚早,但<span style="color: black;">倘若</span>它们依然基于对人类诗歌的模仿,那就不禁让人想起那则著名的童话</span>——<span style="color: black;">小红帽听到了外婆的声音,打开门却看见了一头狼。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;"><span style="color: black;">3、</span>人工智能创作诗歌的创造性问题</span></strong></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">诗歌<span style="color: black;">区别</span>于类型文学,<span style="color: black;">能够</span>提炼出某种模式<span style="color: black;">海量</span>复制。相反,诗歌强调独创性,强调“陈言之务去”,一首好诗的价值<span style="color: black;">常常</span><span style="color: black;">表现</span>在前所未有上,一个好诗人的价值<span style="color: black;">常常</span><span style="color: black;">表现</span>在独树一帜上。试将诗歌与围棋做一对比。围棋是人工智能<span style="color: black;">善于</span>的<span style="color: black;">行业</span>,自从</span>2016<span style="color: black;">年阿尔法狗(</span>AlphaGo<span style="color: black;">)击败李世石,人类便失去了这一<span style="color: black;">行业</span>的王冠。阿尔法狗采用卷积神经网络(</span>Convolutional Neural Networks<span style="color: black;">),<span style="color: black;">经过</span>蒙特卡洛搜索树(</span>MonteCarlo tree search<span style="color: black;">)进行决策优化,在模仿人类和自我对弈中<span style="color: black;">提高</span>棋力。</span>2017<span style="color: black;">年,</span>Google<span style="color: black;">推出阿尔法狗的进化版阿尔法元(</span>AlphaGo Zero<span style="color: black;">)。阿尔法元<span style="color: black;">运用</span>残差网络(</span>Residual Network<span style="color: black;">),不须学习人类旧谱,仅<span style="color: black;">经过</span>自我对弈训练棋力。它在与阿尔法狗的对弈中创造了</span>100<span style="color: black;">:</span>0<span style="color: black;">的压倒性战绩。无论是阿尔法狗还是阿尔法元,其原理都是对棋路的破解</span>——<span style="color: black;">无论棋路来自人类还是人工智能本身。这与诗歌<span style="color: black;">区别</span>。创造一流的诗歌,<span style="color: black;">必须</span>避免<span style="color: black;">招数</span>;换言之,围棋是在道路中寻找道路,诗歌是在道路外开辟道路。文首所举的两例</span>AI<span style="color: black;">“诗人”采用的途径却是对已有诗歌的模仿,这与真正优秀的诗作<span style="color: black;">必须</span>创新这一点显然背道而驰。<span style="color: black;">另一</span>,诗歌的创新未必<span style="color: black;">寓意</span>着一空依傍。杰出的诗人多在继承前人优点的<span style="color: black;">基本</span>上别开生面,<span style="color: black;">怎样</span>把握继承和创造之间的“微妙分寸”(</span>nuance<span style="color: black;">),对</span>AI<span style="color: black;">“诗人”<span style="color: black;">来讲</span>是一个严峻的挑战。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">围棋与诗歌面对的信息量<span style="color: black;">亦</span>有巨大的差别。围棋是<span style="color: black;">繁杂</span>的简单问题,诗歌是简单的<span style="color: black;">繁杂</span>问题。围棋本身是一种抽象</span>——<span style="color: black;">棋子和棋盘上的交叉点是仅仅<span style="color: black;">表率</span>相对位置的几何点,不<span style="color: black;">拥有</span>任何其他内涵。<span style="color: black;">因此</span>尽管围棋的计算量<span style="color: black;">非常</span>浩大,但全都<span style="color: black;">出现</span>在边界清楚、数量固定的规则层面,比无限且难以分割的现实容易计算得多。一言以蔽之,人工智能高速运算的<span style="color: black;">优良</span>正适合围棋这种规则清晰、计量标准统一的游戏。诗歌牵涉的层面则极为<span style="color: black;">繁杂</span>。<span style="color: black;">首要</span>,诗歌由符号<span style="color: black;">构成</span>,但正如上文所示,“味在咸酸之外”,符号的指涉并不止于符号层面</span>——<span style="color: black;">它们通向并反馈现实,现实又随时间、标准、个体、情感的变化而变化,换言之,诗歌创作<span style="color: black;">不仅</span><span style="color: black;">无</span>清晰的规则,<span style="color: black;">乃至</span><span style="color: black;">亦</span>缺乏稳定的规则(除了“格律”那样某类诗体所需的规则)。优秀的诗人之<span style="color: black;">因此</span>能够在<span style="color: black;">这般</span>的“混沌”(</span>chaos<span style="color: black;">)中游刃有余,依靠的是微妙的直觉,而就<span style="color: black;">日前</span>来看,人工智能尚无望达到<span style="color: black;">这般</span>的水准。除非人类<span style="color: black;">把握</span>了某种“黑箱”(</span>black box<span style="color: black;">),能够自发性地模仿“<span style="color: black;">繁杂</span>适应系统”(</span>complex adaptive systems<span style="color: black;">)</span>——<span style="color: black;"><span style="color: black;">咱们</span><span style="color: black;">能够</span>推想“强人工智能”<span style="color: black;">特别有</span>可能正是这种“黑箱”</span>——<span style="color: black;">否则以人类当前的信息处理能力,设计出真正<span style="color: black;">道理</span>上的人工智能“诗人”几乎是不可能的。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;"><span style="color: black;">4、</span>人工智能创作诗歌的必要性问题</span></strong></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">诗歌并非生活的<span style="color: black;">必要</span>品,<span style="color: black;">亦</span>不以量取胜,又无<span style="color: black;">非常</span><span style="color: black;">知道</span>的<span style="color: black;">评估</span>标准,有<span style="color: black;">无</span>必要花费巨大的代价<span style="color: black;">研发</span>人工智能写诗的能力?人类<span style="color: black;">研发</span>人工智能<span style="color: black;">常常</span>基于工具性的目的,而诗歌创作<span style="color: black;">常常</span>基于情感的<span style="color: black;">拜托</span>。这一问题或许可做如下类比:前者如<span style="color: black;">保母</span>,后者如母亲,<span style="color: black;">研发</span></span>AI<span style="color: black;"><span style="color: black;">保母</span>有必要,<span style="color: black;">研发</span></span>AI<span style="color: black;">母亲<span style="color: black;">是不是</span>有必要?后者本身与人类的属性相嵌合:人们寄望于后者的<span style="color: black;">重点</span>不是工具价值,而是情感价值,<span style="color: black;">因此</span>对后者<span style="color: black;">来讲</span>,人类的身份至关重要,绝非可有可无</span>——<span style="color: black;">一个具备<span style="color: black;">保母</span>功能的</span>AI<span style="color: black;">能够代替<span style="color: black;">保母</span>,一个具备母亲功能的</span>AI<span style="color: black;">则不可能代替母亲。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">更重要的是,<span style="color: black;">科研</span>人工智能创作诗歌有什么好处呢?<span style="color: black;">日前</span>人工智能“诗人”尚<span style="color: black;">处在</span>模仿已有诗歌的<span style="color: black;">周期</span>,对诗歌的开拓与发展<span style="color: black;">来讲</span>并<span style="color: black;">无</span><span style="color: black;">显著</span>的助益。而对人工智能<span style="color: black;">来讲</span>,能够在真正<span style="color: black;">道理</span>上创作诗歌的前提是主体<span style="color: black;">认识</span>、纯熟地理解和<span style="color: black;">运用</span>自然语言的能力和情感体验的能力。<span style="color: black;">咱们</span><span style="color: black;">能够</span>设想,一旦<span style="color: black;">持有</span>主体<span style="color: black;">认识</span>、具备卓越的语言和情感能力的强人工智能<span style="color: black;">显现</span>,诗歌创作便是题中应有之义。“古诗云:‘谁能思不歌,谁能饥不食。’诗词者,物之不得其平而鸣者<span style="color: black;">亦</span>。”</span><span style="color: black;"></span><span style="color: black;"><span style="color: black;">倘若</span><span style="color: black;">咱们</span>认可王国维的这一观点,<span style="color: black;">那样</span>诗歌本质上正是情感“不得其平”而生成的表达欲望。正如肉体会饥饿<span style="color: black;">同样</span>,<span style="color: black;">这般</span>的欲望是“不教而能”的,而一旦<span style="color: black;">持有</span><span style="color: black;">这般</span>的欲望,凭借语言能力和人工智能本身的存储和运算<span style="color: black;">优良</span>,</span>AI<span style="color: black;">距离创作真正的诗歌<span style="color: black;">仅有</span>一步之遥,举足可至。换言之,写诗的能力<span style="color: black;">附庸</span>于强人工智能,<span style="color: black;">无</span>必要花费太大的代价单独<span style="color: black;">研发</span>。在强人工智能<span style="color: black;">显现</span>前一味关注这一能力,恐怕有事倍功半<span style="color: black;">乃至</span>无解的<span style="color: black;">危害</span>。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">即便强人工智能<span style="color: black;">研发</span>成功,<span style="color: black;">亦</span>如人类的期望写出了真正的诗歌,届时或许会有一个新的问题<span style="color: black;">显现</span>:强人工智能何以<span style="color: black;">必要</span>用人类能够欣赏的方式表达情感?人类固然在有意无意地以人的标准塑造人工智能,但人工智能终究是一种与人类<span style="color: black;">区别</span>的智慧体,它们真实的表达方式<span style="color: black;">特别有</span>可能<span style="color: black;">没法</span>移情于人类,正如<span style="color: black;">咱们</span>很难体会动物的心情(值得<span style="color: black;">重视</span>的是,相<span style="color: black;">针对</span>以无机物为载体的人工智能,人类与动物皆属有机物,关系更近),而<span style="color: black;">倘若</span>在<span style="color: black;">这般</span>的<span style="color: black;">状况</span>下</span>AI<span style="color: black;">仍然采用符合人类预期的方式进行情感表达,<span style="color: black;">那样</span>在很大程度上这仍然是一种缺乏真正意图的模仿。换言之,创造符合人类审美的诗歌对</span>AI<span style="color: black;">未必有“必要性”。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;"><span style="color: black;">5、</span>人工智能创作诗歌的借鉴作用</span></strong></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">上面列举了人工智能创作诗歌所面临的四个挑战,但现<span style="color: black;">周期</span>人工智能的诗歌创作仍可<span style="color: black;">供给</span><span style="color: black;">必定</span>的借鉴<span style="color: black;">道理</span>。它有助于人们重新思考诗歌的好坏和门槛。在缺乏参照的<span style="color: black;">状况</span>下,人类对自己的长处与短处<span style="color: black;">常常</span>是<span style="color: black;">茫然</span>的。再举围棋为例。在阿尔法狗<span style="color: black;">显现</span>前,人们常以为人工智能下棋的强项在依赖计算的微观搏杀部分,而在诉诸直觉的大局观上应该不如人类,<span style="color: black;">因此呢</span>人类的机会集中在布局<span style="color: black;">周期</span>,<span style="color: black;">必须</span>精细计算的中盘和收官才是</span>AI<span style="color: black;">的擅场。阿尔法狗与李世石的实战颠覆了这一想象。被誉为“阿尔法狗之父”的</span>Demis Hassabis<span style="color: black;"><span style="color: black;">重视</span>到,在狗李之战第二局的第</span>37<span style="color: black;">步,阿尔法狗<span style="color: black;">无</span><span style="color: black;">根据</span>人类行棋的惯例在右起的第三线或第四线落子,而出人意料地在第五线落子。此线离中部位置较近,<span style="color: black;">寓意</span>着阿尔法狗在战略上更“<span style="color: black;">注重</span>”中部区域。<span style="color: black;">这般</span>的“胜负手”层出不穷,可见阿尔法狗大局观之高明。相反,在微观计算部分,却不乏恶手与俗手</span><span style="color: black;"></span><span style="color: black;">。同理,以</span>AI<span style="color: black;">诗歌为镜,<span style="color: black;">亦</span>能映照出人类诗歌的问题,<span style="color: black;">促进</span>人们分辨出诗歌中真正可贵的神理,改进创作的技巧,转移阅读的重点。<span style="color: black;">例如</span>,<span style="color: black;">经过</span>微软小冰创作的诗歌,人们会<span style="color: black;">发掘</span>部分汉语现代诗歌意象堆垛、情感泛滥、比喻莫名其妙,小冰<span style="color: black;">能够</span>轻易模仿、批量生产。相对而言,当前的人工智能模仿旧体诗词较为困难。</span>AI<span style="color: black;">“诗人”所作旧体诗词虽然符合格律,但稍有古典修养的读者一眼可辨其伪。这是<span style="color: black;">由于</span>旧体诗词不仅在形式上有较为严格的限制,<span style="color: black;">况且</span>经过<span style="color: black;">长期</span>的发展,形<span style="color: black;">成为了</span>“默认”(</span>default<span style="color: black;">)的字法、句法、章法</span>——<span style="color: black;"><span style="color: black;">虽然说</span>只要有法可学,人工智能模仿并非难事,但旧体诗词的法度多与<span style="color: black;">详细</span>情境结合,分寸殊难把握,当前的人工智能<span style="color: black;">没法</span>体察人情世故,几乎无从入手</span>——<span style="color: black;"><span style="color: black;">针对</span><span style="color: black;">熟练</span>旧体诗词文字气味的读者而言,很容易就能觉察正宗与否。新诗则<span style="color: black;">否则</span>,<span style="color: black;">更加多</span>地追求语言的“疏离化”(</span>defamiliarization<span style="color: black;">),与基于经验的<span style="color: black;">平常</span>语言截然异趣。这增大了人类理解的难度,却降低了人工智能模仿的门槛。从这个<span style="color: black;">道理</span>上说,当下人工智能创作的诗歌或许能在<span style="color: black;">必定</span>程度上改变诗歌的<span style="color: black;">评估</span>标准</span>——<span style="color: black;">如上例就<span style="color: black;">表现</span>了旧体诗词相<span style="color: black;">针对</span>新诗的某种可能的<span style="color: black;">优良</span>。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;"><span style="color: black;">6、</span>人文学科<span style="color: black;">必须</span>做些什么?</span></strong></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">人工智能的语言能力<span style="color: black;">研发</span>本质上是一个技术问题,<span style="color: black;">重点</span>牵涉到计算机学、神经学、语言学、<span style="color: black;">规律</span>学等非人文学科<span style="color: black;">行业</span>,但人工智能诗歌创作能力的<span style="color: black;">研发</span>则<span style="color: black;">必要</span>有人文学科参与其间。然而,当下该<span style="color: black;">行业</span>却较少见到人文学者的身影。这一方面是<span style="color: black;">由于</span><span style="color: black;">研发</span>者将审美的重要性置于技术之下,另一方面<span style="color: black;">亦</span>是<span style="color: black;">由于</span>当代人文学科体系以学理性的<span style="color: black;">科研</span>为主,把创作和鉴赏放在了次要的位置。事实上,就诗歌一道而言,审美是极为重要的。一件<span style="color: black;">制品</span>只要合格就<span style="color: black;">能够</span>用,一道菜肴即便无味<span style="color: black;">亦</span>能果腹,然而复制平庸的诗歌几乎<span style="color: black;">无</span><span style="color: black;">道理</span>。换言之,一首诗的价值<span style="color: black;">首要</span>在于它的好。<span style="color: black;">针对</span>诗歌<span style="color: black;">来讲</span>,重要的是审美性而非实用性。这就<span style="color: black;">必须</span>人文学者的介入。遗憾的是,人文缺席严重制约了</span>AI<span style="color: black;">诗歌创作的发展。以前文所举微软小冰为例,其诗集《阳光失了玻璃窗》遭到了诗歌界的<span style="color: black;">消极</span><span style="color: black;">评估</span>,如诗人于坚说:“(小冰的诗作)冷酷、无心,修辞的空转,东一句西一句随意组合,意象缺乏内在<span style="color: black;">规律</span>,软语浮词,令人生厌的油腔滑调,原材料来自平庸之句。这个软件对诗的理解是电视台诗歌朗诵会的水平。这个软件设计不出灵性,设计不了‘诗成泣鬼神’。”</span><span style="color: black;"></span><span style="color: black;">诗人欧阳江河则认为:“(小冰的作品)不是人话,<span style="color: black;">无</span>人味。”“我当然认为,<span style="color: black;">咱们</span><span style="color: black;">必要</span>要承认人工智能,小冰<span style="color: black;">亦</span>会继续发展下去。在十万个二流诗人里面,它<span style="color: black;">能够</span>排到</span>30<span style="color: black;">、</span>50<span style="color: black;">,<span style="color: black;">然则</span>这毫无<span style="color: black;">道理</span>。它永远成不了<span style="color: black;">咱们</span>所说的原创性<span style="color: black;">道理</span>上的伟大诗人。伟大的诗人<span style="color: black;">必定</span>针对的是文明<span style="color: black;">道理</span>上的东西。而这种东西绝不是学学已有的大诗人的句子,写出<span style="color: black;">有些</span><span style="color: black;">美丽</span>的、优美的句子就能达到的。”</span><span style="color: black;"></span><span style="color: black;">这些酷评固然有当前技术限制的<span style="color: black;">原由</span>,但<span style="color: black;">亦</span>有“原材料来自平庸之句”<span style="color: black;">这般</span>的设计者<span style="color: black;">导致</span>的缺憾,而“软语浮词,令人生厌的油腔滑调”<span style="color: black;">更加是</span><span style="color: black;">显著</span>的人为后果。欧阳江河所说“伟大的诗人<span style="color: black;">必定</span>针对的是文明<span style="color: black;">道理</span>上的东西。而这种东西绝不是学学已有的大诗人的句子,写出<span style="color: black;">有些</span><span style="color: black;">美丽</span>的、优美的句子就能达到的”,揭示了诗歌<span style="color: black;">必要</span>注入比文学技巧更广阔、更宏伟的人文内涵。即便回到基本的文学技巧层面,当前对人工智能诗歌<span style="color: black;">研发</span>中<span style="color: black;">表现</span>的人文水平<span style="color: black;">亦</span>是远远<span style="color: black;">不足</span>的。例如,不加择别地<span style="color: black;">供给</span><span style="color: black;">海量</span>诗歌<span style="color: black;">做为</span></span>AI<span style="color: black;">的学习和训练资料,而<span style="color: black;">忽略</span>了<span style="color: black;">各样</span>风格之间的潜在矛盾。<span style="color: black;">通常</span>而言,人类学习诗歌总是从某个风格入手</span>——<span style="color: black;">古人<span style="color: black;">所说</span>“专攻一家”</span>——<span style="color: black;">某些<span style="color: black;">研发</span>者让人工智能囫囵吞枣,自然难企高境。理想的<span style="color: black;">状况</span>是,具备创作和鉴赏经验的文学<span style="color: black;">科研</span>者与技术人员合作,<span style="color: black;">按照</span>人工智能的特性和认知科学的成果,精心制定循序渐进的学习<span style="color: black;">过程</span>。深度学习(</span>deep learning<span style="color: black;">)的发展使得人工智能有可能模拟人类大脑皮层的活动,<span style="color: black;">从而</span>以一种类似于人类的方式学习和思考,<span style="color: black;">那样</span>人类的教育方式或许仍可经过<span style="color: black;">必定</span>的改造应用于人工智能。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;"><span style="color: black;">7、</span>结语</span></strong></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">人工智能的发展是大势所趋,据统计,截止</span>2019<span style="color: black;">年</span>6<span style="color: black;">月,仅上海一地就有人工智能核心企业</span>1000<span style="color: black;">余家</span><span style="color: black;"></span><span style="color: black;">。<span style="color: black;">咱们</span><span style="color: black;">乃至</span><span style="color: black;">能够</span>说,人工智能不仅是一个行业,<span style="color: black;">况且</span>是所有行业的<span style="color: black;">将来</span>。本文讨论的仅是它无数可能具备的技能中的一种</span>——<span style="color: black;">诗歌创作。这一技能有<span style="color: black;">必定</span>特殊性,<span style="color: black;">由于</span>诗歌是人类最独特的精神产物之一。</span>2019<span style="color: black;">年</span>4<span style="color: black;">月,笔者应邀参加了贵阳孔学堂举办的关于人工智能能否进行诗歌创作的辩论,<span style="color: black;">做为</span>反方,对人工智能创作诗歌的可能性做出了质疑。然而,笔者更愿意相信,人工智能的可能性是无限的。这一信念<span style="color: black;">创立</span>在一个简单的假设上:人工智能能够在极短的时间内走完人类漫长的进化历程,如阿尔法狗从</span>2014<span style="color: black;">年<span style="color: black;">起始</span>研发,不到两年就打败了人类最优秀的棋手之一,而阿尔法元仅训练三天就完胜阿尔法狗。<span style="color: black;">能够</span>毫不夸张地设想,<span style="color: black;">倘若</span></span> AI<span style="color: black;">能够在一天内从无到有,赶上人类的水准,<span style="color: black;">那样</span><span style="color: black;">次日</span>它就有可能进化到人类望尘莫及的境界。<span style="color: black;">咱们</span><span style="color: black;">此刻</span>的认识受锢于人类的大脑,而一旦技术奇点(</span>technologicalsingularity<span style="color: black;">)来临,</span>AI<span style="color: black;">进化出了更卓越的“大脑”,<span style="color: black;">那样</span>当下的“不可能”或将变成“可能”。这曾经只是一个科幻色彩浓郁的臆想,而<span style="color: black;">此刻</span><span style="color: black;">成为了</span>一个日益临近的现实。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">(原载于《华南师范大学学报》2019年第5期)</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;"></span><span style="color: black;">《微软小冰首席<span style="color: black;">专家</span>宋睿华:小冰是<span style="color: black;">怎样</span>写诗的?》,九州连线网,</span>2018-10-31<span style="color: black;">,</span>http://www.sohu.com/a/272437756_634549<span style="color: black;">。</span> </span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;"></span><span style="color: black;">《人工智能微软小冰出首部“诗集”》,新华网,</span>2017-05-20<span style="color: black;">,</span>http://www.xinhuanet.com/local/2017-05/20/c_129609705.htm<span style="color: black;">。</span></span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;"></span>“Scientists trained an AI to write poetry, and now it’s standingtoe-to-toe with Shakespeare”<span style="color: black;">,</span>Boy Genius Report<span style="color: black;">,</span>2018-8-8<span style="color: black;">,</span>https://bgr.com/2018/08/08/poetry-ai-bot-shakespeare-human-research/?from=singlemessage&amp;isappinstalled=0<span style="color: black;">。</span></span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;"></span><span style="color: black;">【美】乔姆斯基:《句法结构》,第</span>8<span style="color: black;">页,黄长著等译,中国社会科学出版社</span>1984<span style="color: black;">年版。</span></span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;"></span><span style="color: black;">《苏轼文集》,第</span>2552<span style="color: black;">页,中华书局</span>1986<span style="color: black;">年版。</span></span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;"></span><span style="color: black;">《苏轼文集》,第</span>2124<span style="color: black;">页。</span></span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;"></span><span style="color: black;">杜牧:《樊川文集》,第</span>152<span style="color: black;">页,中华书局</span>1979<span style="color: black;">年版。</span></span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;"></span><span style="color: black;">【英】威廉</span>·<span style="color: black;">燕卜荪:《朦胧的七种类型》,第</span>57<span style="color: black;">页,周邦宪等译,中国美术学院出版社</span>1996<span style="color: black;">年版。</span></span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;"></span><span style="color: black;">《毛诗正义》,第6页,北京大学出版社</span>1999<span style="color: black;">年版。</span></span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;"></span><span style="color: black;">《孟子正义》,第</span>638<span style="color: black;">页,中华书局</span>1987<span style="color: black;">年版。</span></span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;"></span><span style="color: black;">《王国维全集》第</span>1<span style="color: black;">册,第</span>497<span style="color: black;">页,浙江教育出版社</span>2009<span style="color: black;">年版。</span></span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;"></span> “The mind in the machine: Demis Hassabis on artificial intelligence”,FinancialTimes<span style="color: black;">,</span>2017-4-21<span style="color: black;">,</span>https://www.ft.com/content/048f418c-2487-11e7-a34a-538b4cb30025<span style="color: black;">。</span></span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;"></span><span style="color: black;">于坚:《一种可怕的美<span style="color: black;">已然</span>诞生》,载《南方周末》,</span>2017-6-15<span style="color: black;">。</span></span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;"></span><span style="color: black;">《人工智能将<span style="color: black;">解决</span>诗歌?欧阳江河</span>:<span style="color: black;">不可能完成的任务》,中新网,</span>2017-7-3<span style="color: black;">,</span>http://www.xinhuanet.com//book/2017-07/03/c_129645856.htm<span style="color: black;">。</span></span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;"></span><span style="color: black;">《上海已<span style="color: black;">持有</span>人工智能核心企业</span>1000<span style="color: black;">余家</span> <span style="color: black;">初步建成中国人工智能发展领先地区》,凤凰财经,</span>2019-6-18<span style="color: black;">,</span>https://finance.ifeng.com/c/7nbd11k0OP6<span style="color: black;">。</span></span></p>




wrjc1hod 发表于 5 天前

楼主发的这篇帖子,我觉得非常有道理。

nqkk58 发表于 前天 14:54

楼主的文章非常有意义,提升了我的知识水平。
页: [1]
查看完整版本: 程羽黑:​人工智能诗歌论