AI 的語言世界：台大語言學研究所 謝舒凱

Outline

1. AI, Deep Learning, and the (unknown) Future
2. Cognition, Computation, Linguistics and AI : a (mis)match in Heaven?
3. Semantic Network: Representation, Memory and Processing
4. AI, AAI and AW (Artificial Wisdom)?

時代背景：我的 ** 不是人

廚師、司機、醫師、健身教練、理專、律師、老師、情人(?)、。。。

常被大眾談論（與誤解）的例子

• 語音辨識 = 語言理解
• (語音辨識 + 字串關聯比對的) 聊天機器人 = AI
• 人形機器人(android/humanoid) = AI

ASIMO, Boston Dynamics, up and coming

恐怖谷理論 (uncanny valley)，準嗎？

人類對機器人的好感度會隨著相似度增加，相似度高達 85% 時，會讓人心生恐懼與感到詭異。(森政弘,1970)

還是，可愛與需要打敗一切恐懼

(source: http://www.xenosystems.net/uncanny-valley/)

語音辨識

Google Web Speech API Demonstration

繼 語音 之後，語意（理解）是關鍵。

其他的 AI 相關領域

• 知識搜尋處理 決策、推薦、預測、推理、、
• 感知 (辨識) 視覺（人臉、圖形、物件）、語音與語言、情緒、、
• 生成 圖文詩詞自動生成、語音與音樂合成、娛樂商務與療癒對話、、 https://www.slideshare.net/ckmarkohchang/neural-art-english-version

其他參考

• 藏頭詩
• 歌詞/產生器
• 畫作
• Predicting Future Human Behavior with Deep Learning (Vondrick, 2016)

Deep Learning and Language Technology

Real Time American Sign Language Video Captioning using Deep Neural Networks

機器可以自學？可以!

機器可以自學？可以嗎?

• Elon Musk (Tesla) vs Mark Zuckerberg (FB)
• Singularity: AI 的高鐵過站隱喻。

當萬物皆可量度，我們都也只是一堆感測器 (Hsieh, 2017)

AI HERTORY: It all begins with ELIZA...

ELIZA 加強版誰都可以實作

• 聊天機器人很多 https://www.chatbots.org/
• 真正處理「語言理解」的系統不多 (對於急需商業應用的人來說不是重點)
  • Language-based AI

這不是我們所理解的「理解」吧？

YOU: My _1_ is _2_
ELIZA: How long has your _1_ been _2_ ?

YOU: _1_ 覺得我很 _2_
CELIZA: 那妳覺得妳很 _2_ 嗎？

資料夠大，即便神經網路模式也不一定穩定

AI, Neural Network (aka Deep Learning)

deep learning ＝ deep representation, 其梯度可以透過 chain rule 來反向傳播。

AI, Neural Network (aka Deep Learning)

按照能夠使得損失函數 L(W) 減小的方向 (e.g. 隨機梯度下降) 來調整參數訓練神經網絡。

Neural Network (aka Deep Learning) Typology

(End-to-End) AI 實作快變成中學生的課外活動

Tensorflow playground

Keras: Deep Learning library in R or Python within 30 seconds

library(keras)
mnist <- dataset_mnist()
model <- keras_model_sequential() 
model %>% 
  layer_dense(units = 256, activation = 'relu', 
              input_shape = c(784)) %>% 
  layer_dropout(rate = 0.4) %>% 
  layer_dense(units = 128, activation = 'relu') %>%
  layer_dropout(rate = 0.3) %>%
  layer_dense(units = 10, activation = 'softmax')

但，我們應該假定這才是理解能力嗎

• natural language processing to enable it to communicate successfully;
• knowledge representation to store what it knows or hears;
• automated reasoning to use the stored information to answer questions and to draw new conclusions;
• machine learning to adapt to new circumstances and to detect and extrapolate patterns.

題外話

測試一下班上有沒有機器人:)

蜘蛛為何是白色的？ 水手的工作？

Deep Learning: a hype or go-to algorithm?

• main criticism
• concerns

Issues

• 不是每個領域都有可得的大數據 (big data)。
• 不是每個問題都有標準答案 (labeled data)。
• 理解活動可能不是單純的模型或是序列匹配 (pattern matching/seq2seq chatbot)
• AI 中的科學與工程 Science v.s./with Engineering?

Last miles of the way ?

• Language understanding (vs. comprehension) and Linguistic complexity

"human language is one of the most complex processes to be found anywhere on our planet" (Tomasello, 2008).

Natural language understanding is sometimes referred to as an AI-complete or AI-hard problem, implying that the difficulty of these computational problems is equivalent to solving the central artificial intelligence problem.

[哲學思考] 丟一個問題給你，請想久一點

如果有機器通過了圖林測試，它還是機器嗎？

Loebner Prize Competition in Artificial Intelligence

Talking to itself/themselves (娛樂效果之外的反思.....)

Outline

1. AI, Deep Learning, and the (unknown) Future
2. Cognition, Computation, Linguistics and AI : a (mis)match in Heaven?
3. Semantic Network: Representation, Memory and Processing
4. AI, AAI and AW (Artificial Wisdom)?

人的歷程

認識、知識、常識、意識、心識

人的語言

What is Linguistics?

• 目的：語言學要回答語言的習得與發展,結構與功能,神經與心理機制,社會變異與演化過程等。
• 應用： (大數據中的) 語言數據(語料)蘊含了文化歷史記憶,社會心理趨勢,政治輿情傾向,情緒偏好分佈,人格特質與決策行為,疾病前期徵兆等等。

順便澄清一個觀念：

Asking a linguist how many languages they speak is like asking a doctor how many diseases they have (Lynne Murphy).

What is NLP?

經驗/計算語言學 (empirical/computational linguistics) [a.k.a. Natural Language Processing] 用電腦來幫助我們回答上述問題，並產生應用。

• Natural Language Processing (NLP)} is a field of computer science and linguistics concerned with the interactions between computers and human (natural) languages. It began as a branch of artificial intelligence, as a very attractive methodology of human–computer interaction.

語言複雜度的處理

• 表徵與標記 (representation and annotation)
• 計算與學習 (modeling and machine learning)
• 模擬 (simulation)

以計算詞彙語意學 (computational lexical semantics) 為例, 參考 2014 講義 ai-lecture2014.pdf

語言表徵與理解

Semantic memory and categorization

• Semantic memory: our organized knowledge about the world.
• Concepts and Categories:
  • The feature approach
  • The prototype approach
  • The exemplar approach (your previous experience lumped together into a category)
  • The network models

定義遊戲練習

Semantic memory

• general (common-sense) knowledge
• linguistic knowledge
• conceptual knowledge
  • Psychologists use the term concept to refer to your mental representations of a category (Murphy, 2010; Rips et al., 2012; Wisniewski, 2002).

Cognitive scientists have found it useful to draw a distinction between declarative (factual) knowledge and procedural knowledge.

「知識救援」的個人看法

• Human annotation/involvement driven
• Language and Knowwledge Resource embedded

Machine Learning w/o Knowledge Representation and Annotation

• 爭辯熱度最高的主題
  • 標記（特徵工程）很貴、費工、太主觀
  • 標記（就算沒用）反映了科學的努力與限制；至少我們知道他們在幹什麼

More on Levels of Annotation

Units (crossing the sentence boundary) reflect the communicative function of the sentence

• Topic-focus articulation/rhetorical structures and discourse connectives/anaphora and coreference

Topic-Focus Articulation (TFA)

• Topic: What is the sentence about?
  • The topic (or theme) is the part of the proposition that is being talked about (predicated). Once stated, the topic is therefore old news, i.e. the things already mentioned and understood.
• Focus: What information about the topic is asserted?
  • The focus determines which part of the sentence contributes the most important information. The focus may be highlighted either prosodically or syntactically or both, depending on the language.
• An important role is played by the position of the intonation marker.

Deep (Linguistic) Learning: batteries included?

But, which Semantics?

• Lexical Semantics
• Neural Semantics
• Vector Semantics

Network Representation of Semantics

Network Analysis

micro motifs and macro behavior

• focuses on (in)dependent actors at the micro level and the consequences of their behavior at the macro level.
• ideographic approach stresses the uniqueness of a phenomenon; nomothetic approach stresses law-likeness.

Small worlds:

'six degrees of separation' (Watts and Strogatz, 1998)

sw <- sample_smallworld(dim=2, size=10, nei=1, p=0.1)
plot(sw, vertex.size=6, vertex.label=NA, layout=layout_in_circle)

Lexical network (Wordnet)

Chinese Wordnet

• A network representation of lexical semantic knowledge in Chinese.
• Toward an OntoLex resource for Natural Language Processing.

Vector Semantics

• Like the neural language models, the word2vec models learn embeddings by training a network to predict neighboring words.

Vector Semantics

Word Embeddings

Neural Semantic Network

Gallant lab Brain Viewer

Human Connectome Project (HCP)

a consortium whose goal is to map “human brain circuitry in a target number of 1200 healthy adults using cutting-edge methods of noninvasive neuroimaging”

Neural WordNet

• Neurologically-based Ontology and Semantic Space (Huth et al. 2012; Huth et al. 2016, Nature ) Brain altas: same with Chinese?

https://www.youtube.com/watch?v=0FDtsbLZBuM

(Words/Chunks?) on the Brain: A Semantic Map of the Cortex

DeepMind 的 leader 要跟神經科學學習 / Hinton 認為 back-propagation 要打掉重練

Chinese QIEs

• Is Construction Grammar Neuroscientifically Plausible?

Chinese QIEs

• Idiom, Near, Far conditions were contrasted against random condition.
• BOLD signal increase:
  • Idiom condition: left and right AG (angular gyrus), PCC (posterior cingulate cortex)
  • Near conditions at left AG and IFG (left Inferior frontal gyrus )
  • No significant difference found in Far vs. Random condition.

Chinese QIEs

結論，有嗎之一

AI needs greater representation from the humanities

• 人的問題涉及行為、經驗、心靈、意識、生理變化等。
• 語言裡反映了慾望、記憶與遺忘、錯誤、情緒、愛、性格。

再確認妳是人嗎? 下面這句話一秒鐘看完並說出意思 賣女孩的小火柴 請上獎領台

結論，有嗎之二

call for 同理計算 (computational empathetic communication)

人格面具,自我實現、善心、惻隱之心、同情心、助人、愛心、民主、創造、幽默、風趣、詼諧、恐懼、自卑、氣質、非理性決策（直覺

結論，有嗎之三

From AI (人工智能) to AW (人工智慧)

• 馬失業之後應該比較開心過馬的生活。
• 人是什麼？機器是什麼？是變動的定義。

Kurzweil預計到 2030 年人類都會進化成半機械人（Elon Musk的 人機合一正是這項實驗， Neurolink），通過神經系統進入虛擬現實（VR）世界。事實上，AI早已改變人類行為。傳播學者 McLuhan 提出「人的延伸」，智慧手機就是「大腦延伸」。

結論，有嗎之四

Reference

[1] Huth AG, de Heer WA, Griffiths TL, Theunissen FE, & Gallant JL (2016). Natural speech reveals the semantic maps that tile human cerebral cortex. Nature, 532 (7600), 453-8 PMID: 27121839

[2] Friedemann Pulvermüller, Bert Cappelle and Yury Shtyrov. (2013). Brain basis of meaning, words, constructions, and grammar. In: Graeme Trousdale and Thomas Hoffmann (eds.), Oxford Handbook of Construction Grammar. Oxford: Oxford University Press, 397-416.