问Wentian Notes
NLP Note 4 of 12

Applications of NLP

912 wordsAbout 4 min

Applications of Natural Language Processing

Word vectors are used to describe the relationships between words, or in other words, what a piece of text means

Assisted Medical Consultation

Analyzing pathology texts, text structuring, text annotation. An excellent doctor is the accumulation of experience and professional knowledge, and AI can learn both kinds of knowledge

Education

Adaptive Learning: teaching students according to their aptitude

Reading platform Newsela

Adaptive platform Newton, in K12 and higher education

English Liulishuo

Media

In content production, AI can take part in basic work such as screening, proofreading and editing. Representative examples include the foreign writing robot Heliograf, Toutiao’s writing robot xiaomingbot, Yicai’s DT稿王 (DT Writing King), Tencent Finance’s Dreamwriter, and Xinhua News Agency’s 快笔小新 (Kuaibi Xiaoxin). Their advantages are high efficiency, timeliness and fast updates.

Besides intelligent writing, AI can also be used for intelligent recommendation, such as The New York Times’ robot editor Blossomblot. According to statistics, the articles recommended by this robot received 40 times more reads than articles selected by people

Finance

Applications in finance are very broad, because the financial sector has accumulated a large amount of data and its degree of digitalization is already very high; with suitable big-data processing techniques, it can be turned into a large amount of structured data

  1. Smart banking: AI + business —> intelligent customer service, call data between the bank and customers
  2. Intelligent wealth advisors: recommend suitable products by analyzing customers’ preferences
  3. Intelligent investment research: process and analyze data to generate investment decisions, professional knowledge, research reports, knowledge graphs

Law

There is also a large amount of text data: indictments, judgments, case records. China has relatively few lawyers and fees are high

AI can recommend similar cases to lawyers, and can also enable automated adjudication

For courts, it can handle some document-processing workflows

Common Tasks in Natural Language Processing

Mostly Solved: Tasks That Are Basically Solved

  1. Spam Detection
  2. Word Segmentation
  3. Part Of Speech (POS) Tagging
  4. Named Entity Recognition (NER)

Making good Progress: Tasks That Have Made Great Progress

  1. sentiment Analysis
  2. Coreference Resolution, e.g. Manning is a great NLP professor,he worked in academia for over 24 years. Here, “he” is the coreference
  3. Word Sense Disambiguation (WSD), e.g. I went to the bank to get some money,here bank means a financial institution, not the land beside a river.
  4. Machine translation

still a bit hard: Tasks That Are Still Somewhat Difficult

  1. Dialogue agents and chat-bots. There are several categories here: 1. chit-chat; 2. task-oriented, where a task must be completed, such as booking a flight—but with multi-party conversations, understanding context becomes involved, which is harder and requires memory; 3. knowledge Q&A, which can look up some basic common knowledge
  2. Summarization. There are generally two approaches: 1. extractive, which extracts some of the sentences and is relatively simple; 2. abstractive (generative), which selects some key words from the text, and is comparatively more difficult
  3. NLP for low resource languages. For some specific tasks the corpus is relatively small, which makes the task harder too. Since we do many tasks with deep learning, insufficient corpus data has a very large impact on the specific task

Challenges of Natural Language Processing

  1. Word sense disambiguation: the same word may take on extended meanings in different contexts. For example, when Curry lunges forward, the audience might say “大家可别让他伤了” (“don’t let him get hurt”); when O’Neal lunges forward, the audience might also say “大家可别让他伤了” (here meaning “don’t get hurt by him”). Other examples: 口味清淡 (light taste) vs. 生意清淡 (slow business); Li Mi, leader of the Wagang army, vs. Li Mi, author of the “Chen Qing Biao”; “You’re a good person?” vs. “You’re a good person!”
    To give a practical scenario: “Xiao Ming goes to buy an Apple phone” and “Xiao Ming likes eating apples” are two known sentences; have the machine understand the following sentence: “One drawback of the Apple is that its screen breaks easily, so it needs a phone case”
  2. Coreference resolution: coreference is a relatively advanced linguistic-cognitive activity. In children’s books, coreference is used relatively rarely and repetition is common, while in adult reading material coreference appears much more often. This is also a very hard task for machines
  3. Context understanding: here we use three sentences as examples: 1. “I have an Apple, and also a Samsung” 2. “Are you hungry? I have an apple” 3. “I have an apple”
  4. Pragmatics: this involves the difference between semantics and pragmatics. “It’s really hot today” and “Please turn on the air conditioner” have different semantics but the same pragmatics—both want the robot to turn on the air conditioner. This may be a relatively difficult point even for humans, and it is a challenge for robots as well

The problem with today’s chat-bots is that they have no memory and a monotonous way of speaking, which is what makes robots seem like robots

Tools for Natural Language Processing: Packages for Basic Tasks

These are tools that an NLP engineer needs to master thoroughly

  1. NLTK (Natural language toolkit): a package for foreign (non-Chinese) languages that can do word segmentation, POS tagging and NER
  2. HaNLP, FNLP, jieba: these are all good tools for Chinese

Basic Frameworks for Scientific Computing and Machine Learning

numpy pandas scipy scikit-learn

Deep Learning Frameworks

Pytorch (the most important), Tensorflow, Keras (this one is simple and easy to pick up)