Class 1: Overview, Introductions, and a Text Investigation
Tuesday, October 6, 2026 · 14:40–16:10 · DH Lab B307
1. NLP tasks and methods
Natural language processing (NLP) includes many tasks. Even with the same text, deciding word boundaries, grammatical relationships, word meanings, and who an expression refers to requires different judgments.
| Task | Example output | Specialized approaches |
|---|---|---|
| Morphological analysis and POS tagging | Word boundaries and parts of speech | Dictionaries and statistical analysers |
| Dependency parsing | Links between words in “Tanaka reads a book” | Grammars and trained parsers |
| Semantic similarity | A similarity score for “read a book” and “do some reading” | Word co-occurrence and embedding models |
| Word-sense disambiguation | Whether “bank” means a financial institution or a river bank | Dictionaries and classifiers using surrounding words |
| Named entity recognition | Names and their categories | Rules and trained NER models |
| Coreference resolution | Linking “Natsume Sōseki” and “he” | Rules and models for linking mentions |
| Text classification | An article labelled politics or literature | Word frequencies and trained classifiers |
| Machine translation | An English translation of a Japanese paragraph | Translation models trained on paired texts |
| Question answering | An answer to “Who climbed?” | Search and models that extract an answer passage |
| Summarization | A short account of the main events | Scoring and selecting source sentences |
| Speech recognition | A transcript of a recorded introduction | Acoustic models and language models |
Specialized systems often combine several models, dictionaries, and rules.
Changes in NLP and software
| Resource | Excerpt and question |
|---|---|
| Andrej Karpathy: Software Is Changing (Again) (Y Combinator, 2025, English) | Watch 1:25–6:10, lasting 4 minutes 45 seconds. Compare writing code, training a model, and instructing an LLM in words. |
| Christopher Manning: History of Natural Language Processing (Stanford CS224N, 2026, English). Course website · YouTube lectures (2024) | Slide 2, “Four eras of NLP.” Locate rule-based methods, statistical machine learning, and deep learning in the history of NLP. |
- NLP has long used models and systems tailored to individual tasks.
- Shared pretrained models and LLMs can handle multiple tasks.
- Each task still needs clear output criteria and its own evaluation.
Optional: 3Blue1Brown on LLMs
- English original: Large Language Models Explained Briefly: 7 minutes 58 seconds
- Official Japanese version: 「LLMの仕組み（簡単バージョン）」: 7 minutes 44 seconds
2. Named entity recognition: criteria and methods
Decide where a person’s name begins and ends
田中さんは東京で夏目漱石の『こころ』を読んだ。
Tanaka read Natsume Sōseki’s Kokoro in Tokyo.
- Does the honorific
さんbelong in the extracted name田中さん? - Does mentioning Natsume Sōseki mean he is present in the scene?
Comparing rules and models requires shared criteria for name boundaries and a clear definition of what to extract.
Compare NER methods
Compare these methods for named entity recognition (NER).
| Method | Questions to investigate |
|---|---|
| String, dictionary, and rule-based search | Can we read the rule and explain the result? Does it miss another spelling or context? |
| Fixed-category NER models (GiNZA, BERT, and RoBERTa models) | Does the training language and material fit our source? Does the trained category set include what we want to study? |
| NER models with supplied categories (GLiNER) | Can we specify what we want to extract? What changes when we change the category descriptions? |
| Generative large language models (LLMs) | What changes when we change the instruction? Can we verify its quotations and explanations in the source? |
Compare the same four paragraphs of 「蜘蛛の糸」 using people/places, then deities/sinners/places of punishment. Predict whether 極楽 belongs in each place category. Which methods can accommodate a change of categories?
3. Investigate the text and record evidence
Reading and source
- Akutagawa Ryūnosuke, 「蜘蛛の糸」: read the text at Soranoha
- Soranoha’s plain text: text to search in the browser
- Soranoha’s work page: work and release information
- Aozora Bunko’s work card: edition, base text, contributors, and downloadable text
The card labels this version 新字新仮名 (modern character forms and kana spelling). Soranoha’s reading page displays pronunciation guides; its plain text leaves out the guides and editorial notes.
Search matches and events
Suppose a search finds 下りろ (“Descend!”). Did an actual descent happen, not happen, or can we not tell from this information alone?
Open the plain text. Use Command+F on a Mac, or Ctrl+F on Windows/Linux, to search for 下りろ. Read the surrounding passage and identify who asks whom to do what. Then revisit your initial prediction.
This expression is a command to descend. Finding it does not establish that the people addressed actually descended. Record an expression about downward movement separately from evidence of an actual descent.
Your investigation: upward and downward movement
Our shared question is how the story describes movement upward and downward.
- Before searching, predict one expression for upward movement and one for downward movement.
- Read the story and try your expressions in the plain text. If you need a starting point, use
のぼfor upward movement and落ちfor downward movement. - Choose one upward and one downward passage, then read their surrounding paragraphs. Record the part (一, 二, or 三), actor, short expression, and direction.
- Explain whether each passage describes actual movement, a condition, a negation, a command, or another context. If you cannot decide, record why.
- Compare one judgment in pairs, then revisit the passage or search expression.
Browser search finds the string you entered. It does not decide whether inflected forms or compound words belong to the same verb. If you get no matches, check the spelling in the text and change one search expression.
If Soranoha does not load, use this copy of the text.
Classify the same passage independently
If you agree, did you use the same evidence? If you disagree, which expression or surrounding context explains the difference? If you cannot decide, compare your reasons for leaving the judgment open.
Notes to bring to the next class
Keep today’s notes and bring them to the next class. There is no submission today.
| What to save | Contents |
|---|---|
| Question and prediction | A small question the text could answer, and your prediction before searching |
| Source and search | Work URL and search expressions |
| Two supporting passages | Part, actor, short expression, direction, and judgment based on context |
| Revision and uncertainty | What you changed, why, and what remains unresolved |
Finish with one sentence each: what you investigated, what the text supports, and what you would change next. For example, you could ask whether a search can distinguish actual climbing from imagined climbing in part 二.