ELIZA

A Java implementation of Joseph Weizenbaum's ELIZA chatbot (1966), running the DOCTOR script — a Rogerian psychotherapist simulation. Seasoned with a dash of Dr. Sbaitso personality.

Features

• Keyword-based pattern matching with decomposition/reassembly rules, following Weizenbaum's original design
• Priority system — when multiple keywords match, the highest-priority rule wins; if its patterns don't fit, the engine falls through to the next matching rule
• Pronoun reflection — captured text is reflected (I → you, my → your, je → vous, etc.)
• Round-robin responses — each pattern cycles through its reassembly templates
• Memory — certain rules store a response for later recall when no keyword matches
• Parity error — repeated insults trigger a Dr. Sbaitso-style fake crash (PARITY ERROR!!! SESSION TERMINATED.)
• YAML-driven rules — all rules, reflections, and UI messages are defined in external YAML files, easy to extend
• Multilingual — English (default) and French, selectable at startup via a CLI argument
• Accent-insensitive matching — French users can type with or without accents (deprime matches déprimé); ligatures like œ are also handled (sœur matches soeur)

Requirements

• Java 17+
• Maven 3.8+

Build

mvn clean package

This produces an executable fat JAR at target/eliza.jar.

Run

# English (default)
java -jar target/eliza.jar

# French
java -jar target/eliza.jar fr

Or on Windows:

eliza.cmd
eliza.cmd fr

English session

Type anything to talk to ELIZA. Type quit, bye, or exit to end the session.

ELIZA: Hello. I'm ELIZA. How are you feeling today?

You:   I am feeling sad
ELIZA: I'm sorry to hear that you are sad.

You:   Since my mother passed away
ELIZA: Tell me more about your mother.

You:   I dream about her sometimes
ELIZA: What does that dream suggest to you?

You:   I don't know
ELIZA: Why don't you know?

You:   bye
ELIZA: Goodbye. Thank you for talking with me.

French session

Tapez ce que vous voulez. Tapez quitter, quit ou au revoir pour sortir.

ELIZA: Bonjour. Je suis ELIZA. Comment vous sentez-vous aujourd'hui ?

Vous:  je suis triste
ELIZA: Je suis désolé d'apprendre que vous êtes triste.

Vous:  ma mere est decedee
ELIZA: Parlez-moi davantage de votre mère.

Vous:  je reve d'elle parfois
ELIZA: Que vous suggère ce rêve ?

Vous:  quitter
ELIZA: Au revoir. Merci d'avoir parlé avec moi.

Project structure

src/
  main/
    java/com/eliza/
      Eliza.java              # Engine: matching, reflection, memory, insult tracking
      Rule.java               # Rule record with PatternResponse (round-robin)
      Main.java               # Console interface, loads messages from YAML
    resources/
      rules_us.yaml           # ~170 English keyword rules
      rules_fr.yaml           # ~180 French keyword rules
      reflections_us.yaml     # English pronoun/verb reflection mappings
      reflections_fr.yaml     # French pronoun/verb reflection mappings
      messages_us.yaml        # English UI strings (intro, greetings, prompts, etc.)
      messages_fr.yaml        # French UI strings
  test/
    java/com/eliza/
      ElizaTest.java          # 32 JUnit 5 tests

How it works

1. Preprocess — trim, strip trailing punctuation, normalize whitespace, lowercase, strip accents and expand ligatures (œ → oe)
2. Match — find all rules whose keyword appears in the input (accent-insensitive), sorted by descending priority
3. Decompose — for each matching rule, try its regex patterns (accent-insensitive) against the input
4. Reassemble — fill the response template with captured groups, reflecting pronouns; responses keep proper accents
5. Fallback chain — if no pattern fits, try the next rule; then check memory; then use the @none fallback

Rule priority tiers

Priority	Category
7	Safety (die/mourir)
6	Computer/ordinateur
5	Family, "I remember", "I want", "I feel", insults
4	Emotions, topics (dream, sad, angry, ...)
3	Greetings, misc patterns, work, help
2	"I am"/"je suis", "my", absolutes (always, never)
1	Yes/no, uncertainty, sorry
0	@none fallback

Note: "i am" / "je suis" are intentionally low priority (2) so that specific keywords like "sad" (4) or "depressed" (4) win when the input is "I am sad".

Memory

Some rules include a @memory: directive as one of their round-robin reassemblies. When triggered, the response is stored in a queue instead of being returned. On a later turn where no keyword matches, the stored memory is recalled — giving ELIZA the appearance of remembering earlier topics.

Parity error

Rules tagged with insult: true increment a counter when matched. After 4 insults, ELIZA returns a PARITY ERROR!!! message and the console displays a fake system crash — a nod to Dr. Sbaitso's behavior from 1991.

Tests

mvn test

32 tests covering keyword matching, priority ordering, pronoun reflection, preprocessing, round-robin cycling, fallback, memory storage/recall, parity error, French language loading, accent-insensitive matching, and ligature handling.

Adding a language

1. Create rules_{lang}.yaml — keyword rules with decomposition/reassembly patterns
2. Create reflections_{lang}.yaml — pronoun/verb reflection mappings
3. Create messages_{lang}.yaml — UI strings (intro, greetings, prompt, goodbye, quit_words, crash)
4. Add the language code to the CLI check in Main.java

Extending the rules

Edit src/main/resources/rules_{lang}.yaml to add new rules:

- keyword: "weather"
  priority: 3
  patterns:
    - decomposition: ".*weather(.*)"
      reassemblies:
        - "Does the weather affect your mood?"
        - "Tell me how{1} weather makes you feel."

• keyword — substring matched against the preprocessed input (accent-insensitive)
• priority — higher wins; ties go to the first rule defined
• insult: true — optional flag to count toward parity error
• decomposition — Java regex, matched accent-insensitively; capture groups are referenced as {1}, {2}, ...
• reassemblies — cycled round-robin; prefix with @memory: to store instead of respond; keep proper accents here