📄 Improving the CSIEC Project and Adapting It to the English Teaching and Learning in China

Summary

This paper is a progress report on CSIEC (Computer Simulator in Educational Communication), a web-based chatbot that the authors started in 2003 as a conversation partner for learners of English. The system parses user input with an English parser into NLML markup, represents it as Java objects (NLOMJ), keeps the dialogue history in NLDB, and generates replies through a Communicational Response (CR) module. CR takes into account context, a world model and the personalities of both the user and the system. The report describes two additions. The first is five Microsoft Agent animated characters, each with a different conversational style. The second is a guided scenario mode, demonstrated with a mock graduate job interview. The authors then describe the Chinese setting: English is a compulsory, exam-driven subject, grammar instruction matters because Chinese grammar is so different, and native-speaking teachers are scarce. They argue that a natural-language partner can partly fill this gap. No quantitative evaluation is reported. The planned next steps are classroom trials with teachers and adding speech recognition.

Key points

• Architecture:
  • An English parser produces NLML markup.
  • NLOMJ holds the parsed grammar objects.
  • NLDB stores the dialogue history.
  • The CR mechanism builds replies.
  • Microsoft Agent provides synthesized speech and animation.
  • The public website had received over 40,000 homepage visits.
• Five characters, each tied to an agent type:
  • Christine is active: she tells stories, jokes and news.
  • Stephan is reactive: he mostly just listens.
  • Emina is proactive: she asks questions about what the user said.
  • Christoph is also proactive: he offers comments and advice.
  • Ingrid is a hybrid that draws on both the current input and the discourse history.
• Where the characters differ: all five handle questions and commands the same way. They differ only in how they respond to declarative statements. Users can also pick a character to voice their own lines.
• Job-interview scenario:
  • Interviewer questions are stored in an NLDB table, grouped by topic.
  • The topics run in order: college and specialty, student life, motivation for applying.
  • Within each topic, some questions are drawn at random and others must be asked in a fixed sequence.
  • At the end, the system lists the user's sentences that contain spelling or grammar problems but does not correct them, leaving correction to teachers.
• Context: a 2004 Ministry of Education report counted roughly 20 million Chinese English learners who must pass an English exam.

Interesting ideas

• I1. Routing by speech act first and only then applying a "personality" to plain statements gives distinct behaviours while keeping one shared core for questions and commands.
• I2. Emina forms follow-up questions by grammatically transforming the user's statement, for example by shifting its tense or turning it into a question with an auxiliary.
• I3. The bot remembers earlier user claims and points out when a new statement contradicts one ("you said you can't sing").
• I4. Scripted scenarios group questions by topic in a fixed topic order, with each question marked as either random or mandatory-sequence.
• I5. The system flags faulty sentences without auto-correcting them, on the deliberate reasoning that a wrong fix is worse than none.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (frames): classify each clause's speech act in prag, then let different reply policies attach only to assertions, so behaviour remains traceable to the act type.
• I2 (back): generate probing questions directly from a parsed statement's UD tree and frame, which also gives a cheap self-check that generation preserves the statement's predicate and arguments.
• I3 (world): store assertions per speaker as world states and flag conflicts with later assertions, with the earlier source sentence shown as the explanation.
• I4 (answer): school-course tests and tale-QA drills could use the same two-tier schedule, giving controlled variety without losing the necessary order of questions.
• I5 (learn): when self-induced repair rules have low confidence, emit a flagged sentence for review instead of a silent rewrite, and use these flags as a negative-control set.