Swahili Chatbot with RegEx · Projects

A rule-based chatbot for Swahili, built in the style of ELIZA. It holds a conversation with no learned model at all: the single eliza function runs the input through a fixed ladder of eight regular expressions, each with capture groups, and the first one to match builds the reply. An input that matches nothing falls through to a neutral prompt.

The patterns are tried top to bottom in an if/elif chain, so their written order is their priority. Each one targets a family of Swahili sentences:

Naming. Naitwa X or Jina langu ni X (my name is X) returns the respectful greeting Shikamoo, X. Umeshindaje?
State of mind. nimefurahi / sijahuzunika and kin, the nime/sija prefix plus a mood stem, get reflected back as a question.
Self-description. Mimi ni … (I am …) becomes Mbona wewe …? (why are you …?).
Family. A mama/baba/kaka/dada noun with the -ngu (my) suffix draws out Niambie mengine kumhusu …ko (tell me more about your …).
Intent. Modal stems Nataka/Sitaki/Naweza/Siwezi/Nahitaji and the obligation Lazima ni… are echoed with the person flipped.

The remaining rules catch thoughts (Nadhani/Natumai), ask for an example on any generic verb, and deflect a short list of insults; the default reply Niambie mengine… (tell me more) closes the ladder.

Algorithm:

\textsc{Respond}(u, R)

— first-match pattern reply

1
input: a user utterance $u$ , an ordered list $R$ of (pattern $p$ , templates $T$ ) rules
2
$x \gets$ normalize and lowercase $u$
3
for each $(p, T)$ in $R$ , in order, do
4
$m \gets$ match $p$ against $x$
5
if $m$ succeeds then
6
$t \gets$ choose a template from $T$
7
return $t$ with capture groups of $m$ substituted in
8
return the default fallback reply

Reflection keeps the replies coherent. When a captured group is spliced back into a reply, its person markers are rewritten so the bot answers from its own point of view. The possessive -angu (my) is swapped to -ako (your), and the subject prefixes on verbs flip — a first-person nime/nataka returns as second-person ume/unataka. Without this step, echoing the user's words reads as nonsense; with it, a statement about "me" comes back as a question about "you", so the reply stays on topic without any parsing of meaning.

You can read the full report.

References

Project repository
Reference notes: Dialogue and Chatbots

╌╌ END ╌╌