ovos-yes-no-plugin

A heuristic yes/no answer classifier for OpenVoiceOS.

Given a question and a spoken or typed response, it returns True (yes), False (no), or None (unclear). It handles common variants (yeah, sure, nah), double negatives (not a lie → yes), and language-specific word lists for 20+ locales.

Role in OVOS

OVOS skills ask yes/no questions through OVOSSkill.ask_yesno() in ovos-workshop:

# ovos_workshop/skills/ovos.py
def ask_yesno(self, prompt: str, data=None) -> Optional[str]:
    resp = self.get_response(dialog=prompt, data=data)
    engine = self._get_yesno_engine()
    answer = engine.yes_or_no(question=prompt, response=resp, lang=self.lang)
    if answer is True:
        return "yes"
    elif answer is False:
        return "no"
    else:
        return resp   # pass-through when answer is ambiguous

_get_yesno_engine() loads the plugin named ovos-yes-no-plugin by default. This package provides that plugin. The active engine can be overridden per-skill in settings.json or system-wide in mycroft.conf:

# mycroft.conf
{
  "skills": {
    "ask_yesno_plugin": "ovos-yes-no-plugin"
  }
}

# skill settings.json  (overrides the system default for this skill only)
{
  "ask_yesno_plugin": "my-custom-yesno-plugin"
}

Install

pip install ovos-yes-no-plugin

Usage

Inside an OVOS skill

class MySkill(OVOSSkill):
    def handle_intent(self, message):
        answer = self.ask_yesno("confirm.action")  # plays TTS, waits for reply
        if answer == "yes":
            self.speak_dialog("confirmed")
        elif answer == "no":
            self.speak_dialog("cancelled")
        else:
            self.speak_dialog("didnt.understand")

ask_yesno returns "yes", "no", or the raw utterance when the engine returns None.

Direct use

from ovos_yes_no import HeuristicYesNoEngine

engine = HeuristicYesNoEngine()
engine.yes_or_no("Do you want to continue?", "yes", "en-us")          # True
engine.yes_or_no("Do you want to continue?", "no way", "en-us")       # False
engine.yes_or_no("Do you want to continue?", "beans", "en-us")        # None
engine.yes_or_no("Do you want to continue?", "it's not a lie", "en-us")  # True  (double negative)
engine.yes_or_no("Do you want to continue?", "yes, but actually, no", "en-us")  # False (last word wins)

Via ovos-plugin-manager

from ovos_plugin_manager.agents import load_yesno_plugin

cls = load_yesno_plugin("ovos-yes-no-plugin")
engine = cls()
result = engine.yes_or_no("Shall I set a reminder?", "please do", "en-us")  # True

Configuration

No mandatory configuration. An optional config dict is accepted by the constructor. The only supported key is lang, which sets the default language when yes_or_no is called without an explicit lang argument:

engine = HeuristicYesNoEngine(config={"lang": "de-de"})

Algorithm

Source: HeuristicYesNoEngine.yes_or_no — ovos_yes_no/__init__.py:71

The engine scans the normalised response against a per-language locale/<lang>/yesno.json resource file containing four word lists:

Decision rules (in order):

1. Last word wins. When both a yes word and a no word appear, the one at the higher character index is taken as the user's final intent.
2. Double negatives. A no-category word immediately followed by a neutral_no word (e.g., not + lie → "not a lie") is interpreted as affirmative.
3. Neutral words. If neither a yes nor a no word was found, neutral_yes and neutral_no words are considered as weak signals.
4. Default. No recognised word → return None.

Language matching uses langcodes.tag_distance to find the closest available locale, so en-AU silently falls back to en-US.

Supported Languages

an, az, ca-ES, cs-CZ, da-DK, de-DE, en-US, es-ES, eu-ES, fa-IR, fr-FR, hu-HU, it-IT, nl-NL, pl-PL, pt-BR, ru-RU, sv-SE, tr-TR, uk-UA

Adding a Language

Create ovos_yes_no/locale/<lang-TAG>/yesno.json with the four word lists described above:

• neutral_yes: mild agreement words that are positive but not direct synonyms of "yes" (e.g., French "bien sur", Portuguese "claro").
• neutral_no: words that imply disapproval indirectly (e.g., French "mensonge", Portuguese "errado").
• Double-negative structures vary widely between languages — test them explicitly.

Limitations

• No sarcasm or idiom detection.
• Vocabulary is limited to the words in the resource files; unlisted slang is ignored.
• Complex nested negations beyond one level may yield incorrect results.
• Missing or incomplete resource files cause the engine to return None for that language rather than raising an error.

License

Apache 2.0