Skip to content

About

Portuguese interaction research for Vbot: phrase sandbox, deliberate abstentions, confusion matrix and per-intent metrics. Five synthetic benchmarks, English/Portuguese.

Topics

Resources

Contributing

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

Portuguese Interaction Bench

Every phrase deserves a clear response.
Cada frase merece uma resposta clara.

Checks Node.js 22+ Five scenarios MIT license

Open the interactive study · Abrir estudo em português · Guided walkthrough · Reproduce the examples

Actual workbench screenshot

English · Português

English

Inspect a small Portuguese text baseline before a robot trial. Explore successful matches, ambiguous requests and deliberate abstentions.

Research question: Which phrases does an explicit dictionary understand, and where does it abstain?

This is a working, independent software study for a proposed Vbot companion-robotics program. You can run every example without a robot, account, package installation or API key. Physical validation has not been performed.

Explore in one minute

  1. Open Everyday requests.
  2. Inspect the chart and the explanation below it. The starting result is 12/16 correct · 4 abstentions.
  3. Adjust a control or select a contrasting example. Download a printable HTML report, a CSV evidence table or the full JSON analysis.

View a portable example report.

The JSON editor is available under Advanced; the main study uses guided controls.

What you can investigate

  • A deliberately small, inspectable Portuguese phrase dictionary and full-string matching.
  • A live text sandbox that explains the operator's next step without issuing commands.
  • A confusion matrix, per-intent precision/recall/F1 and supported-class macro F1.
  • Known phrases, normalization, negation/compound cases and a difficult held-out fixture.
  • Visible abstentions and errors, including examples where the dictionary performs poorly.

All five studies include a responsive English/Portuguese interface, keyboard focus styles, synthetic narrated examples and local report exports. Analysis runs in your browser after the page loads.

Five examples, five questions

Example Question it exposes Reproducible core result
First six requests A small deterministic fixture with known and negated requests. 6/6 correct · 2 abstentions
Everyday requests Declared dictionary variants and phrases it does not cover. 12/16 correct · 4 abstentions
Negation and compound requests Unrecognized requests should ask for clarification. 9/10 correct · 8 abstentions
Accents and punctuation Test normalization with explicit expected labels. 9/10 correct · 2 abstentions
Beyond the dictionary A deliberately difficult fixture separates example correctness from generalization. 2/10 correct · 10 abstentions

These are deterministic scenario results, not field measurements. Experiment notes distinguish expected regression outputs from independent arithmetic and failure-case tests.

Run locally

Install Node.js 22 or newer. No npm dependencies are required.

git clone https://github.com/galafis/vbot-ptbr-interaction-bench.git
cd vbot-ptbr-interaction-bench
npm run serve

Open http://127.0.0.1:8765. The local server binds to your computer only. Stop with Ctrl+C. For another port, use node server.mjs 8766.

npm run check                                  # syntax, domain tests, examples and server checks
node cli.mjs examples/everyday-requests.json            # core analysis
node study.mjs examples/everyday-requests.json          # analysis plus derived insights
node cli.mjs examples/baseline.json new-result.json

The CLI refuses to overwrite an existing output file. Exit 1 means invalid input or an I/O failure; review findings are valid analytical outputs and do not change the exit code.

How the study works

flowchart LR
  A[Portuguese text and expected intent] --> B[Normalize accents and punctuation]
  B --> C[Exact declared phrase rules]
  C --> D[Intent or abstention]
  D --> E[Confusion matrix]
  E --> F[Precision, recall and macro F1]
  D --> G[Operator review only]
Loading

Method: Exact phrase rules, a confusion matrix and per-intent precision/recall. The core engine and the insight layer are pure local calculations shared by browser and CLI. Architecture · Domain method · Input contract.

Evidence and scope

Evidence available now Next evidence needed
19 automated tests, five reproducible fixtures and a real interface screenshot Measurements from a specific supplied Vbot unit
Deterministic calculations and explicit counterexamples Repeated permitted exercises with operator observations
Bilingual methods, reports and example inputs Appropriate authorization before institutional or participant studies

A standard Vbot without an open SDK is welcome. The proposed first hardware stage uses native features, the official app and manual observations. Any later programmatic integration is optional and must be permitted by the manufacturer. Native workflow · 90-day study plan.

Project map

src/        Pure domain engine, contracts and derived insights
examples/   Five narrated inputs and reproducible expected outputs
app.mjs     Guided bilingual browser workbench
styles.css  Responsive, printable visual system
cli.mjs     Core analysis from a local JSON file
study.mjs   Complete study from a local JSON file
tests/      Domain, insight, reproduction and HTTP checks
docs/       Walkthrough, experiments, method and study plan

Português

Examine uma baseline textual em português antes de um ensaio com o robô. Explore acertos, pedidos ambíguos e abstenções deliberadas.

Pergunta do estudo: Quais frases um dicionário explícito reconhece e quando ele se abstém?

O projeto está pronto para exploração em software. Os exemplos são fictícios e a validação física ainda depende de uma unidade. Uma unidade comum, sem SDK aberto, é suficiente para iniciar a fase proposta com recursos nativos e observação manual.

Comece por um exemplo concreto

Abra Pedidos do cotidiano. O resultado inicial é 12/16 acertos · 4 abstenções. Observe o gráfico, mude um parâmetro ou escolha outro cenário. A interface explica os resultados e exporta relatório HTML imprimível, tabela CSV e análise JSON.

Funcionalidades

  • Dicionário pequeno e inspecionável de frases em português, com correspondência integral.
  • Teste de frases na interface, com explicação da próxima ação do operador.
  • Matriz de confusão, precisão/recall/F1 por intenção e macro F1 das classes com suporte ou previsão.
  • Casos conhecidos, normalização, negação e frases difíceis fora do dicionário.
  • Abstenções e erros ficam visíveis, inclusive quando o desempenho é baixo.

Exemplos disponíveis

Exemplo O que permite examinar Resultado reproduzível
Primeiros seis pedidos Uma pequena verificação determinística com pedidos conhecidos e negados. 6/6 acertos · 2 abstenções
Pedidos do cotidiano Variações previstas no dicionário e frases que ele ainda não cobre. 12/16 acertos · 4 abstenções
Negação e pedidos compostos Pedidos não reconhecidos devem pedir esclarecimento. 9/10 acertos · 8 abstenções
Acentos e pontuação Teste a normalização mantendo os rótulos esperados explícitos. 9/10 acertos · 2 abstenções
Fora do dicionário Uma avaliação deliberadamente difícil evita confundir acerto no exemplo com generalização. 2/10 acertos · 10 abstenções

Executar e contribuir

Use os comandos da seção em inglês com Node.js 22 ou superior. Não é necessário instalar pacotes. npm run check executa verificações de sintaxe, testes e reprodução dos cenários. O guia DEMO explica um percurso de exploração; EXPERIMENTS registra os resultados e suas limitações.

Sugestões úteis incluem novos cenários fictícios, cálculos independentes, melhorias de acessibilidade da interface e protocolos de observação autorizados. Veja CONTRIBUTING.

Study collection · Coleção de estudos

01 · Companion Follow Lab · 02 · Portuguese Interaction Bench · 03 · Community Cargo Planner · 04 · Inclusive Encounter Kit · 05 · Community Observation Lab


Created by Gabriel Demetrios Lafis · Applied AI and robotics portfolio

MIT-licensed independent work. Vbot is a third-party brand; this is not an official SDK or an announced Vbot/USP partnership.

About

Portuguese interaction research for Vbot: phrase sandbox, deliberate abstentions, confusion matrix and per-intent metrics. Five synthetic benchmarks, English/Portuguese.

Topics

Resources

Contributing

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages