RecWizard

A Plug-n-Play Toolkit for Conversational Recommenders with Modular, Portable Models and Interactive User Interface

(accepted at AAAI 2024 Demo Track)

Overview

The RecWizard libary is built for easy deployment/development of Conversational Recommender Systems(CRS). The core idea of RecWizard is to use a "Text-in, Text-out" interface for all kinds of different modules used in a CRS pipeline. The text-based interface gives our package a plug-and-play nature that allows users to easily experiment with new combinations of modules.

Base Concept

We abstract a CRS model into two levels as shown in the figure below:

• Module Level: We typically provide a recwizard.BaseModule for making recommendations and another rectbot.BaseModule for generating natural-language responses. We can further introduce processor rectbot.BaseModule to extract important information (e.g., entity linking) from users' raw features;
• Model Level: We treat a recwizard.BasePipeline as a high-level manager to decide when and how to call the lower-level modules, and how to combine the results from such modules.

Quick start

Take ExpansionPipeline for example, you can build a ChatGPT-based expansion model with the classic AutoRec module ¹ in a few lines:

1.Setup the pipeline

from recwizard import ExpansionConfig, ExpansionPipeline
from recwizard import ChatgptGen, RedialRec

pipeline = ExpansionPipeline(
  ExpansionConfig(),
  rec_module=RedialRec.from_pretrained('recwizard/redial-rec'),
  gen_module=ChatgptGen.from_pretrained('recwizard/chatgpt-expansion')
)

2. Format the input

you would want to format your dialogue history like this before passing it to the pipeline:

context = "<sep>".join([
  "User: Hello!",
  "System: Hello, I have some movie ideas for you. Have you watched the movie <entity>Forever My Girl (2018)</entity> ?",
  "User: Looking for movies in the comedy category. I like Adam Sandler movies like <entity>Billy Madison (1995)</entity> Oh no is that good?"
])

Currently. The formatting contains 3 aspects:

1. We prepend the sender of the message to the beggining of the sentence. Either "User: " or "System: "
2. We join the utterances with the <sep> token. which can be handled differently by different tokenizers
3. (Optional) We mark the entities in the text manually so that the named entities can be extracted by the downstream modules.

3. Get Response!

print(pipeline.response(context))

System: If you're a fan of Adam Sandler, you might enjoy "50 First Dates (2004)" which is a romantic comedy featuring him and Drew Barrymore. 
Another one is "Big Daddy (1999)" where Sandler's character adopts a kid to impress his girlfriend, but things don't go as planned. 
Both movies have Sandler's signature humor and are quite entertaining.

Implemented Modules

Base Model	Recommender Module	Generator Module
ReDIAL	redial-rec	redial-gen
KBRD	kbrd-rec-redial	kbrd-rec-redial
UniCRS	unicrs-rec-redial/unicrs-rec-inspired	unicrs-gen-redial/unicrs-gen-inspired
KGSF	kgsf-rec	kgsf-gen
ChatGPT	chatgpt-rec-fillblank	chatgpt-gen-fillblank/chatgpt-gen-expansion
Llama2		llama-expansion

Code Structure

recwizard
├── __init__.py
├── configuration_utils.py
├── model_utils.py
├── module_utils.py
├── modules
│   ├── __init__.py
│   ├── llm
│   │   ├── __init__.py
│   │   ├── configuration_chatgpt_gen.py
│   │   ├── configuration_chatgpt_rec.py
│   │   ├── modeling_chatgpt_gen.py
│   │   ├── modeling_chatgpt_rec.py
│   │   └── tokenizer_chatgpt.py
│   ├── kbrd
│   │   ├── configuration_kbrd_gen.py
│   │   ├── configuration_kbrd_rec.py
│   │   ├── modeling_kbrd_rec.py
│   │   ├── shared_encoder.py
│   │   └── tokenizer_kbrd_rec.py
│   ├── redial
│   │   ├── __init__.py
│   │   ├── autorec.py
│   │   ├── beam_search.py
│   │   ├── configuration_redial_gen.py
│   │   ├── configuration_redial_rec.py
│   │   ├── hrnn.py
│   │   ├── hrnn_for_classification.py
│   │   ├── modeling_redial_gen.py
│   │   ├── modeling_redial_rec.py
│   │   ├── params.py
│   │   ├── tokenizer_redial_gen.py
│   │   ├── tokenizer_redial_rec.py
│   │   └── tokenizer_rnn.py
│   └── unicrs
│       ├── __init__.py
│       ├── configuration_unicrs_gen.py
│       ├── configuration_unicrs_rec.py
│       ├── kg_prompt.py
│       ├── modeling_unicrs_gen.py
│       ├── modeling_unicrs_rec.py
│       ├── prompt_gpt2.py
│       ├── tokenizer_unicrs_gen.py
│       └── tokenizer_unicrs_rec.py
├── pipelines
│   ├── __init__.py
│   ├── chatgpt
│   │   ├── __init__.py
│   │   ├── configuration_chatgpt_agent.py
│   │   └── modeling_chatgpt_agent.py
│   ├── expansion
│   │   ├── __init__.py
│   │   ├── configuration_expansion.py
│   │   └── modeling_expansion.py
│   ├── fill_blank
│   │   ├── __init__.py
│   │   ├── configuration_fill_blank.py
│   │   └── modeling_fill_blank.py
│   └── switch_decode
│       ├── __init__.py
│       ├── configuration_switch_decode.py
│       └── modeling_switch_decode.py
├── tokenizer_utils.py
└── utility
    ├── __init__.py
    ├── constants.py
    ├── device_manager.py
    ├── entity_linking.py
    ├── singleton.py
    ├── utils.py
    └── monitor.py

RoadMap

• Implement KBRD Recommender
• Implement KBRD Generator
• Improve the BaseTokenizer API (simplify)
• Provide detailed examples of tokenization