Skip to content

Latest commit

 

History

History
106 lines (74 loc) · 5.21 KB

Meena.md

File metadata and controls

106 lines (74 loc) · 5.21 KB
Raw

Towards a Human-like Open-Domain Chatbot

a.k.a. Meena

Adiwardana et al., Google Research, Brain Team

arXiv 2020

References

Summary

  • Presents Meena, a 2.6B-parameter Transformer language model trained on public domain social media conversations.
    • Data: 40B words (61B subwords), most likely sourced from Reddit.
    • Model: A seq2seq Evolved Transformer model with a 1-layer encoder (for context) and a 14-layer decoder (for response). 2.6B is 1.7x the number of parameters in the largest GPT-2 model.
    • 7% increase in SSA after filtering and tuned decoding.
  • Introduces a human evaluation metric called Sensibleness and Specificity Average (SSA), which closely correlates with model perplexity.
    • Human: 86%, Best Meena: 79%, End-To-End Meena: 72%
    • Both Meena models have significantly higher SSA than existing neural chatbots (Cleverbot, Mitsuku, DialoGPT, XiaoIce).

Sensibleness and Specificity Average (SSA)

Definition

  • Sensibleness: Does this response make sense?
    • (Claim) includes common sense, logical coherence, and coherence.
  • Specificity: Is this response specific to context (i.e., cannot be used in "dozens of different contexts")?
    • (Claim) includes diversity of responses.
  • SSA = (sensibleness + specificity) / 2

Human Evaluation

SSA is a human evaluation method. Given a context and a response, humans are asked:

  1. Is this response sensible?
    1. (If Yes) Is this response specific?
    2. (If No) considered not specific either.

Justifications

  • Humans "reasonably" agree with each other on their sensibleness and specificity labeling.
  • DialoGPT (51%), made of GPT, scores higher than GenericBot (39%), which only answers "I don't know" and "ok".
  • SSA correlates (R^2=0.96) with human likeness, measured simply by asking humans "is this response humanlike?".

Static vs. Interactive

  • Static: 1,477 conversational contexts between 1 and 3 turns are provided, response is evaluated (= Mini-Turing Benchmark or MTB)
  • Interactive: crowdworkers converse with bots for 14-28 turns about anything they wanted. Each bot is evaluated 100 times and each response is evaluated (i.e., 700+ responses evaluated for SSA).
    • Unlike Turing test, crowdworkers know they're talking with bots.
      • (Claim) This shifts the focus from detecting deception to detecting human-like qualities.
    • Ghandeharioun et al. (2019) instead evaluates the conversation as a whole.

Training Data

  • "Public Domain Social Conversations" (probably Reddit?)
    • Essentially, message trees involving multiple speakers.
    • Treat each message as a response and all previous turns (up to 7) as context.
    • After filtering, this amounts to 867M (context, response) pairs.
    • This amounts to 341GB of text, containing 40B words (or 61B subwords).
    • Best PPL: 10.2 (ET), 10.7 (Transformer)

Model Architecture

  • An uneven seq2seq model with Evolved Transformer blocks.
    • This is to understand the context bidirectionally while focusing on learning to generate.
  • 1 encoder and 13 decoder blocks
  • 2,560 hidden size with 32 attention heads
  • Maximum of 128 sequence length for encoder and decoder (each)
  • 8K SentencePiece vocabulary
  • Decoding (at test time): sample-and-rank
    • NOT beam search.
    • top-k (k=40) nucleus sampling (Holtzman et al., 2019) is used for the full model (2%+ SSA improvement).

Comparing Transformer and Evolved Transformer Blocks

An Evolved Transformer block is about as twice as large as a Transformer block and contains convolutional layers before self-attention. This block architecture was discovered using Neural Architecture Search (NAS) applied to WMT machine translation. Note that lightweight and dynamic convolutions also involve convolutional layers in a Transformer-like architecture, but the LightConv or DynamicConv layer replaces self-attention layers instead.

Results

  • See samples (in paper and in the link above).
  • Both interactive sensibleness and specificity correlate highly with perplexity, which is better for larger models.
    • Both static SSA (std 0.02) and interactive SSA (std 0.01) were consistent enough across 7 different evaluations.
  • The largest (2.6B) model actually overfits the 61B-token dataset, after 164 epochs.
  • For decoding, sample-then-rank (temperature-based sampling + ppl-based reranking) performs much better than beam search and is "good enough" given large enough models.
    • Sampling: N = 20, T = 0.88 works the best.

Further Improvements (base → full)

Two improvements to push SSA from 72% to 79%:

  • Top-k nucleus sampling (k=40, N=20, T=1.0): 72% → 74%
  • Rule-based restriction on cross-turn repetitions: 74% → 79%

Also added a "safety layer" that classifies potentially sensitive or toxic responses.