Predict and Use Latent Patterns for Short-Text Conversation

Hung-Ting Chen, Yu-Chieh Chao, Ta-Hsuan Chao, Wei-Yun MaarXiv 2020 · OpenReview

TL;DR

This paper introduces an end-to-end approach for generating more controlled and informative short-text conversations by leveraging complex latent semantic patterns, specifically latent sentences and part-of-speech (POS) sequences. By predicting and incorporating these richer semantic forms, their model significantly improves response quality, outperforming existing methods in relevance and informativeness.

What Problem Does This Paper Address?

Current neural network models for conversational AI, while capable of generating fluent text, often produce generic or meaningless responses (e.g., "I don't know"). This stems from a lack of fine-grained control over the generation process. While some prior work has used simple semantic controls like keywords or latent words, these offer limited guidance. The challenge is to imbue these models with a deeper understanding of desired response characteristics, enabling them to generate responses that are not only fluent but also relevant, informative, and contextually appropriate, without resorting to non-end-to-end retrieval-based methods.

Methods

The core innovation lies in predicting and utilizing "latent semantic sequences" to guide the dialogue generation. The proposed model has two main components:

  1. Latent Sequence Predictor: This module predicts latent semantic forms. The paper explores two types:
  2. Latent Sentences: Instead of single words, they use entire sentences as semantic guides, providing richer contextual and word usage information.
  3. Latent Part-of-Speech (POS) Sequences: Inspired by prior work in poetry generation, they use POS sequences to guide the grammatical structure and flow of the generated response. This predictor is trained to classify or generate these latent sequences based on the input post.

  4. Dialogue Generator: This component takes the input post and the predicted latent semantic sequence to generate the final response. They employ a sequence-to-sequence architecture, adapting it to incorporate the latent semantics. For latent sentences, they use a pointer-generator network with attention over both the post and the latent sentence. For POS sequences, they concatenate the POS sequence with the input post and feed it into a Transformer model.

Crucially, the model is trained end-to-end using reinforcement learning (RL), specifically the REINFORCE algorithm. This allows the latent sequence predictor to learn policies that maximize rewards based on the quality of the generated response, effectively optimizing for better latent semantic choices.

Datasets

The primary dataset used for training and evaluation is the Weibo Benchmark Dataset, an open-domain Chinese dialogue dataset containing over 400 million training pairs. A testing set of 3,200 data pairs is used. The paper notes they use a modified version of this dataset, as released by Gao et al. (2019a).

Key Insights and Ideas

  • Richer Semantics for Better Control: The paper strongly advocates for using more complex latent forms (sentences, POS sequences) over simple keywords or single latent words. This richer guidance provides a more robust framework for controlling response generation.
  • POS Sequences as a Sweet Spot: While latent sentences show promise, the POS sequence approach appears to strike a better balance. It provides structural guidance without being overly restrictive, leading to strong performance.
  • End-to-End RL for Latent Optimization: Fine-tuning the entire system with RL is critical. It allows the latent sequence predictor to learn how to choose latent forms that actually lead to better responses, rather than just relying on pre-trained semantic predictors.
  • Balancing Faithfulness and Freedom: The RL fine-tuning process helps the model learn to balance strictly following the latent semantic guidance with generating more freely, leading to more natural and less repetitive outputs.

Key Results

  • Automatic Evaluation (BLEU Scores): The "Ours-RL (sample POS)" model achieved the highest BLEU-1 and BLEU-2 scores among all tested models, indicating strong performance in word-level matching. This suggests the POS sequences effectively narrow down word choices.
  • Human Evaluation: The model using latent POS sequences (specifically, the "Ours(select POS)" variant) received the highest scores for relevance and informativeness, outperforming strong baselines like HGFU and GMDR. This highlights the model's ability to generate more meaningful and contextually appropriate responses.
  • Faithfulness Analysis: For latent sentence models, they found that the generated responses copied over two-thirds of words and about one-fourth of n-grams from the latent sentence, demonstrating significant influence. For POS sequence models, analysis showed that the model initially strictly followed the POS sequence but later generated more freely during fine-tuning, striking a balance.

Limitations and Future Work

  • Latent Sentence Sensitivity: The latent sentence model's performance was sensitive to the quality of the chosen latent sentence. If the selected sentence was dissimilar to the target response, BLEU scores dropped. This suggests a need for better latent sentence selection mechanisms.
  • Search Space for Latent Sentences: The search space for latent sentences was found to be smaller than some retrieval-based methods, potentially limiting its exploration.
  • Fluency vs. Informativeness Trade-off: While the POS-guided model excelled in relevance and informativeness, human evaluation suggested it might slightly sacrifice fluency compared to some baselines.
  • Future Work: The authors propose exploring a mixed latent sequence approach, combining POS tags and words. They also suggest investigating other semantic forms like Semantic Role Labeling (SRL) for even richer control.

Summary updated

Abstract

Many neural network models nowadays have achieved promising performances in Chit-chat settings. The majority of them rely on an encoder for understanding the post and a decoder for generating the response. Without given assigned semantics, the models lack the fine-grained control over responses as the semantic mapping between posts and responses is hidden on the fly within the end-to-end manners. Some previous works utilize sampled latent words as a controllable semantic form to drive the generated response around the work, but few works attempt to use more complex semantic patterns to guide the generation. In this paper, we propose to use more detailed semantic forms, including latent responses and part-of-speech sequences sampled from the corresponding distributions, as the controllable semantics to guide the generation. Our results show that the richer semantics are not only able to provide informative and diverse responses, but also increase the overall performance of response quality, including fluency and coherence.
BibTeX
@misc{chen2020predict,
  author = {Hung-Ting Chen and Yu-Chieh Chao and Ta-Hsuan Chao and Wei-Yun Ma},
  title = {{Predict and Use Latent Patterns for Short-Text Conversation}},
  year = {2020},
  eprint = {2010.13982},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2010.13982},
}

Loading related content...

Figure