# Frugal Prompting for Dialog Models

**Authors:** Bishal Santra, Sakya Basak, Abhinandan De, Manish Gupta, Pawan Goyal  
**arXiv:** [2305.14919](https://arxiv.org/abs/2305.14919)  
**Date:** 2023-05-24  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.14919v1/

## Key points

- Study explores different approaches for building dialog systems using large language models (LLMs) in NLP tasks
- Experimentation with various aspects of the prompt, including instructions, exemplars, current query, and additional context
- Analysis of representations of dialog history with optimal usable-information density
- Use of Sentence Transformers to measure overall similarity between utterances
- Consideration of using a summary of the full dialog history as an alternative approach
- Finetuning BART and Pegasus models on generic and dialog datasets for generating informative and concise summaries
- Addressing the shortening of background information often included in dialog datasets using BART and Pegasus models
- Utilization of two dialog datasets: Multi-session Chat (MSC) and Topical Chat (TC)
- Normalization of utterances by removing trailing whitespaces and capitalizing the first word of every sentence
- Challenges in dialog summarization due to dynamic and context-dependent conversations.

## Summary

This study explores different approaches for building dialog systems using large language models (LLMs) in natural language processing (NLP) tasks. The researchers aim to understand the conversational abilities of LLMs and how they can be effectively used for interactive systems. They experiment with various aspects of the prompt, including providing instructions, exemplars, current query, and additional context. Additionally, the research analyzes the representations of dialog history that have optimal usable-information density. To measure the overall similarity between utterances, Sentence Transformers are used. An alternative approach considered is using a summary of the full dialog history. Two Transformer-based encoder-decoder abstractive summarization models (BART and Pegasus) are finetuned on generic as well as dialog datasets like CNN/DailyMail, SAMSum, and DialogSum to generate informative and concise summaries. The study also addresses shortening background information often included in dialog datasets such as persona information, reading sets, and knowledge facts. BART and Pegasus models are utilized to shorten this background information. For experimental setup, two dialog datasets are used: Multi-session Chat (MSC) and Topical Chat (TC). These datasets were chosen due to their varying characteristics and length of dialog history. The MSC dataset consists of multiple chat sessions where participants learn about each other's interests and discuss what they have learned from past sessions. Each user plays a role or persona during these conversations. In contrast, the TC dataset assigns users one or more topics along with associated facts or knowledge about those topics for conversation. In both datasets, utterances are normalized by removing trailing whitespaces and capitalizing the first word of every sentence. BART and Pegasus models are employed for summarization of dialog history and background information. Dialog summarization presents challenges due to conversations being dynamic and context-dependent.