# Integrating AI Planning with Natural Language Processing: A Combination of Explicit and Tacit Knowledge

**Authors:** Kebing Jin, Hankz Hankui Zhuo  
**arXiv:** [2202.07138](https://arxiv.org/abs/2202.07138)  
**Date:** 2022-02-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2202.07138v2/

## Key points

- Natural language processing (NLP) analyzes and processes natural language data to investigate interactions between agents and humans
- Large-scale language models are crucial in NLP but pose challenges of explainability and complexity
- Researchers have been exploring ways to introduce logical relations and rules into NLP models through the integration of AI planning
- Integration of AI planning and NLP can improve communication between human and intelligent agents in five areas: planning-based text understanding, planning-based natural language processing, planning-based explainability, text-based human-robot interaction, and applications
- Text summarization involves extracting important information from texts and generating new summaries based on that information
- Text summarization can be categorized into extractive techniques or abstractive techniques
- Researchers have integrated text summarization models with AI planning by using deep learning abstractive summarization with content planning
- Planning-based text summarization methods first plan out skeletons of summaries or compute probability distributions before generating whole sentences based on those plans or predictions
- Current planning-based text summarization approaches rely on deep learning to learn models and probability distributions, but it is challenging to capture the relations implied in texts.

## Summary

Natural language processing (NLP) is a field that focuses on analyzing and processing large amounts of natural language data to investigate the interactions between agents and humans. Large-scale language models play a crucial role in NLP, but they also pose challenges of explainability and complexity. To address these challenges, researchers have been exploring ways to introduce logical relations and rules into natural language processing models, such as through the integration of AI planning. AI planning, or automated planning, involves building symbolic domain models and synthesizing plans to transition initial states to goals based on those models. Recently, there has been significant research on the integration of AI planning and NLP, which can generate explicit knowledge such as preconditions and effects of action models while also learning from tacit knowledge like neural models. This integration can improve communication between human and intelligent agents in five areas: (1) planning-based text understanding, (2) planning-based natural language processing, (3) planning-based explainability, (4) text-based human-robot interaction, and (5) applications. One area where AI planning has shown promise in NLP is text summarization. Text summarization involves extracting important information from texts and generating new summaries based on that information. It requires agents to understand texts, filter information from abundant descriptions, and organize them into summaries. Text summarization can be categorized into extractive techniques that select subsets of words or sentences from input articles to summarize them or abstractive techniques that try to understand the texts and generate summaries. In recent years, researchers have integrated text summarization models with AI planning by using deep learning abstractive summarization with content planning that describes or predicts skeletons of articles. Planning-based text summarization methods first plan out skeletons of summaries or compute probability distributions before generating whole sentences based on those plans or predictions. For example, some researchers have computed plans in the form of entity chains before generating summaries conditioned on those plans. While current planning-based text summarization approaches rely on deep learning to learn models and probability distributions, it is challenging to capture the relations implied in texts. Researchers believe that it would be interesting and challenging to let content planning modules understand these relations, allowing content plans to be more coherent and controllable. Furthermore, there are commonalities between some mainstream text summarization methods and AI planning. For example, tree-based and graph-based text summarization approaches first find the most important information from the text before using trees and graphs to create summaries.