Automatic Extraction of Relationships among Motivations, Emotions and Actions from Natural Language Texts

TL;DR

Addresses the need to explicitly model how motivations, emotions, and actions relate in text, beyond opaque neural models. Proposes MEA-DAG, a graph-based framework built from a manually designed Nature Design and learnable Nurture Belief, with LLM-assisted mappings to map text to directed graphs. Applies the approach to Amazon Fine Foods Reviews to generate 92,990 MEA-DAGs, achieving 63% logical validity and revealing key bottlenecks in event linking, extraction, negation, and word-sense ambiguity. Demonstrates a white-box alternative to annotation-heavy methods and outlines directions for expanding node coverage, improving robustness, and addressing error sources.

Abstract

We propose a new graph-based framework to reveal relationships among motivations, emotions and actions explicitly given natural language texts. A directed acyclic graph is designed to describe human's nature. Nurture beliefs are incorporated to connect outside events and the human's nature graph. No annotation resources are required due to the power of large language models. Amazon Fine Foods Reviews dataset is used as corpus and food-related motivations are focused. Totally 92,990 relationship graphs are generated, of which 63% make logical sense. We make further analysis to investigate error types for optimization direction in future research.

Figures (10)

• Figure 1: (a) Nature Design. This graph reveals the interactive mechanism among motivations, emotions and actions in human's nature. (b) An MEA-DAG example. Events are presented in green color and the activated nodes of Nature Design are in red color. Other nodes are omitted. Nurture Belief participates in linking events to corresponding nodes, and explicitly presented in the MEA-DAG. For instance, "it seriously makes perfect meatballs", has a connection with #food by the belief tuple ("meatball", #food).
• Figure 2: Framework of computing a MEA-DAG. It inputs review texts (top left corner) and outputs a graph (bottom right corner). The green line at the bottom shows the evolution of a MEA-DAG in different processing stages.
• Figure 3: Prompt engineering. We rely on LLM to accelerate the establishment of Nurture Belief, and no longer rely on manual labeling resources.
• Figure 4: Events extracted by ASER are in blue color. We use a green font background to highlight the events incorporated in MEA-DAG. (a): Critical events "bags are more convenient when I'm at work", "it's inexpensive" are not included in MEA-DAG. (b): Critical event "these are fabulous" are not included in MEA-DAG.
• Figure 5: Texts in blue color are the events extracted by ASER. (a): Critical events "the diet kind had that funny taste", "it's way cheaper then the jugs in the stores" and "it's very simple to mix & it stays mixed" are not captured by ASER. (b): Critical events "would definitely purchase them again", "I would also recommend them" and "the cost is more reasonable" are not captured by ASER.