01 · GRASP: Graph-Retrieval Automated Scoring Pipeline for Label-Free Multi-Topic Essay Grading

GRASP：用于无标注多主题问答评分的图检索自动评分流水线

Research record	Details
Authors	Aafreen Husain, Samar Shailendra, Saad Sajid Hashmi
Published	2026-09-03
Sources	arxiv
Focus	graph-augmented retrieval, label-free multi-topic grading, semantic similarity graph, RAG 图增强检索, 无标注多主题评分, 语义相似度图, 检索增强生成（RAG）

English

Recommended for a skim: the paper contains practical graph-augmented retrieval and matching techniques that could inform engineering choices in the thesis, but it lacks the key contributions the thesis targets (persistent triple-based graph memory, entity-resolution experiments, and multi-hop QA).

中文

建议略读：该文提供了可借鉴的图增强检索与匹配工程技术，可能对实现细节有帮助，但它缺乏论文关注的核心贡献（基于三元组的持久图记忆、实体解析实验和多跳问答评估）。

Research synopsis

English

This paper presents GRASP, a Graph-Retrieval Automated Scoring Pipeline for grading label-free multi-topic short-answer science exams. Reference answers are encoded into a FAISS index using Sentence-BERT and organized into a semantic similarity graph. At grading time, sentence-count heuristics (with an LLM to resolve ambiguity) segment student essays; top cosine matches are retrieved and then expanded by graph traversal over strong edges (GRAG). The Hungarian algorithm assigns reference nodes to segments, and each segment is graded with GPT-4.1-mini. The contribution is an unsupervised, graph-augmented retrieval workflow that aims to improve retrieval quality and downstream grading accuracy without training data.

中文

本文提出 GRASP，一种用于无标注多主题简答科学试卷评分的图检索自动评分流水线。参考答案通过 Sentence-BERT 编码并存入 FAISS 索引，同时构建语义相似度图。评分时先基于句子计数启发式方法（并在歧义时用 LLM 解决）对学生段落分段；检索采用余弦相似度取得候选参考节点，随后通过强边的图遍历（GRAG）扩展检索结果。使用匈牙利算法将参考节点分配到段落，最终每段用 GPT-4.1-mini 评分。该工作在无训练数据情况下提出了可提高检索质量和下游评分准确性的图增强检索流程。

Thesis relevance

English

Overlap: GRASP uses graph-augmented retrieval and graph traversal to expand initial embedding-based matches, which directly relates to the thesis’s interest in graph-based retrieval improving multi-step tasks. Differences: GRASP builds a semantic similarity graph over reference answers for single-interaction retrieval and grading, not a persistent knowledge graph of subject-predicate-object triples assembled across interactions; it does not treat multi-hop QA or entity resolution across surface forms. Complementarity: the paper’s GRAG traversal and post-retrieval matching (Hungarian algorithm) offer practical retrieval and assignment techniques that could be adapted to mapping retrieved graph nodes or paths to multi-hop query subgoals in the thesis system.

中文

重合点：GRASP 使用图增强检索和图遍历以扩展初始基于嵌入的匹配，这与论文中利用图记忆改善多步检索的目标直接相关。差异：GRASP 构建的是基于参考答案的语义相似度图用于单次检索与评分，而非跨交互累积的、以主谓宾三元组为单位的持久知识图谱；它未处理多跳问答或不同表面形式的实体消歧。互补性：其 GRAG 遍历策略和检索后匹配（匈牙利算法）是可借鉴的工程方法，可用于将检索到的图节点或路径映射到论文系统中的多跳查询子目标。

Writing and related work

English

Position GRASP as an applied retrieval-methods paper in the automated grading literature: it demonstrates that lightweight graph augmentation of vector indexes can improve downstream task alignment without supervised training. Note that the graph encodes similarity among reference answers rather than typed relations or inferred triples, which should be made explicit when comparing to knowledge-graph approaches.

中文

将 GRASP 定位为自动评分领域的检索方法应用论文：它展示了对向量索引进行轻量图增强可以在无监督情况下改善下游任务的对齐。需要明确的是，该图编码的是参考答案之间的相似性，而非有类型的关系或推断出的三元组，这一点在与知识图谱方法比较时应予以说明。

Method and evaluation

English

Adaptable techniques: (1) seed-and-expand graph traversal over strong similarity edges (GRAG) can be repurposed to expand initial entity or passage matches into multi-step candidate paths; (2) use of the Hungarian algorithm to optimally assign retrieved nodes to query segments suggests a way to align retrieved paths to decomposed subquestions. Empirical evaluation ideas: implement GRAG-style traversal on your constructed triple graph (seeded by initial retrieval) and measure its effect on bridge-entity identification, path precision, and end-to-end QA accuracy versus pure vector RAG.

中文

可迁移的方法： (1) 以种子节点为起点在强相似边上进行扩展的图遍历（GRAG）可以改造为将初始实体或段落匹配扩展为多步候选路径；(2) 使用匈牙利算法将检索节点最优分配到分解后的查询片段，可用于将检索到的路径与子问题对齐。可行的评估思路：在你构建的三元组图上实现类似 GRAG 的遍历（以初始检索为种子），并比较其在桥接实体识别、路径精确度和端到端 QA 准确率上的效果，与纯向量 RAG 进行对比。

Future directions

English

Extend GRASP-style graph augmentation to a persistent, typed graph by extracting triples from references and student text, add entity-resolution to avoid fragmentation, and evaluate on explicit multi-hop benchmarks (e.g. MultiHop-RAG) to test reuse of earlier-discovered intermediate entities and paths. Also compare embedding-only expansion vs. LLM-mediated judgment for ambiguous matches when building or traversing the graph.

中文

将 GRASP 风格的图增强扩展为持久的、有类型边的图：从参考答案和学生文本中提取三元组，加入实体解析以避免图的碎片化，并在显式多跳基准（如 MultiHop-RAG）上评估早先发现的中间实体与路径的重用效果。此外，可比较在构建或遍历图时仅用嵌入扩展与由 LLM 判定的匹配策略的差异。

---