When a paper's abstract never uses the words in a category's name, a classifier that matches words cannot place it. Explicit Semantic Analysis (ESA) helps here. ESA represents a text by how strongly it matches each article of an encyclopedia, which gives a sparse, readable "concept fingerprint". Two texts with no words in common can still activate the same concepts, such as Gamma-ray burst and Quasar. Its known weakness is noise. A fingerprint picks up many loosely related concepts, and adding them unfiltered pulls retrieval off-topic. The fix that works, from Concept-Based Feature Generation and Selection for Information Retrieval (MORAG), is to treat top-retrieved documents as stand-ins for relevant ones and keep only the concepts that separate them from the rest.
This note proposes a different source of that relevance signal for documents that live in a graph. On a citation network, a paper's cited and citing neighbors are a better guess at "documents about the same thing" than a first-pass keyword search. We use the neighborhood in two ways. It prunes concepts that a node's neighbors do not share, which controls drift. It also lets a node borrow concepts its neighbors activate strongly but its own text misses, which bridges the vocabulary gap. The result is a training-free, interpretable concept profile for each node. We test it on classifying graph nodes from their label names alone, where the gap between label vocabulary and node text is the main obstacle.
Core Claim
Graph neighborhoods can do the job that pseudo-relevance feedback does in ESA, and do it better. Neighborhood-guided selection and borrowing of explicit concepts should beat both unfiltered ESA and pseudo-relevance-filtered ESA on label-name ("dataless") node classification over text-attributed graphs, which are graphs whose nodes carry text. The gains should come mainly from nodes whose text shares no content words with their label's description.
The claim against dense methods is more limited. The filtered explicit profiles should match a dense sentence-embedding pipeline that uses the same propagation over the graph, and should add accuracy when fused with it. If they only match it, the contribution reduces to interpretability and drift control, and the experiments are built to say which outcome holds.
Relation to Prior Work
Three lines of work motivate the gap.
Classical ESA, from Gabrilovich and Markovitch, and its use for dataless classification by Chang et al. treat each document as an isolated text. Chang et al. project documents and label names into Wikipedia concept space and classify by nearest neighbor, reaching 85.29% on 20 Newsgroups and 88.62% on Yahoo! Answers with no training data. Their method has no way to use links between documents.
MORAG supplies the method we reuse: concept selection by information gain against pseudo-relevant and pseudo-nonrelevant sets, computed over passages. It also supplies the constraint we must respect. Unfiltered ESA expansion hurt one retrieval engine (PIRCS) by 3.2%, while filtered expansion raised mean average precision by 18.8% on another (Xapian). We keep MORAG's selection rule and replace where its positive set comes from.
Hybrid text-graph models fill the other side. GraphFormers nests graph aggregation inside Transformer layers. GLEM alternates training a language model and a graph network. TAPE has a large language model write explanations of each node, which a smaller model then encodes into vectors for a graph network. All three show that neighbors disambiguate node text, but they do so in uninterpretable dense spaces and need labels.
STAG and PromptGFM are the closest work. Both turn graph nodes into discrete vocabularies: STAG maps nodes onto subword tokens from a frozen language model, and PromptGFM maps them onto natural-language ID strings. Neither uses encyclopedic concepts, and neither targets the vocabulary gap between label names and node text.
Our method needs no gradient training. Every dimension it keeps is a named Wikipedia article, and graph structure enters only as a feature-selection and borrowing signal.
Mechanism
Let $x_v$ be the ESA vector of node $v$, computed with the passage-level indexing used by MORAG. Each coordinate $x_v(c)$ is the BM25 match score between $v$'s text and Wikipedia article $c$.
For each node we score every concept by information gain. The positive set is $v$'s graph neighborhood $N(v)$, out to one or two hops. The negative set is an equal-size random sample of non-neighbors. This gives a mask $m_v(c)=1$ when the information gain exceeds a threshold $\tau$, and $0$ otherwise. The node's profile is then
$$z_v = m_v \odot \Big( (1-\alpha)\, x_v + \alpha \frac{1}{|N(v)|}\sum_{u \in N(v)} x_u \Big),$$
where $\alpha \in [0,1]$ sets how much a node borrows from its neighbors and $\odot$ is coordinate-wise product. The mask removes concepts the neighborhood does not support (pruning). The averaging term adds supported concepts that the node's own text lacks (borrowing).
Labels are fingerprinted in the same space from their names and short descriptions, such as arXiv category blurbs. Classification takes the label with the highest cosine similarity to $z_v$. An optional second step propagates these zero-shot scores over the graph once. The parameters $\tau$ and $\alpha$ are tuned on a small held-out set of labeled nodes, and that same budget is given to every baseline.
A fallback guards against graphs where neighbors often have different labels (heterophily). When few concepts pass the threshold for a node, the mask degrades toward the unfiltered vector rather than an empty one, so a poor neighborhood should cost little.
Research Program
The work has three parts.
The first is infrastructure. We build passage-level ESA indexes over a fixed Wikipedia snapshot in the style of EasyESA. For the 2023 arXiv graph we use a snapshot dated before its papers, so the encyclopedia cannot contain them.
The second separates pruning from borrowing. We run each alone and together, and compare neighbor-derived positive sets with MORAG's retrieval-derived ones.
The third measures where gains come from. We stratify nodes by content-word overlap with their label description, in the spirit of the WikiQA finding that 20.3% of correct answers share no content words with the question. We also hand-judge drift: a sample of each node's top-20 concepts is labeled on-topic or off-topic, before and after selection.
A secondary arm tests the classical complaint that ESA treats related concepts as unrelated axes. It smooths profiles over Wikipedia's own hyperlink graph and asks whether this adds anything once the document graph is used.
Baselines
The naive reference points are label-name keyword matching with BM25 and the majority class. Keyword matching sets the floor that the vocabulary gap imposes.
The direct ancestor is NN-ESA dataless classification from Chang et al., run on the same index and label descriptions. Its reported 85.29% on 20 Newsgroups is the strongest published number for this family. We reproduce it on 20 Newsgroups, which has no graph, to validate our index before moving to graphs.
The selection baseline is ESA with MORAG's pseudo-relevance information-gain filter, using the identical selection rule and threshold search. This isolates our key ingredient, which is replacing retrieved positives with graph neighbors.
The smoothing ablation averages ESA vectors over neighbors without the mask ($m_v \equiv 1$). It tests whether selection matters beyond plain propagation.
The strongest competitor is the dense twin. We take SentenceBERT embeddings of node text and label descriptions, apply the same neighbor averaging, and score labels by cosine similarity. It uses the same graph, the same label text and the same tuning set. It is the strongest competitor because dense embeddings already cluster synonyms, so the vocabulary gap may simply not exist for it.
We also report zero-shot prompting of an LLM on tape-arxiv23, the 2023 arXiv graph built by the TAPE authors, where label leakage from pretraining is reduced. Where STAG's code can be run, we include its zero-shot transfer. The file reports only that STAG beats OFA by 20–30% on hard transfers, with no absolute accuracy.
I found no reported result on Lacuna for label-name zero-shot classification on ogbn-arxiv, tape-arxiv23, Cora, PubMed or WikiCS, so all graph baselines will be run by us under matched conditions.
Beating a baseline means a gain in macro-F1, not only accuracy, over five seeds of the random negative samples and the tuning split, with non-overlapping confidence intervals. Against the MORAG filter and the smoothing ablation, the gain must appear in the zero-overlap stratum. Against the dense twin, we count either parity alone plus a significant gain when the two are fused, or an outright win.
Why not just…?
Why not just use dense embeddings with graph propagation? This is the real threat. Nothing in the file shows explicit concepts beating modern dense encoders on node classification. The failures of SentenceBERT reported by G-Retriever concern subgraph construction for question answering, not label matching. The dense twin is therefore run with the identical propagation operator. If it wins outright, we report that explicit profiles are not worth their cost for accuracy, and fall back to measuring whether fusion helps and whether the readable concepts support auditing. We do not argue around it.
Why not just have an LLM explain each node, as TAPE does? TAPE's explanations are free text, re-encoded into opaque vectors. They need labels downstream and per-dataset prompts, and the authors built tape-arxiv23 precisely because leakage clouds results on older arXiv data. The comparison is run on tape-arxiv23, against a pre-2023 Wikipedia, to settle it on clean ground.
Why not just spread activation over Wikipedia's link graph? That addresses the axes-are-unrelated complaint without any document graph, and it may help. But it smooths every concept in every node, including the noisy ones, which is the drift MORAG warns about. It appears as our secondary arm. If it matches neighbor selection, the document graph is unnecessary and we will say so.
Why not just keep MORAG's retrieval-based feedback? Retrieval-based feedback inherits the vocabulary gap of the first-pass search: if the node's words miss, its pseudo-relevant set is wrong too. Citations are written by authors, not retrieved by words. This contrast is the central ablation.
Proposed Experiments
The graph experiments use ogbn-arxiv from the Open Graph Benchmark, with 40 categories and its time-based split, as the primary graph benchmark. tape-arxiv23 is the leakage-controlled test. Cora, PubMed and WikiCS test breadth. 20 Newsgroups and Yahoo! Answers serve as no-graph controls, where neighbor selection is unavailable and the method should reduce to MORAG-style filtering.
Three checks decide the outcome.
The first is the overlap stratification. Neighbor selection must beat both MORAG filtering and plain smoothing by its largest margin in the zero-overlap bucket. If the margin is uniform across buckets, the method is not specifically closing the vocabulary gap.
The second is a drift audit. Hand-judged off-topic concepts should fall under neighbor masking and rise under plain smoothing. This answers whether graph structure filters semantic noise or spreads it.
The third is edge rewiring. Replacing 0–100% of edges with random ones should erode the advantage over MORAG filtering monotonically, with the fallback keeping performance near unfiltered ESA at full rewiring. If accuracy holds with random edges, the gain was not coming from structure.
Together with the dense-twin comparison, these show whether citation neighborhoods are a usable relevance signal for explicit concept spaces, and how much of the vocabulary gap they close.
References
- Concept-Based Feature Generation and Selection for Information Retrieval
- Computing Semantic Relatedness Using Wikipedia-based Explicit Semantic Analysis
- Importance of Semantic Representation: Dataless Classification
- GraphFormers: GNN-nested Transformers for Representation Learning on Textual Graph
- Learning on Large-scale Text-attributed Graphs via Variational Inference
- Harnessing Explanations: LLM-to-LM Interpreter for Enhanced Text-Attributed Graph Representation Learning
- Quantizing Text-attributed Graphs for Semantic-Structural Integration
- LLM as GNN: Graph Vocabulary Learning for Text-Attributed Graph Foundation Models
- EasyESA: A Low-effort Infrastructure for Explicit Semantic Analysis
- WikiQA: A Challenge Dataset for Open-Domain Question Answering
- G-Retriever: Retrieval-Augmented Generation for Textual Graph Understanding and Question Answering
- Open Graph Benchmark: Datasets for Machine Learning on Graphs