acceptodds
Withdrawn from ICLR 2027

Kernel Change-Point Detection under Dependence for Unsupervised Text Segmentation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Unsupervised text segmentation is important because boundary labels are expensive, subjective, and often fail to transfer across domains or granularity choices. Text segmentation can be viewed as detecting distributional changes in sequences of pretrained representations, yet existing kernel change-point detection (KCPD) theory largely assumes independent observations. To provide a principled analysis beyond the usual i.i.d. setting, we study penalized KCPD under -dependence, a stylized finite-memory model of short-range dependence, and establish an oracle inequality for the population penalized risk together with a conservative localization guarantee. Building on this perspective, we introduce Embed-KCPD, a training-free method that represents text units with pretrained sentence embeddings and estimates boundaries using penalized KCPD. To complement the theory empirically, we introduce an LLM-based simulation framework that generates synthetic documents with known boundaries and controlled short-range dependence. Across standard text segmentation benchmarks, Embed-KCPD is competitive with and often outperforms strong unsupervised baselines. We further show that its segments are useful beyond intrinsic boundary-based evaluation, serving as effective chunks for evidence retrieval in retrieval-augmented generation (RAG) systems.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Kernel change-point detection, unsupervised text segmentation, m-dependence, sentence embeddings, retrieval-augmented generation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_MQjxOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies unsupervised text segmentation. The method proposed is straightforward: 1) use a pretrained embedding model to convert each sentence into a vector, and 2) applies KCPD algorithm to identify positions where the distribution of these vectors changes, thereby achieving unsupervised segmentation.

The paper also considers potential dependencies between neighboring sentences and provides two theoretical results under the assumption of m-dependence. The experiments are mainly conducted on several text segmentation datasets and RAG retrieval tasks, where the proposed method generally outperforms the baselines.

Critical strengths

None.

Critical weaknesses
  1. My first concern is why we should use this method in practice. There are already many simple and intuitive approaches, such as fixed-length chunking or directly splitting text based on the embedding similarity between adjacent sentences. However, the authors do not sufficiently demonstrate the advantages of their method over these simpler alternatives. Moreover, the actual improvements seem relatively small, while the proposed method requires substantially more computation. Therefore, I am not convinced that the additional complexity is justified.

  2. Another concern is that the methodological novelty is relatively limited, as the paper does not introduce a new algorithm. As for the theoretical results, I think the m-dependence assumption is quite strong, which may limit their practical significance. In real language, dependencies can exist even between sentences that are far apart. Furthermore, when the authors use an LLM to generate text, even if they only provide the previous M sentences as context, this does not necessarily guarantee that the generated data actually satisfies the m-dependence assumption.

Questions for authors
  1. Why should we use EMBED-KCPD instead of simpler chunking methods?
  2. Is Theorem 3.11 actually non-vacuous?
  3. How does the theory apply to the actual experiments?
  4. Is the RAG comparison fair?
Reviewer_ZzWHOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies penalised kernel change-point detection under mm-dependence and introduces Embed-KCPD, a training-free text segmentation method using pretrained sentence embeddings. It presents concentration, oracle, and localization results, alongside segmentation benchmarks and retrieval experiments. The application is useful and the empirical results are promising, but the verified theoretical issues substantially weaken the central contribution.

Critical strengths
  • The uniform concentration analysis explicitly accounts for finite-range dependence and provides a meaningful foundation for studying dependent observations.
  • The pipeline accommodates multiple encoders without segmentation-specific training. Table 1 reports strong results on Wiki-300 and arXiv, and the embedding-based TextTiling comparisons help assess gains beyond representation quality. The retrieval evaluation extends the investigation beyond intrinsic boundary metrics
Critical weaknesses
  • Assumption 3.8 includes [s,e]=[τk,τk+1][s,e]=[\tau_k,\tau_{k+1}]. Following the paper, this interval contains one observation from PkP_k followed by gkg_k observations from Pk+1P_{k+1}. Boundedness and mm-dependence imply that every admissible population split gain should satisfy Gt≤2MδT+2(2m+1)M.G_t\le \frac{2M}{\delta_T}+2(2m+1)M. However, Assumptions 3.4–3.8 require
\ge3c_0C_0\beta_T-C_m\longrightarrow\infty.$$ These bounds contradict each other for fixed $m,M,C_m$. Assumption 3.8 enters Lemma A.3 directly, supports Lemma A.4, and is subsequently used in Theorem 3.12’s proof. Thus, Theorem 3.12 is vacuous for $K\ge2$ under the stated asymptotic assumptions. * Under Assumption 3.2, empirical and population fit costs for any partition lie in $[0,M(T-1)]$. Because the same penalty appears in both criteria, empirical optimality already gives the deterministic guarantee $$L^\star(\widehat\tau)-\inf_\tau L^\star(\tau)\le2M(T-1).$$ Thus, it seems that no dependence assumption or restriction on the penalty’s sign is needed. Theorem 3.11 instead provides $2\lambda_TT$, which exceeds this deterministic bound for $T\ge3$ and $M>0$. The proof is algebraically sound, but discards potentially useful segment-count dependence in lines 959–971. Consequently, the interpretation that the estimator performs “almost as well” as the population oracle is unsupported by the displayed bound.
Questions for authors
  • With the stated common-window definitions, WindowDiff should be at least PkP_k, yet Table 1 reports 52.9/42.352.9/42.3 for GraphSeg on Elements and 47.9/40.147.9/40.1 for TextTiling on arXiv. Separately, Table 2 describes Random-matched as having the same chunk count as KCPD Base, whereas Tables 6–7 report different means: 46.446.4 versus 54.754.7on QASPER and 190.3190.3 versus 87.387.3 on NQ. Can the authors explain these inconsistencies?
Reviewer_xVPvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

For unsupervised text segmentation, the paper proposes Embed-kcpd. It combines embedding of texts with the idea of kcpd. It also analyze the theoretical quality of the proposed embed-kcpd and also includes a synthetic experiment. On existing text segmentation datasets, Embed-kcpd outperforms other baselines. The paper also performs RAG experiments to show that Embed-kcpd can help chunking for RAG.

Critical strengths

(1) The theoretical analysis of Embed-kcpd and synthetic data analysis seems interesting.

Critical weaknesses

(1) The method proposed in this paper is not novel enough. The implementation of this paper is based on rupture and it already frames text segmentation as a kcpd problem and using cosine similarity as the kernel and word-count representation. Replacing the word-count representation with a sentence and a different kernel seems not technically novel enough for a conference paper. Besides, it would also be great to include the original implementation with the word-count representation and cosine similartiy kernel as the baseline.

Questions for authors

(1) The bolding in Table 1 seems confusing. Can the author provide more explanations?

(2) I am wondering whether the sentence representation is generated on its own or in the context of whole documents. Since the meaning of each sentence can depend on the whole context.

(3) The random-matched baseline should have the same number of chunks as base, why there is different number of chunks in Table 6 and Table 7?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.