When Lexical Change Misleads: Rethinking Dynamic Topic Model Evaluation with Traditional and LLM-Based Metrics

# When Lexical Change Misleads: Rethinking Dynamic Topic Model Evaluation with Traditional and LLM-Based Metrics
Source: [https://arxiv.org/html/2608.13835](https://arxiv.org/html/2608.13835)
Charu Karakkaparambil JamesAffiliation:RPTU University Kaiserslautern\-LandauAffiliation:Kaiserslautern, GermanyEmail:[surname@cs\.uni\-kl\.de](mailto:)

###### Abstract

Dynamic topic models capture evolving word distributions, but traditional coherence metrics may fail when vocabulary changes while semantic meaning persists\. We evaluate 120 topics from CoNTM and DLDA across NYT, DBLP, and arXiv, using three human annotators and Low, Medium, and High lexical\-change categories\. Traditional temporal coherence shows highly variable agreement with human judgments \(ρ\\rho=\-0\.256 to 0\.614\)\. In contrast, LLM\-based semantic similarity agrees strongly with human semantic judgments for CoNTM on NYT \(ρ\\rho=0\.609\), DBLP \(ρ\\rho=0\.721\), and arXiv \(ρ\\rho=0\.502\), but is less consistent for DLDA\. Lexical\-change stratification reveals variation hidden by aggregate evaluation\. We therefore advocate lexical\-change\-aware evaluation, jointly reporting traditional coherence and LLM\-based semantic measures as complementary rather than interchangeable signals\.

## 1Introduction

Dynamic topic models \(DTMs\) extend conventional topic modeling by allowing latent topics to evolve across time[14](https://arxiv.org/html/2608.13835#bib.bib9);[6](https://arxiv.org/html/2608.13835#bib.bib10);[16](https://arxiv.org/html/2608.13835#bib.bib11);[7](https://arxiv.org/html/2608.13835#bib.bib12)\. Instead of representing a topic using one static set of words, a dynamic topic may contain substantially different vocabulary at different time points while still describing a continuous underlying phenomenon\. The original Dynamic Topic Model explicitly motivated this setting by noting that a scientific subject can look lexically different across decades even when the underlying intellectual trajectory remains connected[2](https://arxiv.org/html/2608.13835#bib.bib1)\.

Evaluating such trajectories remains difficult\. Topic\-model evaluation has traditionally relied on automatic coherence measures based on statistical relationships between topic words\. Human\-centered work has repeatedly shown, however, that likelihood or automated coherence does not always correspond to human interpretability[3](https://arxiv.org/html/2608.13835#bib.bib2);[11](https://arxiv.org/html/2608.13835#bib.bib3)\. More recent analyses have questioned whether coherence measures developed for classical topic models remain valid for newer models and across different application settings[4](https://arxiv.org/html/2608.13835#bib.bib4);[5](https://arxiv.org/html/2608.13835#bib.bib5)\.

The problem is amplified in temporal topic modeling\. Consider a topic whose early vocabulary concerns desktop, computer, software, and internet, while later years emphasize smartphone, app, cloud, and platform\. Lexically, the topic has changed considerably\. Semantically, however, a human reader may recognize a coherent evolution of digital technology\. A metric dominated by lexical co\-occurrence can therefore answer a different question from the one a human asks when interpreting a temporal narrative\.

Large language models provide a possible complementary signal because they can reason over the semantic relationship between different surface realizations of a topic\.[15](https://arxiv.org/html/2608.13835#bib.bib6), for example, showed that LLM judgments can correlate strongly with human assessments in static topic\-model evaluation, although the effectiveness depends on the evaluation task\. We extend this motivation to temporal topic evolution\.

We study two models, CoNTM[8](https://arxiv.org/html/2608.13835#bib.bib7)and DLDA[2](https://arxiv.org/html/2608.13835#bib.bib1), on three temporally structured corpora\. Our main finding is that metric validity is conditional\. Traditional coherence correlates strongly with human coherence in some conditions but weakly or even negatively in others\. Similarly, LLM semantic evaluation performs strongly for CoNTM but is inconsistent for DLDA\. Lexical\-change stratification reveals patterns that aggregate evaluation obscures\.

![Refer to caption](https://arxiv.org/html/2608.13835v1/Lexical_Change.png)Figure 1:Relationship between lexical temporal change and traditional NPMI coherence for CoNTM and DLDA across NYT, arXiv, and DBLP\. Panels \(a–b\) show a negative association between lexical change and coherence across datasets\. Panels \(c–d\) illustrate that topics with relatively low coherence can nevertheless exhibit meaningful and interpretable semantic evolution, motivating evaluation beyond lexical coherence alone\.Our main contributions are:

- •We analyzetraditional temporal coherenceacross two dynamic topic models and three datasets, showing substantial variation in agreement with human judgments\.
- •We introducelexical\-change\-aware evaluationby grouping topics into Low, Medium, and High lexical\-change regimes, revealing patterns hidden by aggregate correlations\.
- •We evaluateLLM\-based semantic similarityagainst human judgments, showing that it provides a complementary semantic signal but is not uniformly reliable\.
- •We advocate amulti\-view evaluation frameworkcombining traditional coherence, LLM\-based semantic similarity, and lexical change\.

This paper is organized as follows\. Section[2](https://arxiv.org/html/2608.13835#S2)reviews related work, Section[3](https://arxiv.org/html/2608.13835#S3)describes the experimental setup, Section[4](https://arxiv.org/html/2608.13835#S4)presents the results, Section[5](https://arxiv.org/html/2608.13835#S5)discusses the findings, and Section[6](https://arxiv.org/html/2608.13835#S6)concludes the paper\.

## 2Related Work

This section reviews prior work on traditional topic\-model evaluation and recent LLM\-based approaches to semantic evaluation, with a focus on their relevance to dynamic topic models\.

#### Topic model evaluation

Traditional topic\-model evaluation commonly relies on coherence measures based on statistical associations among topic words\. However, automated metrics do not always align with human interpretability[3](https://arxiv.org/html/2608.13835#bib.bib2);[11](https://arxiv.org/html/2608.13835#bib.bib3)\. More recent studies show that this mismatch can persist across neural topic models and different domains, questioning the use of coherence as a universal proxy for human judgment[4](https://arxiv.org/html/2608.13835#bib.bib4);[5](https://arxiv.org/html/2608.13835#bib.bib5)\. This issue is especially relevant for dynamic topic models, where vocabulary can change substantially over time while the underlying semantic theme remains coherent\.

#### LLM\-based evaluation

Recent work has explored LLMs as semantic evaluators of topic\-model outputs\.[15](https://arxiv.org/html/2608.13835#bib.bib6)show that LLM judgments can correlate strongly with human topic evaluations, although performance depends on the evaluation formulation\. Existing work primarily considers static topic representations[9](https://arxiv.org/html/2608.13835#bib.bib8)\. In contrast, we examine temporal topic evolution and ask whether the reliability of traditional and LLM\-based metrics changes with Low, Medium, and High lexical change\. Rather than replacing traditional coherence, we study traditional coherence and LLM\-based semantic similarity as complementary signals for evaluating dynamic topics\.

## 3Experimental Setup

This section describes the models and datasets, human evaluation procedure, traditional and LLM\-based metrics, and lexical\-change analysis used in our experiments\.

### 3\.1Models and datasets

We evaluate CoNTM and DLDA on three temporally structured datasets: the New York Times \(NYT\), DBLP, and arXiv\. For each model–dataset combination, we select 20 topics, resulting in 120 temporal topic trajectories in total\. Each topic is represented at three time points: 1987, 1997, and 2007 for NYT[13](https://arxiv.org/html/2608.13835#bib.bib13); 2010, 2015, and 2020 for DBLP[10](https://arxiv.org/html/2608.13835#bib.bib15); and 2012, 2018, and 2024 for arXiv[1](https://arxiv.org/html/2608.13835#bib.bib14)\.

### 3\.2Human evaluation

Three annotators evaluate every topic\. For consistency, we use three responses for all model–dataset combinations; the DLDA–NYT file contains five responses, from which the first three are used in the present experiment\.

Annotators answer five questions on a five\-point scale:

- •Temporal Topic Coherence: Are the topic words coherent and interpretable when consecutive time points are considered?
- •Temporal Topic Smoothness: Does the topic evolve smoothly between the three temporal points?
- •Beginning Theme Accuracy: Does the generated beginning theme accurately represent the initial topic words?
- •Ending Theme Accuracy: Does the generated ending theme accurately represent the final topic words?
- •Semantic Similarity: How semantically similar are the beginning and ending themes?

This yields \(120×\\times3×\\times5=1,800\) scalar human judgments\. We average the three annotators for each topic\.

The present analysis focuses primarily on Temporal Topic Coherence and Semantic Similarity because these correspond directly to the traditional and LLM\-based metrics studied below\.

### 3\.3Traditional and LLM\-Based Evaluation

For traditional evaluation, we use the automatically computedTemporal Topic Coherence[9](https://arxiv.org/html/2608.13835#bib.bib8), which summarizes topic coherence across the temporal trajectory\. We compare this score with the mean humanTemporal Topic Coherence\(TTC\) rating obtained from the three annotators\.

For semantic evaluation, we use GPT\-5\.5[12](https://arxiv.org/html/2608.13835#bib.bib16)through the OpenAI Responses API\. For each temporal topic pair, the model receives the dataset, dynamic topic model, topic identifier, timestamps, and corresponding top topic words\. Using a fixed prompt, GPT\-5\.5 generates concise semantic descriptions of the two topic states and evaluates their semantic similarity on a five\-point scale, ranging from 1 \(semantically unrelated\) to 5 \(semantically equivalent\)\. The prompt explicitly instructs the model to distinguish genuine semantic change from lexical replacement and not to penalize vocabulary change when the underlying topic remains conceptually similar\. The model returns a structured JSON response containing the two semantic themes, a similarity score, and a justification\. We compare this score with the mean humanSemantic Similarity\(SS\) rating\.

Importantly, the traditional and LLM\-based metrics are evaluated against different human judgments\. Specifically, we examine

Ctrad\.↔HTTC,C\_\{\\mathrm\{trad\.\}\}\\leftrightarrow H\_\{\\mathrm\{TTC\}\},\(1\)
and

SLLM↔HSS,S\_\{\\mathrm\{LLM\}\}\\leftrightarrow H\_\{\\mathrm\{SS\}\},\(2\)
whereCtrad\.C\_\{\\mathrm\{trad\.\}\}denotes temporal topic coherence,SLLMS\_\{\\mathrm\{LLM\}\}denotes LLM\-based semantic similarity,HTTCH\_\{\\mathrm\{TTC\}\}denotes the mean human temporal topic coherence judgment, andHSSH\_\{\\mathrm\{SS\}\}denotes the mean human semantic similarity judgment\. Therefore, we do not interpret the two correlation coefficients as a direct comparison between competing metrics\. Instead, we assess how well each metric agrees with human judgments of the corresponding property of temporal topic quality\.

### 3\.4Lexical\-change stratification

Each topic has a Topic Lexical Temporal Change value and an associated Low, Medium, or High lexical\-change category\. Within each model–dataset combination, the 20 topics consist of seven Low, seven Medium, and six High lexical\-change topics\.

This stratification allows us to test whether aggregate metric–human correlations conceal different behavior under different levels of vocabulary evolution\.

Because human responses are ordinal Likert judgments, Spearman’s rank correlation \(ρ\\rho\) is used as the primary agreement statistic\. Category\-level results should be interpreted cautiously because each category contains only six or seven topics per dataset\.

## 4Results

This section presents the overall and lexical change analyzed correlations between automatic metrics and their corresponding human judgments, highlighting how metric reliability varies across models and datasets\.

### 4\.1Traditional coherence does not have uniform human validity

Table[1](https://arxiv.org/html/2608.13835#S4.T1)reports the relationship between traditional temporal coherence and human Temporal Topic Coherence, together with the relationship between the LLM score and human Semantic Similarity\.

Table 1:Spearman’sρ\\rhobetween automatic metrics and corresponding three\-annotator human judgments\.Ctrad\.C\_\{\\mathrm\{trad\.\}\}denotes traditional temporal coherence,HTTCH\_\{\\mathrm\{TTC\}\}denotes human Temporal Topic Coherence, andHSSH\_\{\\mathrm\{SS\}\}denotes human Semantic Similarity\. Bold values indicate statistically significant correlations \(p<0\.05p<0\.05\)\.Traditional temporal coherence behaves very differently across experimental conditions\. For DLDA–DBLP, it agrees strongly with humans \(ρ=0\.614,p=\.004\\rho=0\.614,p=\.004\); DLDA–NYT also exhibits significant positive agreement \(ρ=0\.531,p=\.016\\rho=0\.531,p=\.016\)\. CoNTM–arXiv produces moderate agreement \(ρ=0\.500,p=\.025\\rho=0\.500,p=\.025\)\.

However, these results do not generalize across datasets\. For DLDA–arXiv, correlation is essentially zero \(ρ=0\.007\\rho=0\.007\)\. More strikingly, CoNTM–NYT shows a negative overall correlation \(ρ=−0\.256\\rho=\-0\.256\), although it is not statistically significant\.

Thus, a high traditional coherence score cannot be assumed to have the same relationship with human temporal interpretability across models and corpora\. This supports prior concerns that automated topic coherence is not universally interchangeable with human judgment[4](https://arxiv.org/html/2608.13835#bib.bib4);[5](https://arxiv.org/html/2608.13835#bib.bib5)\.

### 4\.2Lexical change reveals hidden metric behavior

The aggregate results hide substantial differences among lexical\-change categories\. Particularly notable examples include:

Table 2:Selected lexical\-change\-stratified Spearman correlations \(ρ\\rho\)\. Blue rows represent traditional coherence versus human Temporal Topic Coherence \(TTC\), while orange rows represent LLM semantic similarity versus human Semantic Similarity \(SS\)\. Green cells highlight the statistically significant correlations \(p<0\.05p<0\.05\)\. Category sizes areN=7N=7,77, and66for Low, Medium, and High lexical change, respectively\.The strongest example of metric instability occurs for CoNTM–NYT\. Under Low lexical change, traditional coherence is strongly negatively correlated with human coherence \(ρ=−0\.786,p=\.036\\rho=\-0\.786,p=\.036\), whereas the LLM semantic similarity score strongly agrees with human semantic judgments in the same lexical regime \(ρ=0\.801,p=\.030\\rho=0\.801,p=\.030\)\.

DLDA shows a different pattern\. Traditional coherence is particularly informative for Medium lexical change, reachingρ=0\.873\\rho=0\.873on NYT andρ=0\.898\\rho=0\.898on DBLP\. This observation is important because it prevents an overly simple conclusion that increasing lexical change necessarily invalidates traditional coherence\. Instead, the effect depends on the interaction among lexical change, model behavior, and corpus characteristics\.

### 4\.3LLM semantic evaluation is useful but not universal

Across all 120 topics, human Semantic Similarity and the LLM semantic score have a pooled correlation of

\[ρ=0\.479,p<\.001\.\]\[\\rho=0\.479,\\qquad p<\.001\.\]
The pooled correlation is strongest for Low lexical\-change topics \(ρ=0\.556,p<\.001\\rho=0\.556,p<\.001\), remains significant for Medium lexical change \(ρ=0\.405,p=\.008\\rho=0\.405,p=\.008\), and becomes weaker for High lexical change \(ρ=0\.294,p=\.082\\rho=0\.294,p=\.082\)\.

Model\-specific analysis reveals an important asymmetry\. Across the three datasets, CoNTM has strong LLM–human semantic agreement:

\[ρNYT=0\.609,ρDBLP=0\.721,ρarXiv=0\.502\.\]\[\\rho\_\{\\text\{NYT\}\}=0\.609,\\rho\_\{\\text\{DBLP\}\}=0\.721,\\\\ \\rho\_\{\\text\{arXiv\}\}=0\.502\.\]
All three are statistically significant\.

DLDA is substantially less consistent\. NYT shows moderate agreement \(ρ=0\.445,p=\.049\\rho=0\.445,p=\.049\), but DBLP \(ρ=−0\.086\\rho=\-0\.086\) and arXiv \(ρ=−0\.190\\rho=\-0\.190\) do not\.

These results argue against replacing traditional metrics with an LLM score\. Instead, they suggest that LLM judgments form a second evaluation view whose reliability must itself be validated\. This is consistent with existing LLM\-based topic\-model evaluation work showing that performance depends strongly on the evaluation formulation\.

## 5Discussion: When Does Lexical Change Mislead?

Our results \(see Table[2](https://arxiv.org/html/2608.13835#S4.T2)\) suggest that lexical change should be treated as an evaluation context variable, rather than merely as another topic\-quality score\.

Traditional coherence asks whether topic words exhibit statistical or lexical compatibility\. This is useful when lexical association is closely aligned with the semantic organization perceived by humans\. DLDA–NYT and DLDA–DBLP demonstrate that this alignment can be very strong\.

Semantic continuity asks a different question: After the words change, is the topic still about the same underlying theme?

An LLM can compare descriptions such as a Beginning Theme and Ending Theme without requiring the same words to persist\. This capability is particularly valuable when semantic relationships are not reducible to direct lexical overlap\.

However, the results also show that lexical evolution is not equivalent to semantic evolution\. High lexical change does not automatically imply that a traditional metric fails, nor does it guarantee that an LLM succeeds\. Indeed, the LLM–human relationship weakens for several High\-change conditions\. This may occur because highly changed trajectories are genuinely ambiguous: both humans and LLMs must determine whether the trajectory represents meaningful evolution, topic drift, or a transition between unrelated concepts\.

We therefore propose evaluating a dynamic topic using three complementary dimensions:

\[E⁡\(T\)=Ctraditional,SLLM,Lchange\]\[\\boxed\{E\(T\)=\{C\_\{\\text\{traditional\}\},S\_\{\\text\{LLM\}\},L\_\{\\text\{change\}\}\}\}\]
where \(CtraditionalC\_\{\\text\{traditional\}\}\) measures lexical/statistical coherence, \(SLLMS\_\{\\text\{LLM\}\}\) measures semantic continuity, and \(LchangeL\_\{\\text\{change\}\}\) describes the amount of vocabulary evolution\.

Crucially, we do not collapse these values into a single weighted score in the present work\. The human experiments show that their behavior changes across models and lexical regimes; choosing fixed weights without further validation could simply introduce another opaque evaluation metric\.

Instead, we recommend reporting them jointly\. For example, a topic with low traditional coherence but high semantic similarity and high lexical change should be inspected differently from a topic that scores poorly on both coherence and semantic similarity\. The first may represent legitimate semantic evolution; the second is more likely to represent an unstable or incoherent topic trajectory\.

This interpretation leads to the paper’s main conclusion that Lexical change does not make traditional evaluation useless; it makes reliance on a single evaluation unsafe\.

## 6Conclusion

We investigated the relationship among lexical change, traditional coherence, LLM\-based semantic similarity, and human judgments in dynamic topic\-model evaluation\. Across CoNTM and DLDA on NYT, DBLP, and arXiv, traditional coherence exhibits substantial variation in human agreement, including strong positive, near\-zero, and negative correlations\. LLM\-based semantic similarity provides strong complementary evidence for several CoNTM conditions but is itself model dependent\.

These findings suggest that dynamic topic evaluation should move beyond the search for one universal automatic metric\. A temporal topic is simultaneously a lexical object and a semantic trajectory\. Traditional coherence captures whether its words fit together; LLM\-based evaluation captures whether its meaning persists or evolves coherently\. Lexical change tells us when the distinction matters\.

We therefore advocate lexical\-change\-aware, multi\-view evaluation in which traditional and LLM\-based metrics are jointly reported and validated against their corresponding human judgments\.

## Limitations

Our study evaluates 120 topics from two models and three English\-language datasets\. Category\-level experiments contain only six or seven topics, making individual correlations sensitive to outliers\. The LLM analysis uses one LLM configuration, and LLM judgments may vary with prompting, model family, and generated semantic themes\. In addition, traditional coherence and LLM semantic similarity are validated against different human questions, so their correlation coefficients should not be interpreted as direct metric rankings\.

Future work should increase the number of topics and annotators, evaluate additional temporal topic models and LLMs, test prompt robustness, and investigate whether lexical\-change\-aware combinations predict holistic human judgments better than any individual metric\.

## References

- arXiv\.org submittersArXiv dataset\.Kaggle\.External Links:[Link](https://www.kaggle.com/dsv/4852963),[Document](https://dx.doi.org/10.34740/KAGGLE/DSV/4852963)Cited by:[§3\.1](https://arxiv.org/html/2608.13835#S3.SS1.p1.1)\.
- Blei and Lafferty \(2006\)D\. M\. Blei and J\. D\. LaffertyDynamic topic models\.InProceedings of the 23rd International Conference on Machine Learning,ICML ’06,New York, NY, USA,pp\. 113–120\.External Links:ISBN 1595933832,[Link](https://doi.org/10.1145/1143844.1143859),[Document](https://dx.doi.org/10.1145/1143844.1143859)Cited by:[§1](https://arxiv.org/html/2608.13835#S1.p1.1),[§1](https://arxiv.org/html/2608.13835#S1.p5.1)\.
- Changet al\.\(2009\)J\. Chang, J\. Boyd\-Graber, S\. Gerrish, C\. Wang, and D\. M\. BleiReading tea leaves: how humans interpret topic models\.InProceedings of the 23rd International Conference on Neural Information Processing Systems,NIPS’09,Red Hook, NY, USA,pp\. 288–296\.External Links:ISBN 9781615679119Cited by:[§1](https://arxiv.org/html/2608.13835#S1.p2.1),[§2](https://arxiv.org/html/2608.13835#S2.SS0.SSS0.Px1.p1.1)\.
- Doogan and Buntine \(2021\)C\. Doogan and W\. BuntineTopic model or topic twaddle? re\-evaluating semantic interpretability measures\.InProceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,K\. Toutanova, A\. Rumshisky, L\. Zettlemoyer, D\. Hakkani\-Tur, I\. Beltagy, S\. Bethard, R\. Cotterell, T\. Chakraborty, and Y\. Zhou \(Eds\.\),Online,pp\. 3824–3848\.External Links:[Link](https://aclanthology.org/2021.naacl-main.300/),[Document](https://dx.doi.org/10.18653/v1/2021.naacl-main.300)Cited by:[§1](https://arxiv.org/html/2608.13835#S1.p2.1),[§2](https://arxiv.org/html/2608.13835#S2.SS0.SSS0.Px1.p1.1),[§4\.1](https://arxiv.org/html/2608.13835#S4.SS1.p4.1)\.
- Hoyleet al\.\(2021\)A\. Hoyle, P\. Goel, D\. Peskov, A\. Hian\-Cheong, J\. Boyd\-Graber, and P\. ResnikIs automated topic model evaluation broken? the incoherence of coherence\.InProceedings of the 35th International Conference on Neural Information Processing Systems,NIPS ’21,Red Hook, NY, USA\.External Links:ISBN 9781713845393Cited by:[§1](https://arxiv.org/html/2608.13835#S1.p2.1),[§2](https://arxiv.org/html/2608.13835#S2.SS0.SSS0.Px1.p1.1),[§4\.1](https://arxiv.org/html/2608.13835#S4.SS1.p4.1)\.
- Iwataet al\.\(2009\)T\. Iwata, S\. Watanabe, T\. Yamada, and N\. UedaTopic tracking model for analyzing consumer purchase behavior\.InTwenty\-First international joint conference on artificial intelligence,Cited by:[§1](https://arxiv.org/html/2608.13835#S1.p1.1)\.
- Karakkaparambil Jameset al\.\(2026a\)C\. Karakkaparambil James, M\. Monteiro, and S\. FellenzSemantic echo pathways \(SEP\): tracing how medical language propagates and transforms\.InProceedings of the 1st Workshop on Linguistic Analysis for Health \(HeaLing 2026\),V\. Danilova, M\. Kurfalı, Y\. Söderfeldt, J\. Reed, and A\. Burchell \(Eds\.\),Rabat, Morocco,pp\. 55–66\.External Links:[Link](https://aclanthology.org/2026.healing-1.5/),[Document](https://dx.doi.org/10.18653/v1/2026.healing-1.5),ISBN 979\-8\-89176\-367\-8Cited by:[§1](https://arxiv.org/html/2608.13835#S1.p1.1)\.
- Karakkaparambil Jameset al\.\(2026b\)C\. Karakkaparambil James, W\. Mustafa, M\. Monteiro, M\. Kloft, and S\. FellenzContinual neural topic model\.InProceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics \(Volume 1: Long Papers\),V\. Demberg, K\. Inui, and L\. Marquez \(Eds\.\),Rabat, Morocco,pp\. 6636–6658\.External Links:[Link](https://aclanthology.org/2026.eacl-long.312/),[Document](https://dx.doi.org/10.18653/v1/2026.eacl-long.312),ISBN 979\-8\-89176\-380\-7Cited by:[§1](https://arxiv.org/html/2608.13835#S1.p5.1)\.
- Karakkaparambil Jameset al\.\(2024\)C\. Karakkaparambil James, M\. Nagda, N\. Haji Ghassemi, M\. Kloft, and S\. FellenzEvaluating dynamic topic models\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Bangkok, Thailand,pp\. 160–176\.External Links:[Link](https://aclanthology.org/2024.acl-long.11/),[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.11)Cited by:[§2](https://arxiv.org/html/2608.13835#S2.SS0.SSS0.Px2.p1.1),[§3\.3](https://arxiv.org/html/2608.13835#S3.SS3.p1.1)\.
- Ley \(2002\)M\. LeyThe dblp computer science bibliography: evolution, research issues, perspectives\.InInternational symposium on string processing and information retrieval,pp\. 1–10\.Cited by:[§3\.1](https://arxiv.org/html/2608.13835#S3.SS1.p1.1)\.
- Newmanet al\.\(2010\)D\. Newman, J\. H\. Lau, K\. Grieser, and T\. BaldwinAutomatic evaluation of topic coherence\.InHuman Language Technologies: The 2010 Annual Conference of the North American Chapter of the Association for Computational Linguistics,HLT ’10,USA,pp\. 100–108\.External Links:ISBN 1932432655Cited by:[§1](https://arxiv.org/html/2608.13835#S1.p2.1),[§2](https://arxiv.org/html/2608.13835#S2.SS0.SSS0.Px1.p1.1)\.
- OpenAI \(2026\)OpenAIIntroducing GPT\-5\.5\.Note:OpenAICited by:[§3\.3](https://arxiv.org/html/2608.13835#S3.SS3.p2.1)\.
- Sandhaus \(2008\)E\. SandhausThe New York Times Annotated Corpus\.Abacus Data Network\.External Links:[Document](https://dx.doi.org/11272.1/AB2/GZC6PL),[Link](https://hdl.handle.net/11272.1/AB2/GZC6PL)Cited by:[§3\.1](https://arxiv.org/html/2608.13835#S3.SS1.p1.1)\.
- Sasakiet al\.\(2014\)K\. Sasaki, T\. Yoshikawa, and T\. FuruhashiOnline topic model for twitter considering dynamics of user interests and topic trends\.InProceedings of the 2014 conference on empirical methods in natural language processing \(EMNLP\),pp\. 1977–1985\.Cited by:[§1](https://arxiv.org/html/2608.13835#S1.p1.1)\.
- Stammbachet al\.\(2023\)D\. Stammbach, V\. Zouhar, A\. Hoyle, M\. Sachan, and E\. AshRevisiting automated topic model evaluation with large language models\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,H\. Bouamor, J\. Pino, and K\. Bali \(Eds\.\),Singapore,pp\. 9348–9357\.External Links:[Link](https://aclanthology.org/2023.emnlp-main.581/),[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.581)Cited by:[§1](https://arxiv.org/html/2608.13835#S1.p4.1),[§2](https://arxiv.org/html/2608.13835#S2.SS0.SSS0.Px2.p1.1)\.
- Yao and Wang \(2020\)F\. Yao and Y\. WangTracking urban geo\-topics based on dynamic topic model\.Computers, Environment and Urban Systems79,pp\. 101419\.Cited by:[§1](https://arxiv.org/html/2608.13835#S1.p1.1)\.