Stylistic Evolution and LLM Neutrality in Singlish Language

Linus Tze En Foo; Lynnette Hui Xian Ng; Weihan Angela Ng; Wenkai Li

REVIEW 5 major objections 6 minor 7 references

Stylistic Evolution and LLM Neutrality in Singlish Language

T0 review · 5 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that Singlish changed measurably from 2012 to 2021 and that no large language model tested can produce Singlish that is both authentic and free of a specific time period's style.

desk verdict Diachronic Singlish result is intriguing but rests on unexamined corpus-stability assumptions; the LLM trade-off claim is not supported by the reported statistics as written. read the letter →

arxiv 2601.06580 v2 pith:PLJEL4PP submitted 2026-01-10 cs.CL

Linus Tze En Foo , Weihan Angela Ng , Wenkai Li , Lynnette Hui Xian Ng This is my paper · ORCID

classification cs.CL

keywords SinglishcreolediachroniclanguagechangestylisticsimilaritytemporalneutralitylargemodelsLLMevaluationsociolect

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that Singlish, the English-based creole of Singapore, changed in measurable, cumulative ways between 2012 and 2021, and that the large language models it tested cannot generate Singlish that is both convincingly real and free of a particular year's fingerprint. The evidence is a pairwise classifier that distinguishes messages from different years; similarity drops as the year gap grows, and feature attribution ties the drop mainly to message length and complexity. The same classifier, used between each model's output and each year's real messages, shows that all models carry year-specific bias; fine-tuning and prompting shift realism but do not remove the bias, producing a trade-off between authenticity and temporal neutrality. A sympathetic reader would care because this gives a concrete, quantitative diagnostic for whether a generative model has absorbed a sociolect's stable core, and it suggests current models have not.

What carries the argument

The key instrument is a pairwise text classifier that is asked to tell which of two years a message came from; its accuracy is rescaled to a similarity score S = 1 - 2|acc - 0.5|. Running this classifier for all year pairs yields a curve of separability versus time gap, and running it between each model's outputs and each year's real messages yields a per-model profile whose standard deviation is the paper's 'temporal neutrality'. Feature-attribution on the handcrafted features shows message-length features carrying most of the temporal signal.

What would settle it

Rerun the year-pair classification after truncating all messages to a common maximum length and balancing the number of messages per year; if the separability curve becomes flat, the central claim of cumulative stylistic change is an artifact of message-length drift rather than language change.

Watch

Extended reading notes

Core claim

The paper's central claim is that the stylistic distance between Singlish messages grows monotonically with the time gap between them, that this drift is driven mostly by text length and complexity rather than by iconic Singlish particles, and that every large language model examined shows detectable year-specific bias in the Singlish it produces, so none achieves both authenticity and temporal neutrality at once. The authors propose using temporal neutrality as a diagnostic metric for assessing how well a generative model has grounded itself in a sociolect.

Load-bearing premise

The whole diachronic trend rests on the assumption that the messages from 2012 and 2021 are drawn from the same kind of writers using the same kind of platforms, so that year-to-year classifier separability reflects genuine language change rather than a change in who wrote or how messages were constrained (e.g., SMS character limits).

Editorial extensions

If this is right

If the trend is real, Singlish's evolution is gradual and cumulative, so forensic dating of informal text is feasible at the level of a decade.
Temporal neutrality becomes a practical metric: an LLM that scores high authenticity but high variance across years is not sociolectally grounded, just period-mimetic.
Fine-tuning on aggregated multi-year data can raise realism but does not reliably flatten the year-to-year variance, suggesting the temporal signal lives in the training distribution itself.
The framework extends to other low-resource or creole varieties for measuring whether generative models have captured their stable core.
Length-related features, not particles, are the best predictors for diachronic classification of this kind of data.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A testable extension: for each year, subsample messages to equalize message length distributions, then rerun the separability curve; if the monotonic trend collapses, the 'decreolisation' reading is largely a length artifact.
If LLM outputs carry period fingerprints, then any social-science pipeline that uses LLM-generated text as a proxy for community language should either request a target era or explicitly model the temporal bias.
The same pairwise-separability design could be applied to other time series of informal writing (e.g., SMS vs chat platforms) to see whether the length-driven dominance is specific to this corpus or a general property of digital language change.
Temporal neutrality might be incorporated as a training objective—for instance by fine-tuning on year-balanced data and then optimizing for low variance across years—which this paper's evidence suggests would be hard but not impossible.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The two things you should know about this paper: the diachronic separability result is plausible enough to take seriously, but it depends on an assumption about the CoSEM corpus that the manuscript never checks. And the headline LLM claim—a structural trade-off between authenticity and temporal neutrality—is not established by the reported numbers; the appendix actually contradicts the main text.

What's genuinely new: this is the first computational attempt I've seen to track Singlish style across a decade, and the design is sensible. Pairwise year classification with both encoder embeddings and handcrafted features, then measuring similarity via distance from chance accuracy, is straightforward and interpretable. The two feature sets converge (Spearman rho = 1.00), the trend is monotonic in year-gap, and the SHAP analysis cleanly identifies length-related features as dominant, with particles playing a weaker role. That is a real empirical contribution, and the particle-frequency analysis is more nuanced than most.

The soft spots are concentrated where the claims get strong. The corpus: the paper reports no per-year message counts, no user or platform breakdown, and no discussion of how the collection protocol changed over 2012–2021. The strongest single feature in the diachronic analysis is message length—character length, word length, average word length. If later years happen to contain longer messages because of SMS character limits, migration to chat apps, or a different contributor mix, the entire monotonic trend could be a data artifact rather than language change. The paper needs to rule that out, or at least show the trend holds within comparable message-length strata.

The LLM half is more problematic. Section 5.2 states that variance tests confirmed significant variation across years, but Appendix C's Table 4 reports chi-square p = 1.0 for every model. With p = 1.0, you cannot claim significance. There is also no description of the null distribution used for the test, which is basic. On top of that, generation used a single fixed seed, there is no memorization check for the fine-tuned models, and no code or hyperparameters are released. These are all fixable, but as written the 'structural trade-off' is an assertion, not a demonstrated result.

My overall read: the diachronic half could survive a serious revision, the LLM half likely needs to be redone with proper stats and controls. The paper is worth a referee's time—it's asking the right questions and has a novel empirical foundation—but it should not be accepted without major changes. I'd send it to review with a strong request for corpus metadata, a properly specified null, multi-seed generation, and a memorization check. If those come back clean, this could be a useful paper in computational sociolinguistics. As it stands, cite with caution.

Recommendation: accept for peer review with major revision expected, and flag the corpus-composition issue as load-bearing.

Referee Report

5 major / 6 minor

Summary. The paper introduces a stylistic similarity framework that combines handcrafted linguistic features, LIWC psycholinguistic dimensions, and encoder-derived embeddings to study diachronic change in Singlish using the CoSEM corpus (2012–2021). It reports that classifier separability between year-pairs grows monotonically with temporal gap, with length-related features as the dominant driver. It then evaluates four LLMs under multiple prompting strategies and two LoRA-finetuned variants, measuring authenticity via similarity to CoSEM and temporal neutrality via the standard deviation of similarity across years. The paper claims that most LLMs fail to achieve both authenticity and temporal neutrality, implying a structural trade-off. The central diachronic and LLM-neutrality claims are currently undermined by several methodological and statistical issues, detailed below.

Significance. If the findings were robust, the paper would make a useful contribution to computational sociolinguistics: it applies an interpretable, multi-feature framework to a relatively understudied creole variety, and it raises a novel question about whether LLMs can model temporal neutrality in a sociolect. The paper's strengths include the use of a publicly available corpus, convergent evidence from encoder and handcrafted features, and SHAP-based interpretability. However, the load-bearing evidence for both research questions rests on shaky methodological ground: the corpus-comparability assumption is unverified, the LIWC features appear to leak year information into the classifier, the variance tests contradict the reported significance, and the 'structural trade-off' is not demonstrated by the results as presented. These issues are substantive and require reanalysis rather than copy-editing.

major comments (5)

[§4.1 / Fig. 3 / Table 3] The diachronic claim requires that CoSEM yearly slices are comparable samples of the same Singlish-writing population. The paper reports only that messages span 2012–2021 and that 2022 was dropped; no per-year message counts, user/platform composition, or collection protocol are given. Since the SHAP analysis identifies length_char, length_word, and avg_word_len as the top discriminative features, this is not a minor omission: SMS 160-character limits, migration from SMS to chat apps, or shifts in contributor mix could mechanically lengthen later messages and produce the observed monotonic separability. Please add per-year descriptive statistics and run control analyses—for example, restricting to a stable platform subset, or regressing out length and re-testing separability.
[§3.1 / §4.2] LIWC features are aggregated by year and then included in the message-level classification pipeline. Every message in a given year therefore receives the same LIWC values, so the classifier can use these constant-per-year features to separate years based on aggregate trends, inflating accuracy and contaminating the similarity scores. This affects the handcrafted-feature results and any model comparisons using those features. LIWC should be excluded from the message-level classifier and used only for the descriptive yearly trend analysis (Fig. 2/Table 2), or the analysis should be conducted at the year level with appropriate sample-size corrections.
[§5.2 / Table 4 / Appendix C] There is a direct internal contradiction: the text states that 'variance tests confirmed that similarity scores varied significantly across years for all models,' yet Table 4 reports χ² p = 1.0 for every model, which would normally indicate no significant variation. The null hypothesis, degrees of freedom, and test statistic construction are not described in the methods or appendix. As written, the temporal-neutrality finding is statistically unsupported. Please specify the null distribution, the test used, and correct the interpretation; if p = 1.0 is correct, the claim that temporal signals persist after prompting/fine-tuning must be withdrawn or re-argued with a properly specified test.
[§4.3 / §4.4] The fine-tuned models (Qwen-FT, Mistral-FT) are trained on the full CoSEM corpus and then evaluated for similarity to CoSEM yearly subsets. There is no check for memorization or near-duplicate generation. If the models reproduce training messages, their similarity to CoSEM data would be artificially inflated, especially for Qwen-FT, which shows the highest realism (S=0.673). Please add exact-match or n-gram-overlap analyses between generated texts and CoSEM, or retrain on a time-based held-out split (e.g., train on 2012–2019, evaluate against 2020–2021) to rule out memorization.
[Abstract / §5.2 / §7] The abstract and conclusion claim a 'structural trade-off' between authenticity and temporal neutrality, but Table 1 does not show such a trade-off. For example, Qwen-FT has the highest mean similarity (0.673) and a mid-range std (0.043), while SeaLLM has lower realism (0.417) and lower std (0.033); this is not a monotonic negative relationship. No correlation analysis or statistical test of the trade-off is presented. Either provide direct evidence for the trade-off or soften the claim to say that most models fail to jointly achieve high realism and low temporal variation, without asserting a structural trade-off.

minor comments (6)

[Throughout] The paper uses σ² to denote standard deviation (e.g., §3.3, Table 1), but σ² conventionally denotes variance. Please use 'std' or σ throughout.
[§3.3] The similarity transform S = 1 − 2|acc − 0.5| is fine, but the interpretation 'S=0.5' corresponds to chance-level accuracy (acc=0.75) when the classifier is perfectly symmetric; this should be stated explicitly because readers may misread the scale.
[§A / Table 1] Table 1 has a 'Baseline' column that is not defined in the prompting schemes (Zero-Shot, CoT, Diverse-Decoding, Self-Consistency). Please clarify whether 'Baseline' is the Zero-Shot setting or a separate default generation.
[§5.2 / Appendix C] The text references 'Appendix B' for χ² test analyses, but the variance tests appear in Appendix C. Please fix the cross-reference.
[§5.1] The word 'interoperability' should be 'interpretability'; the sentence in §4.4 'linguistic draft' should be 'linguistic drift'; 'real-word' in §7 should be 'real-world'; 'closed to' in §5.2 should be 'close to'; 'We shown' in the contributions should be 'We show'.
[§4.4] The caption of Figure 5 says similarity scores 'do not have obvious trends' — this is acceptable, but the figure itself would benefit from error bars or confidence intervals, since each similarity score is computed from a single 100-message generated set.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: LLM trade-off is empirical, not definitional; self-citations are non-load-bearing.

full rationale

The derivation chain is not circular. The diachronic measurement uses an external corpus (CoSEM, Gonzales et al. 2023) and reports classifier accuracy between year-pairs; the monotonic decline in similarity with temporal gap is an empirical finding, not a consequence of the similarity definition (S = 1 − 2|acc − 0.5|). The LLM evaluation compares generated texts to yearly CoSEM subsets; realism (mean similarity) and temporal neutrality (standard deviation of the same year-wise similarities) are distinct summaries of the same measurements, and the claimed trade-off is an observed pattern across models, not forced by the metric's definition—a model with uniformly high similarity to all years would score high on both. The only self-citations (Ng and Chan 2024; Ng et al. 2025) appear in related work and are not load-bearing for the paper's central argument; no uniqueness theorem or fitted parameter is imported from them. The main concerns—per-year corpus composition, possible memorization in fine-tuned models, and the uninterpretable χ² p-values in Table 4—are validity/statistical reporting issues rather than circularity, so they do not raise the circularity score.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The paper introduces no new entities; 'temporal neutrality' is a derived evaluation metric (Section 3.4), not an entity with independent evidence. The central claims rest on domain assumptions about corpus representativeness, metric validity, and the transferability of pre-trained embeddings to Singlish, plus several unreported hyperparameter choices.

free parameters (4)

GradientBoosting hyperparameters = sklearn defaults (unreported)
Section 3.2 says the classifier was chosen by pre-testing for predictive power, but actual hyperparameters are not given; accuracy—and hence every similarity score—depends on them.
LoRA fine-tuning hyperparameters = LLaMAFactory defaults (unreported)
Section 4.3 states default hyper-parameters are adopted but does not list rank, alpha, learning rate, epochs; these govern how much temporal signal the fine-tuned models retain.
Decoding parameters for generation = model defaults (unreported)
Section 4.3 uses default generation settings; temperature, top-p, max tokens affect message length, the dominant feature in the analysis.
LIWC aggregation window = 1 year
Section 3.1 aggregates LIWC by year because short messages are unreliable; the choice of a one-year window fixes temporal resolution and shapes the PCA/LIWC trends.

assumptions (6)

domain assumption CoSEM yearly subsets sample the same evolving Singlish-speaking population; year labels are not confounded with platform or author changes.
Invoked in Section 4.1 (dataset) and Section 6.1 (decreolisation interpretation); no per-year metadata or controls are given.
ad hoc to paper Year-pair classifier accuracy is a valid measure of stylistic similarity between text collections.
Section 3.3 maps accuracy to S=1−2|acc−0.5| and treats S as similarity; this operationalization is introduced by the paper and not independently validated.
domain assumption LIWC scores aggregated at the year level give reliable psycholinguistic estimates for short messages.
Section 3.1 cites Zhao et al. (2016) for unreliability on short texts and aggregates by year; this relies on external validity claims and the aggregation window.
domain assumption all-MiniLM-L6-V2 embeddings capture stylistic and temporal properties of short Singlish messages.
Section 3.1 assumes the pre-trained encoder's representations encode subtle stylistic and temporal variation; no Singlish-specific validation is provided.
ad hoc to paper The standard deviation of year-wise similarity scores is an interpretable temporal-neutrality metric.
Section 3.4 defines neutrality as low σ of S across years; no benchmark or validity check links this variance to human judgments of period-free Singlish.
standard math SHAP values provide accurate additive feature attributions for the trained GradientBoosting model.
Section 5.1 uses mean absolute SHAP as feature importance; relies on the SHAP framework's additivity assumptions.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Stylistic Evolution and LLM Neutrality in Singlish Language." pith.science (2026). https://pith.science/paper/PLJEL4PP

@misc{pith2026260106580,
  author       = {Pith},
  title        = {Pith review of: Stylistic Evolution and LLM Neutrality in Singlish Language},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PLJEL4PP}},
  note         = {Machine review of arXiv:2601.06580}
}

read the original abstract

Singlish is a creole rooted in Singapore's multilingual environment that continues to evolve alongside social and technological change. We examine diachronic stylistic change across a decade of informal digital messages and ask whether Large Language Models (LLMs) can generate temporally neutral outputs approximating the stable essence of the variety. Using lexical, pragmatic, psycholinguistic, and encoder-based features, we find that stylistic separability increases with temporal distance, driven primarily by structural features such as length and complexity. Evaluated against a null distribution baseline, most LLMs fail to achieve both authenticity and temporal neutrality simultaneously, revealing a structural trade-off: models generating realistic Singlish inherit its temporal biases, while temporally neutral models produce inauthentic outputs. These findings position temporal neutrality as a diagnostic metric for assessing sociolectal grounding in LLMs.

Figures

Figures reproduced from arXiv: 2601.06580 by the authors.

**Figure 2.** Trend of top LIWC features across categories. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Feature Importance of each of the handcrafted features plotted against year-gap. The feature importance is [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: On CoSEM data, similarity score decreases [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Similarity score of generated texts with CoSEM data vary across the years do not have obvious trends. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: Normalized frequency (per 1000 words) of [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

7 extracted references · 3 linked inside Pith

[5]

InPro- ceedings of the 29th International Conference on Computational Linguistics, pages 3924–3936

Singlish message paraphrasing: A joint task of creole translation and text normalization. InPro- ceedings of the 29th International Conference on Computational Linguistics, pages 3924–3936. Scott M Lundberg and Su-In Lee. 2017. A unified ap- proach to interpreting model predictions.Advances in neural information processing systems, 30. Ho Mian-Lian and Jo...

arXiv 2017
[6]

InProceedings of the International AAAI Conference on Web and Social Media, volume 19, pages 1315–1328

Examining the influence of political bias on large language model performance in stance classi- fication. InProceedings of the International AAAI Conference on Web and Social Media, volume 19, pages 1315–1328. Ri Chi Ng, Nirmalendu Prakash, Ming Shan Hee, Kenny Tsu Wei Choo, and Roy Ka-Wei Lee. 2024. Sghate- check: Functional tests for detecting hate spee...

arXiv 2024
[2018]

Jacob Eisenstein, Brendan O’Connor, Noah A Smith, and Eric P Xing

Singaporean internet chit chat compared to informal spoken language: Linguistic variation and indexicality in a language contact situation.Journal of Pidgin and Creole Languages, 33:48–90. Jacob Eisenstein, Brendan O’Connor, Noah A Smith, and Eric P Xing. 2014. Diffusion of lexical change in social media.PloS one, 9(11):e113114. Wilkinson Daniel Wong Gonz...

2014
[2022]

InInterna- tional Conference on Asian Digital Libraries, pages 115–124

Singlish checker: A tool for understanding and analysing an english creole language. InInterna- tional Conference on Asian Digital Libraries, pages 115–124. Springer. Wesley Lee. 2016. In defense of singlish: A cultural interpretation of singapore english.The Morningside Review, 12. Zhengyuan Liu, Shikang Ni, Aiti Aw, and Nancy Chen

2016
[2023]

Nan Zhao, Dongdong Jiao, Shuotian Bai, and Ting- shao Zhu

Prompting multilingual large language models to generate code-mixed texts: The case of south east asian languages.arXiv preprint arXiv:2303.13592. Nan Zhao, Dongdong Jiao, Shuotian Bai, and Ting- shao Zhu. 2016. Evaluating the validity of simplified chinese version of liwc in detecting psychological expressions in short texts on social network services. P...

arXiv 2016
[2024]

InProceedings of the 2nd Workshop on Resources and Technologies for Indige- nous, Endangered and Lesser-resourced Languages in Eurasia (EURALI)@ LREC-COLING 2024, pages 41–50

This word mean what: Constructing a singlish dictionary with chatgpt. InProceedings of the 2nd Workshop on Resources and Technologies for Indige- nous, Endangered and Lesser-resourced Languages in Eurasia (EURALI)@ LREC-COLING 2024, pages 41–50. Dagmar Deuber, Jakob Leimgruber, and Andrea Sand

2024
[2025]

Rajiv Bajpai, Danyuan Ho, and Erik Cambria

Open-source llms for text annotation: a practi- cal guide for model setting and fine-tuning.Journal of Computational Social Science, 8(1):17. Rajiv Bajpai, Danyuan Ho, and Erik Cambria. 2016. Developing a concept-level knowledge base for sen- timent analysis in singlish. InInternational Confer- ence on Intelligent Text Processing and Computa- tional Lingu...

2016

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [5]

InPro- ceedings of the 29th International Conference on Computational Linguistics, pages 3924–3936

Singlish message paraphrasing: A joint task of creole translation and text normalization. InPro- ceedings of the 29th International Conference on Computational Linguistics, pages 3924–3936. Scott M Lundberg and Su-In Lee. 2017. A unified ap- proach to interpreting model predictions.Advances in neural information processing systems, 30. Ho Mian-Lian and Jo...

arXiv 2017

[2] [6]

InProceedings of the International AAAI Conference on Web and Social Media, volume 19, pages 1315–1328

Examining the influence of political bias on large language model performance in stance classi- fication. InProceedings of the International AAAI Conference on Web and Social Media, volume 19, pages 1315–1328. Ri Chi Ng, Nirmalendu Prakash, Ming Shan Hee, Kenny Tsu Wei Choo, and Roy Ka-Wei Lee. 2024. Sghate- check: Functional tests for detecting hate spee...

arXiv 2024

[3] [2018]

Jacob Eisenstein, Brendan O’Connor, Noah A Smith, and Eric P Xing

Singaporean internet chit chat compared to informal spoken language: Linguistic variation and indexicality in a language contact situation.Journal of Pidgin and Creole Languages, 33:48–90. Jacob Eisenstein, Brendan O’Connor, Noah A Smith, and Eric P Xing. 2014. Diffusion of lexical change in social media.PloS one, 9(11):e113114. Wilkinson Daniel Wong Gonz...

2014

[4] [2022]

InInterna- tional Conference on Asian Digital Libraries, pages 115–124

Singlish checker: A tool for understanding and analysing an english creole language. InInterna- tional Conference on Asian Digital Libraries, pages 115–124. Springer. Wesley Lee. 2016. In defense of singlish: A cultural interpretation of singapore english.The Morningside Review, 12. Zhengyuan Liu, Shikang Ni, Aiti Aw, and Nancy Chen

2016

[5] [2023]

Nan Zhao, Dongdong Jiao, Shuotian Bai, and Ting- shao Zhu

Prompting multilingual large language models to generate code-mixed texts: The case of south east asian languages.arXiv preprint arXiv:2303.13592. Nan Zhao, Dongdong Jiao, Shuotian Bai, and Ting- shao Zhu. 2016. Evaluating the validity of simplified chinese version of liwc in detecting psychological expressions in short texts on social network services. P...

arXiv 2016

[6] [2024]

InProceedings of the 2nd Workshop on Resources and Technologies for Indige- nous, Endangered and Lesser-resourced Languages in Eurasia (EURALI)@ LREC-COLING 2024, pages 41–50

This word mean what: Constructing a singlish dictionary with chatgpt. InProceedings of the 2nd Workshop on Resources and Technologies for Indige- nous, Endangered and Lesser-resourced Languages in Eurasia (EURALI)@ LREC-COLING 2024, pages 41–50. Dagmar Deuber, Jakob Leimgruber, and Andrea Sand

2024

[7] [2025]

Rajiv Bajpai, Danyuan Ho, and Erik Cambria

Open-source llms for text annotation: a practi- cal guide for model setting and fine-tuning.Journal of Computational Social Science, 8(1):17. Rajiv Bajpai, Danyuan Ho, and Erik Cambria. 2016. Developing a concept-level knowledge base for sen- timent analysis in singlish. InInternational Confer- ence on Intelligent Text Processing and Computa- tional Lingu...

2016