Semantically Cohesive Word Grouping in Indian Languages

Adyasha Patra; Arnab Bhattacharya; Chaitali Dangarikar; Ganesh Ramakrishnan; Nagasai Saketh Naidu; N J Karthika

REVIEW 3 major objections 6 minor 3 references

Semantically Cohesive Word Grouping in Indian Languages

T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Grouping Hindi words into semantic units makes its sentence structure match other Indian languages and improves machine translation.

desk verdict A concrete Hindi word-grouping recipe with a useful linguistic motivation, but the central MT evidence is confounded by segmentation statistics and needs a matched-length control before the downstream claim can be trusted. read the letter →

arxiv 2501.03988 v1 pith:X4TK75ZC submitted 2025-01-07 cs.CL

N J Karthika , Adyasha Patra , Nagasai Saketh Naidu , Arnab Bhattacharya , Ganesh Ramakrishnan , Chaitali Dangarikar This is my paper · ORCID

classification cs.CL

keywords wordgroupingsemanticunitsHindiIndianlanguagesdependencyparsingmachinetranslationfew-shotpromptingagglutination

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 argues that the apparent syntactic differences among Indian languages are largely an artifact of tokenization: the same semantic unit appears as one word in agglutinative languages like Kannada or Malayalam but as several whitespace-separated words in Hindi. It proposes treating the smallest indivisible semantic unit—called a word group—as the basic unit of processing, and gives three grouping criteria (noun plus postposition, verb plus auxiliary, and multi-word named entities) along with a rule-based automatic grouper for Hindi. The paper tries to show that grouping makes Hindi dependency parse trees resemble those of other Indian languages and improves downstream NLP, specifically few-shot machine translation. The authors demonstrate this with a perturbation experiment, where shuffled sentences stay closer to the original when word groups are preserved, and an extrinsic DecoMT experiment, where grouped chunks outperform fixed-length chunks when translating Hindi into five languages.

What carries the argument

The load-bearing object is the word group: the smallest indivisible, semantically complete unit of a sentence that expresses a single linguistic function, a concept the paper links to the Indian linguistic tradition of 'ēkārthībhāva'. Three grouping criteria carry the argument: inflectional unity (noun plus postposition or case marker), derivational unity (verb plus auxiliary), and named-entity unity. To apply these automatically, the paper mines frequent dependency relations and part-of-speech tags from a Hindi treebank annotated with kāraka relations, has linguists verify the resulting rules, and then runs a transformer-based NLP toolkit (trankit) over Hindi sentences to obtain the POS and dependency values the rules act on. The same word-group units drive both evaluation instruments: the perturbation setup preserves groups during shuffling, and the DecoMT setup replaces fixed-length chunks with word-group chunks.

What would settle it

Compare the perturbation and DecoMT results when the input is grouped by the automatic rules against grouping by a gold standard produced independently by linguists; if the improvements vanish or reverse with gold grouping, the reported gains are an artifact of the rule-generation pipeline rather than of semantic units.

Watch

Extended reading notes

Core claim

The central claim is that a whitespace-separated word is the wrong granularity for computational processing of Indian languages; the right unit is a semantically cohesive word group, defined as the smallest indivisible unit expressing a single linguistic function. Grouping nouns with postpositions (such as राम ने), verbs with auxiliaries (such as जा रहा है), and multi-word named entities (such as श्री ए.पी.जे. अब्दुल कलाम) makes Hindi's dependency structures align with those of its more agglutinative relatives, because the apparent structural differences were due to how many typographic words each language uses for one semantic unit. The paper further claims that using these groups as chunks in decomposed few-shot prompting (DecoMT) improves translation quality from Hindi to Malayalam, Kannada, Sanskrit, Bengali, and Marathi, and that perturbing sentences while preserving groups keeps sentence embeddings closer to the original than perturbing at the word level.

Load-bearing premise

The automatic grouping rules, learned from Hindi treebank statistics and applied through a parser's POS and dependency output, actually identify true semantic units; any parser or tagger error, or a rule misapplied to a sentence, corrupts the grouped input and could create the apparent benefits without them being linguistically real.

Editorial extensions

If this is right

Word grouping should become a standard preprocessing step for Hindi before dependency parsing, cross-lingual alignment, or other structural NLP tasks, since it produces the same parse structure that other Indian languages have without grouping.
Using grouped chunks instead of fixed-size chunks in DecoMT gives consistent spBLEU and chrF++ gains for Hindi-to-Malayalam, Kannada, Sanskrit, Bengali, and Marathi translation, with the largest chrF++ improvements at longer sentence lengths.
Grouping reduces Hindi's apparent word-count deviation in parallel data: grouped Hindi has 18,980 words versus 25,643 ungrouped in FLORES-200 devtest, bringing it close to Kannada, Sanskrit, Bengali, and Marathi.
Preserving word groups during shuffling keeps sentence embedding similarity higher than shuffling individual words, which the paper reads as evidence that groups are the units carrying semantic roles.
For highly agglutinated languages, the paper's limitation note implies the converse operation—splitting a single word into constituents—may be needed, and the proposal is designed to be extended to splitting as well as grouping.

Reading between the lines

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

If the grouping effect is real, the same preprocessing could reduce the token-count imbalance that language models exhibit across languages, since the paper's own word-count table shows grouping shrinks cross-lingual disparities; a testable extension would measure grouped Hindi tokens against tokens of other Indian languages in a multilingual tokenizer.
The perturbation result suggests word-group-preserving shuffling could serve as a cheap data-augmentation strategy for Indian-language sentence encoders, because the authors use it only as an evaluation signal, not as a training technique.
A stricter test would apply the same rule-based grouping to other Indo-Aryan languages such as Marathi and Bengali and check whether their parse trees align with Dravidian languages as well as Hindi's do; the paper only reports Hindi-to-others results.
The dependence on trankit's POS and dependency predictions means the claimed gains may partly reflect parser behavior rather than linguistic units; a human-verified gold grouping benchmark would separate these explanations.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Dear Colleague,

The paper proposes a rule-based word grouping procedure for Hindi, motivated by the observation that whitespace-separated words in Hindi often correspond to a single semantic unit in more agglutinative Indian languages. The contribution is real but modest: the grouping rules are concrete, derived from dependency/POS statistics of a Hindi treebank and verified by linguists, and the authors evaluate grouped chunks in two ways, an intrinsic perturbation test and an extrinsic DecoMT translation experiment. The paper is honest in its limitations section and does not overclaim. The qualitative parse-tree alignment in Figure 2 is nice and supports the linguistic motivation.

The main soft spot is the translation experiment in Table 2. The comparison is between fixed-length word chunks and word-group-preserving chunks, and these differ simultaneously in boundary placement and length distribution. The reported gains are small (0.2–0.6 spBLEU, 0.29–1.43 chrF++), with no variance or significance testing, so they could easily come from mechanical segmentation statistics rather than semantic cohesion. The paper needs a control that redraws chunk boundaries randomly while matching the per-sentence length distribution of the grouped condition. Without that, the claim that word grouping "aids underlying NLP tasks" is not fully established. The intrinsic perturbation test is also partly tautological: if word groups are defined as semantic units, then preserving them during shuffling naturally keeps sentence embeddings closer. That test supports the grouping quality but is not strong independent evidence.

There is also a minor internal inconsistency in A.4: the note says "we chose the three target languages" while the experiments actually use five. That looks like a leftover from an earlier version.

The automatic grouping depends on trankit outputs, so parser errors can corrupt the groups, but the paper acknowledges this dependence in its limitations. The rules are specified in Table 4, which is a step toward reproducibility, though no code or data is released.

Overall, this is a plausible paper with a useful linguistic insight and a concrete grouping recipe for Hindi. The central claim is not yet settled because of the confounded MT comparison. The paper deserves a serious referee, but only after the authors add a matched-length random-boundary control and report significance or at least variance. If the gains survive that control, the contribution would be modest but real.

Recommendation: send to peer review with a request for a controlled MT experiment. I would not cite it in its current form for the downstream claim, but I might cite the grouping rules once the evidence is cleaner.

Best.

Referee Report

3 major / 6 minor

Summary. This paper argues that whitespace-separated words are not the appropriate atomic units for processing Indian languages, and proposes semantically cohesive word groups, defined as the smallest indivisible semantic units. The authors derive Hindi grouping rules from dependency and POS statistics of the Hindi treebank, apply them via trankit, and evaluate the grouped units in two ways: an intrinsic sentence-perturbation study comparing cosine similarities of shuffled sentences with and without preserved groups, and an extrinsic few-shot machine translation experiment using decomposed prompting (DecoMT) on FLORES-200 from Hindi into five languages. The paper reports consistent improvements for grouping in both evaluations, plus qualitative evidence that grouped Hindi parse trees align more closely with parallel trees in Sanskrit, Malayalam, Kannada, and other languages.

Significance. If the central claim holds, the proposed word grouping is a conceptually useful preprocessing step for Hindi and potentially other Indian languages: it would reduce typographic granularity mismatches in cross-lingual tasks, align dependency structures across related languages, and improve downstream neural MT with decomposed prompting. The paper's strengths include a clear linguistic motivation, a linguist-verified rule set (Appendix A.2), a multilingual qualitative analysis (Figure 2), and an extrinsic evaluation on a standard benchmark (FLORES-200). However, the current experimental evidence does not yet separate the effect of semantic grouping from that of arbitrary chunking statistics, so the practical significance, while plausible, is not established by the reported numbers.

major comments (3)

[§4.2, Table 2] The central extrinsic evidence for the claim that word grouping 'aids underlying NLP tasks' is the DecoMT comparison in Table 2, but the two conditions differ simultaneously in whether chunk boundaries respect semantic units and in the mechanical segmentation statistics (chunk lengths, number of chunks, boundary positions). The reported gains are small (0.2–0.6 spBLEU, 0.29–1.43 chrF++) and are presented without variance, confidence intervals, or significance tests. A control condition that redraws chunk boundaries while preserving the same per-sentence length distribution as the grouped condition is required to attribute the improvement to semantic cohesion rather than to a different segmentation granularity; without it, the main downstream claim is not established.
[§4.1, Table 1] The perturbation experiment is partly circular with respect to the definition of a word group. Since a word group is defined in Section 3 as the smallest indivisible semantic unit, preserving these groups during shuffling is expected to keep sentence embeddings closer to the original than shuffling individual words, even if the grouping rules had no special semantic status. To support the specific claim that the proposed grouping is semantically meaningful, the experiment needs a control in which random word groups of the same size and count are preserved; the same comparison should also report variance or significance, as the differences in Table 1 are small in several language rows (e.g., 0.004 for Sanskrit and 0.007 for Telugu in setting (i)).
[§3.1, Appendix A.2, Table 4] The grouping rules are applied to trankit dependency and POS outputs, so parser or tagger errors will propagate directly into the grouped units used in both evaluations. The paper's Limitations section acknowledges dependence on a deep learning model, but no accuracy measure or error analysis of the automatically grouped output is provided. Since the semantic-unit interpretation of both experiments depends on the quality of this preprocessing, a quantitative assessment of grouping quality (e.g., agreement with the linguist-verified rules on a held-out sample) is needed.

minor comments (6)

[Abstract and Section 1] The phrase 'clause-free word order' should presumably be 'free word order'; please correct this and check for similar typos throughout.
[Table 2] The table formatting is broken: entries such as 'Hindi→Malayalam 18.9 36.87 19.4 37.29Hindi→Kannada' run together and need appropriate spacing or line breaks.
[Appendix A.4] The note says 'we chose the three target languages, which are agglutinative in nature,' but Table 2 lists five target languages (Malayalam, Kannada, Sanskrit, Bengali, Marathi); this is inconsistent and should be clarified.
[References] The trankit reference lists the author as 'V an Nguyen'; this should be 'Van Nguyen' to match the actual author name.
[Section 4.1] The phrase 'In most most of the cases' should be 'In most of the cases'.
[Appendix A.3] Figure 7 is referenced as showing length-bucketed chrF++ scores, but no plot appears in the provided text; either include the figure or remove the reference.

Circularity Check

1 steps flagged · score 4.0 of 10

Intrinsic perturbation support is partly self-definitional; MT experiment provides independent but confounded grounding.

self definitional [Section 2.1 (definition of word group) and Section 4.1 (Sentence Perturbation, Table 1)]
"we define a word group to be the smallest indivisible, semantically complete and meaningful unit of a sentence ... Our hypothesis is that word grouping allows sentences to preserve semantic roles/identities of its components, even on random shuffling. ... In most most of the cases, jumbled sentences with word groups preserved show higher similarity to original sentences. This establishes the significance of word grouping."

The claimed confirmation follows from the construction, not from independent evidence of semantic cohesion. Word groups are defined as semantic units (Sec. 2.1) and are always contiguous word spans (Sec. 3, examples such as rāma nē and jā rahā hai). Inter-group shuffling preserves these contiguous spans exactly, whereas word-level shuffling destroys all local n-gram order. Since the similarity is measured with Sentence-BERT, an order-sensitive encoder, the grouped condition is closer to the original by construction; any arbitrary contiguous chunking would yield the same ordering effect. The intrinsic experiment therefore restates the definition of a group as a semantic unit and observes the necessary consequence, rather than testing whether the learned groups are semantic.

full rationale

The central downstream claim, that word grouping aids MT, is evaluated by an external DecoMT experiment (Sec. 4.2, Table 2) with consistent improvements over fixed-length chunks; that result is not forced by the definition of a group, though it is confounded (grouped and baseline chunks differ in length distribution and boundary placement, with no significance testing), which is a validity concern rather than circularity. The intrinsic perturbation experiment in Sec. 4.1 is partly self-definitional as detailed above: groups are defined as semantic units and implemented as contiguous spans, so preserving them trivially preserves local order that the order-sensitive Sentence-BERT encoder rewards. Self-citations to Dangarikar et al. (2024) motivate the grouping criteria and the Hindi focus but do not by themselves force the MT outcomes; external evidence (Gerz et al., 2018; Pimpale et al., 2014) also supports the Hindi-specific premise. The Limitations section candidly notes the dependence on another deep-learning model (trankit) for automatic grouping, which is a robustness limitation, not a circularity. Overall, the paper's strongest independent evidence is the MT experiment, which is not equivalent to its inputs by construction; the intrinsic pillar is partly tautological, so the score is 4 rather than 0.

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

The central claims rest on the assumption that Hindi is sufficiently less agglutinative that grouping (rather than splitting) is the right operation, on the validity of the perturbation metric, and on the accuracy of the rule-based grouping pipeline. No explicit numerical free parameters are fitted to the reported results.

assumptions (4)

domain assumption Hindi is one of the least agglutinative Indian languages, so grouping whitespace-separated words (not splitting them) is the corrective operation.
Based on Dangarikar et al., 2024 and Gerz et al., 2018 word-count statistics; not independently verified here, and it drives the entire Hindi focus.
domain assumption Dependency parse tree structure is similar across Indian languages once semantic units are aligned; apparent differences are mostly granularity.
Stated in the introduction and illustrated with Figure 2; no quantitative cross-language parse comparison is provided.
domain assumption Sentence-BERT cosine similarity of shuffled sentences is a valid proxy for semantic coherence.
Section 4.1 uses this as the intrinsic metric; if the embedding model is insensitive to word order inside groups, the result may be an artifact.
domain assumption The rule-based grouping generated from Hindi treebank statistics plus trankit annotations is accurate.
Section 3.1 and Appendix A.2; no error analysis of the grouping component is given.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Semantically Cohesive Word Grouping in Indian Languages." pith.science (2026). https://pith.science/paper/X4TK75ZC

@misc{pith2026250103988,
  author       = {Pith},
  title        = {Pith review of: Semantically Cohesive Word Grouping in Indian Languages},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X4TK75ZC}},
  note         = {Machine review of arXiv:2501.03988}
}

read the original abstract

Indian languages are inflectional and agglutinative and typically follow clause-free word order. The structure of sentences across most major Indian languages are similar when their dependency parse trees are considered. While some differences in the parsing structure occur due to peculiarities of a language or its preferred natural way of conveying meaning, several apparent differences are simply due to the granularity of representation of the smallest semantic unit of processing in a sentence. The semantic unit is typically a word, typographically separated by whitespaces. A single whitespace-separated word in one language may correspond to a group of words in another. Hence, grouping of words based on semantics helps unify the parsing structure of parallel sentences across languages and, in the process, morphology. In this work, we propose word grouping as a major preprocessing step for any computational or linguistic processing of sentences for Indian languages. Among Indian languages, since Hindi is one of the least agglutinative, we expect it to benefit the most from word-grouping. Hence, in this paper, we focus on Hindi to study the effects of grouping. We perform quantitative assessment of our proposal with an intrinsic method that perturbs sentences by shuffling words as well as an extrinsic evaluation that verifies the importance of word grouping for the task of Machine Translation (MT) using decomposed prompting. We also qualitatively analyze certain aspects of the syntactic structure of sentences. Our experiments and analyses show that the proposed grouping technique brings uniformity in the syntactic structures, as well as aids underlying NLP tasks.

Figures

Figures reproduced from arXiv: 2501.03988 by the authors.

**Figure 2.** Dependency Parse Trees • Named entities: A named entity (NE) with multiple words form a single group. Example groups: श्री ए पी जे अब्दुल कलाम (śrī ē pī jē abdula kalāma), अरुणाचल प्रदेश (aruṇācala ṇ pradēśa) 3.1 Similar Syntactic Structures A dependency parse tree is a syntactic structural representation of a sentence, where the words or phrases form the nodes, and the edges show the dependencies between the nodes … view at source ↗

**Figure 4.** shows the total number of words in different Indian languages for the parallel sentences, representing the same content in FLORES200 devtest data (Costajussà et al., 2022). The graph also contains some nonIndian languages to show the extent to which the number of words can vary across languages to convey the same information. Note that, the number of words in Jingpho is over 3.6× the number of words in the cor… view at source ↗

Figures from the paper (2 more)

**Figure 6.** Figure 6: Proposed Prompt Template for Independent [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

**Figure 7.** Figure 7: The plots show the relationship between source sentence length and chrF++ scores for translation from [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

3 extracted references · 2 canonical work pages

[1991]

Frontiers in Knowledge Based Computing (KBCS90), VP Bhatkar and KM Rege (eds.), Narosa Publishing House, New Delhi , pages 277–296

Local word grouping and its relevance to indian languages. Frontiers in Knowledge Based Computing (KBCS90), VP Bhatkar and KM Rege (eds.), Narosa Publishing House, New Delhi , pages 277–296. Kenneth Ward Church. 1989. A stochastic parts pro gram and noun phrase parser for unrestricted text. In International Conference on Acoustics, Speech, and Signal Pro...

arXiv 1989
[2014]

Decomposed Prompting for Machine Translation Between Related Languages using Large Language Models

Smt from agglutinative languages: use of suf fix separation and word splitting. In Proceedings of the 11th International Conference on Natural Lan guage Processing, pages 2–10. Ratish Puduppully, Anoop Kunchukuttan, Raj Dabre, Ai Ti Aw, and Nancy F Chen. 2023. Decom posed prompting for machine translation between related languages using large language ...

work page Pith review arXiv 2018
[2023]

Open Information Extraction via Chunks

Open information extraction via chunks. arXiv preprint arXiv:2305.03299. Daniela Gerz, Ivan Vulić, Edoardo Ponti, Jason Narad owsky, Roi Reichart, and Anna Korhonen. 2018. Lan guage modeling for morphologically rich languages: Characteraware modeling for wordlevel prediction. Transactions of the Association for Computational Linguistics, 6:451–465. Pr...

work page Pith review arXiv 2018

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1991]

Frontiers in Knowledge Based Computing (KBCS90), VP Bhatkar and KM Rege (eds.), Narosa Publishing House, New Delhi , pages 277–296

Local word grouping and its relevance to indian languages. Frontiers in Knowledge Based Computing (KBCS90), VP Bhatkar and KM Rege (eds.), Narosa Publishing House, New Delhi , pages 277–296. Kenneth Ward Church. 1989. A stochastic parts pro gram and noun phrase parser for unrestricted text. In International Conference on Acoustics, Speech, and Signal Pro...

arXiv 1989

[2] [2014]

Decomposed Prompting for Machine Translation Between Related Languages using Large Language Models

Smt from agglutinative languages: use of suf fix separation and word splitting. In Proceedings of the 11th International Conference on Natural Lan guage Processing, pages 2–10. Ratish Puduppully, Anoop Kunchukuttan, Raj Dabre, Ai Ti Aw, and Nancy F Chen. 2023. Decom posed prompting for machine translation between related languages using large language ...

work page Pith review arXiv 2018

[3] [2023]

Open Information Extraction via Chunks

Open information extraction via chunks. arXiv preprint arXiv:2305.03299. Daniela Gerz, Ivan Vulić, Edoardo Ponti, Jason Narad owsky, Roi Reichart, and Anna Korhonen. 2018. Lan guage modeling for morphologically rich languages: Characteraware modeling for wordlevel prediction. Transactions of the Association for Computational Linguistics, 6:451–465. Pr...

work page Pith review arXiv 2018