pith. sign in

arxiv: 2604.14397 · v1 · submitted 2026-04-15 · 💻 cs.CL · cs.AI

Generating Concept Lexicalizations via Dictionary-Based Cross-Lingual Sense Projection

David Basil , Chirooth Girigowda , Bradley Hauer , Sahir Momin , Ning Shi , Grzegorz Kondrak This is my paper

Pith reviewed 2026-05-10 13:10 UTC · model grok-4.3

classification 💻 cs.CL cs.AI
keywords cross-lingual projectionsense generationWordNet expansionbilingual dictionarylexical resourcesmultilingual NLPword alignmentsemantic projection
0
0 comments X

The pith

Projecting English senses onto aligned target-language tokens and filtering them with a bilingual dictionary produces more precise lexicalizations for new languages.

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

The paper aims to expand WordNet-style resources to additional languages by generating new senses through cross-lingual projection rather than building them from scratch. It starts from a sense-tagged English corpus and its translation, aligns the tokens, and assigns target lemmas to existing English synsets. The core innovation augments a base aligner with a bilingual dictionary both to improve alignment quality and to remove erroneous projections. A sympathetic reader would care because the approach stays interpretable, uses limited external resources, and shows higher precision than prior projection methods or large-language-model baselines across multiple languages. If the claim holds, lexical resources could be grown more reliably without massive new annotation efforts.

Core claim

Given a sense-tagged English corpus and its translation, the method projects English synsets onto aligned target-language tokens and assigns the corresponding lemmas to those synsets. Alignments are generated by augmenting a pre-trained base aligner with a bilingual dictionary, which is also used to filter out incorrect sense projections. Evaluations on multiple languages demonstrate that this project-and-filter strategy improves precision relative to earlier methods, dictionary-only baselines, and large-language-model baselines while remaining interpretable and resource-light.

What carries the argument

The project-and-filter strategy: English synsets are projected onto dictionary-augmented alignments and then filtered to retain only high-quality target lemmas.

If this is right

  • The generated sense inventories can be released and used directly to enlarge multilingual lexical resources.
  • The approach supplies an interpretable alternative when large language models are unavailable or undesirable.
  • Precision gains hold across several tested languages when the same dictionary-augmented pipeline is applied.
  • Fewer external resources are needed than either pure dictionary lookup or full model fine-tuning.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If dictionary coverage is uneven across language pairs, performance on low-resource languages may drop unless the filter is adjusted.
  • The same projection pipeline could be tested on downstream tasks such as cross-lingual semantic parsing to measure real-world utility.
  • Combining the filter with newer alignment models might raise precision further without losing interpretability.

Load-bearing premise

The bilingual dictionary and the resulting word alignments must be accurate enough that the projected senses contain few errors and that the same process works across different language pairs.

What would settle it

Running the method on a held-out language pair and finding that its precision falls below the plain dictionary baseline or that many filtered projections are still incorrect would show the strategy does not deliver the claimed improvement.

Figures

Figures reproduced from arXiv: 2604.14397 by Bradley Hauer, Chirooth Girigowda, David Basil, Grzegorz Kondrak, Ning Shi, Sahir Momin.

Figure 1
Figure 1. Figure 1: An illustration of our project-and-filter algorithm. candidates for a single token. Second, word senses are represented in a multilingual wordnet as synset-lemma pairs, each corresponding to a single lexical concept. Synonymy therefore functions as a test of conceptual identity: lemmas that are synonymous realize the same sense. Third, when the alignment of words in a sentence and its translation is suppor… view at source ↗
Figure 2
Figure 2. Figure 2: Our dictionary-based alignment procedure, DBAlign. B denotes links proposed by the base aligner, D denotes links between synonyms in the bilingual dictionary. S represents the source sentence, and T the target sentence. Links are added in three passes, subject to a non-conflict constraint. When multiple candidate links are available, the diagonal heuristic selects the link closest to monotonic word order. … view at source ↗
Figure 3
Figure 3. Figure 3: Our algorithmic wordnet expansion algorithm, ExpandNet, where Dict is a set of pairs (x, y) such that y in the target language can express the same meaning as x in the source language in some context, DBAlign(S, T, w) returns the word in sentence T aligned to w as found by the alignment algorithm described in Section 2.2, and Synset(S, x): returns the ID of the synset corresponding to the concept expressed… view at source ↗
read the original abstract

We study the task of automatically expanding WordNet-style lexical resources to new languages through sense generation. We generate senses by associating target-language lemmas with existing lexical concepts via semantic projection. Given a sense-tagged English corpus and its translation, our method projects English synsets onto aligned target-language tokens and assigns the corresponding lemmas to those synsets. To generate these alignments and ensure their quality, we augment a pre-trained base aligner with a bilingual dictionary, which is also used to filter out incorrect sense projections. We evaluate the method on multiple languages, comparing it to prior methods, as well as dictionary-based and large language model baselines. Results show that the proposed project-and-filter strategy improves precision while remaining interpretable and requiring few external resources. We plan to make our code, documentation, and generated sense inventories accessible.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a project-and-filter method for expanding WordNet-style resources to new languages: English synsets from a sense-tagged corpus are projected onto target-language lemmas via word alignments produced by a pre-trained aligner augmented with a bilingual dictionary; the same dictionary then filters incorrect projections. The approach is evaluated on multiple languages against prior projection methods, dictionary-only baselines, and LLM baselines, with the claim that it improves precision while remaining interpretable and requiring few external resources.

Significance. If the empirical claims hold after the requested clarifications, the work offers a practical, low-resource, and transparent alternative to neural methods for cross-lingual sense inventory construction. The explicit use of dictionary augmentation for both alignment improvement and error filtering is a clear strength that could aid reproducibility and adaptation to under-resourced languages.

major comments (2)
  1. [§4 and Table 2] §4 (Evaluation) and Table 2: the headline claim of improved precision rests on the dictionary-augmented aligner producing accurate token links and the filter removing erroneous projections without excessive loss of correct ones. However, the section provides only aggregate precision numbers and does not report alignment accuracy (e.g., F1 or AER) before versus after dictionary augmentation, nor precision/recall of the filter step itself. Without these breakdowns it is impossible to verify that the reported gains are attributable to the proposed components rather than other factors.
  2. [§3.1–3.2] §3.1–3.2 (Method): the description of how dictionary entries are merged into the base aligner’s scoring function is high-level. A concrete example or pseudocode showing the effect on alignment scores for a sentence pair containing a polysemous lemma would make the projection step reproducible and would allow readers to assess the risk of systematic sense-projection errors for particular POS or constructions.
minor comments (2)
  1. [Abstract] Abstract: the languages used in the evaluation are not named; listing them would immediately convey the scope of the claimed generalization.
  2. [§5] §5 (Related Work): the comparison to LLM baselines could briefly note the prompting strategy and model size to allow fair replication.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We are grateful to the referee for the thorough review and the recognition of the method's strengths in providing a practical, interpretable alternative. We respond to the major comments point by point below.

read point-by-point responses

  1. Referee: [§4 and Table 2] §4 (Evaluation) and Table 2: the headline claim of improved precision rests on the dictionary-augmented aligner producing accurate token links and the filter removing erroneous projections without excessive loss of correct ones. However, the section provides only aggregate precision numbers and does not report alignment accuracy (e.g., F1 or AER) before versus after dictionary augmentation, nor precision/recall of the filter step itself. Without these breakdowns it is impossible to verify that the reported gains are attributable to the proposed components rather than other factors.

    Authors: We agree that the absence of these intermediate metrics makes it difficult to fully attribute the precision improvements to the dictionary augmentation and filtering components. In the revised manuscript, we will add reports of alignment accuracy (F1 and AER) before and after dictionary augmentation. We will also include precision and recall for the filter step, computed by evaluating the projections against available gold data or by manual inspection on a sample. These additions will be placed in Section 4 or a new subsection to support the claims. revision: yes

  2. Referee: [§3.1–3.2] §3.1–3.2 (Method): the description of how dictionary entries are merged into the base aligner’s scoring function is high-level. A concrete example or pseudocode showing the effect on alignment scores for a sentence pair containing a polysemous lemma would make the projection step reproducible and would allow readers to assess the risk of systematic sense-projection errors for particular POS or constructions.

    Authors: We acknowledge that the method description in Sections 3.1 and 3.2 could benefit from more concrete details for reproducibility. In the revision, we will add a worked example of a sentence pair with a polysemous lemma, showing how the dictionary augments the alignment scores. We will also include pseudocode for the merging process in an appendix. This will help readers understand the mechanism and evaluate potential systematic errors. revision: yes

Circularity Check

0 steps flagged

No circularity: method uses external dictionaries and alignments as independent inputs

full rationale

The paper describes an algorithmic pipeline that augments a pre-trained base aligner with an external bilingual dictionary to generate token alignments, projects English synsets onto target lemmas via those alignments, and applies the same dictionary to filter erroneous projections. This chain relies on independent, externally sourced resources (dictionary, pre-trained aligner, sense-tagged corpus) rather than any self-defined quantities, fitted parameters presented as predictions, or load-bearing self-citations. No equations reduce outputs to inputs by construction, no uniqueness theorems are imported from the authors' prior work, and no known empirical patterns are merely renamed. The evaluation against prior methods and baselines further anchors the claims outside the method's own definitions, rendering the derivation self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The method relies on standard assumptions in cross-lingual NLP about alignment quality and dictionary accuracy, with no new free parameters or invented entities mentioned.

axioms (2)
  • domain assumption Assumes that translations preserve sense information sufficiently for projection.
    Implicit in the projection method.
  • domain assumption Bilingual dictionary provides reliable sense disambiguation.
    Used for augmentation and filtering.

pith-pipeline@v0.9.0 · 5448 in / 1241 out tokens · 44917 ms · 2026-05-10T13:10:52.671111+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

9 extracted references · 9 canonical work pages

  1. [1]

    XL-WSD: An extra-large and cross-lingual evaluation framework for word sense disambiguation

    T. Pasini, A. Raganato, and R. Navigli. “XL-WSD: An extra-large and cross-lingual evaluation framework for word sense disambiguation”. In:Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 35. 2021, pp. 13648–13656

    work page 2021

  2. [2]

    BabelNet Meaning Representation: A Fully Semantic Formalism to Overcome Language Barriers

    R. Navigli, R. Blloshmi, and A. C. Martínez Lorenzo. “BabelNet Meaning Representation: A Fully Semantic Formalism to Overcome Language Barriers”. In:Proceedings of the AAAI Conference on Artificial Intelligence36.11 (June 2022), pp. 12274–12279

    work page 2022

  3. [3]

    DiBiMT: A Novel Benchmark for Measuring Word Sense Disambiguation Biases in Machine Translation

    N. Campolungo, F. Martelli, F. Saina, and R. Navigli. “DiBiMT: A Novel Benchmark for Measuring Word Sense Disambiguation Biases in Machine Translation”. In:Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Ed. by S. Muresan, P. Nakov, and A. Villavicencio. Dublin, Ireland: Association for Comp...

    work page 2022

  4. [4]

    WiC: the Word-in-Context Dataset for Evaluating Context-Sensitive Meaning Representations

    M. T. Pilehvar and J. Camacho-Collados. “WiC: the Word-in-Context Dataset for Evaluating Context-Sensitive Meaning Representations”. In:Proceedings of NAACL-HLT. 2019, pp. 1267– 1273

    work page 2019

  5. [5]

    Building wordnets by machine translation of sense tagged corpora

    A. Oliver and S. Climent. “Building wordnets by machine translation of sense tagged corpora”. In:Proceedings of the Global WordNet Conference 2012. Matsue, Japó 9-13 gener de 2012. 2012

    work page 2012

  6. [6]

    Parallel corpora for wordnet construction: Machine translation vs. automatic sense tagging

    A. Oliver and S. Climent. “Parallel corpora for wordnet construction: Machine translation vs. automatic sense tagging”. In:International Conference on Intelligent Text Processing and Computational Linguistics. Springer. 2012, pp. 110–121

    work page 2012

  7. [7]

    Automatic creation of WordNets from parallel corpora

    A. Oliver and S. Climent. “Automatic creation of WordNets from parallel corpora”. In:Proceed- ings of the Ninth International Conference on Language Resources and Evaluation (LREC’14). Ed. by N. Calzolari, K. Choukri, T. Declerck, H. Loftsson, B. Maegaard, J. Mariani, A. Moreno, J. Odijk, and S. Piperidis. Reykjavik, Iceland: European Language Resources A...

    work page 2014

  8. [8]

    LexicoMatic: Automatic Creation of Multilingual Lexical-Semantic Dictionaries

    F. Martelli, L. Procopio, E. Barba, and R. Navigli. “LexicoMatic: Automatic Creation of Multilingual Lexical-Semantic Dictionaries”. In:Proceedings of the 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). Ed. by J. ...

    work page 2023

  9. [9]

    A Simple, Fast, and Effective Reparameterization of IBM Model 2

    C. Dyer, V. Chahuneau, and N. A. Smith. “A Simple, Fast, and Effective Reparameterization of IBM Model 2”. In:Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Ed. by L. Vanderwende, H. Daumé III, and K. Kirchhoff. Atlanta, Georgia: Association for Computational ...

    work page 2013