💬 NLP

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

This paper proposes a dictionary-augmented, cross-lingual sense projection method that automatically expands WordNet-style lexical resources to new languages by aligning translated corpora with bilingual dictionaries to generate high-precision, interpretable sense inventories.

Original authors: David Basil, Chirooth Girigowda, Bradley Hauer, Sahir Momin, Ning Shi, Grzegorz Kondrak

Published 2026-04-17

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: David Basil, Chirooth Girigowda, Bradley Hauer, Sahir Momin, Ning Shi, Grzegorz Kondrak

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a massive, incredibly detailed library in English called WordNet. In this library, every word isn't just a word; it's a "concept" connected to other words that mean the same thing. For example, the concept of "happiness" might link the words joy, glee, and bliss together. This library is a superpower for computers, helping them understand human language.

But here's the problem: This library only exists in English. If you want a computer to understand Spanish, French, or Korean, you have to build a whole new library from scratch, which is like trying to build a skyscraper by hand, one brick at a time. It takes forever and is incredibly expensive.

This paper introduces a clever, automated way to build these new libraries by "projecting" the English concepts onto other languages. Think of it as using a high-tech translator to copy the English library's structure into a new language, but with a special safety net to make sure the copies are accurate.

Here is how their method, called ExpandNet, works, broken down into simple steps:

1. The "Shadow Puppet" Analogy (The Core Idea)

Imagine you have a puppet show in English. The puppeteer (the computer) knows exactly which hand movement represents "anger" and which represents "fear."

Now, you want to put on a show in Spanish. Instead of teaching a new puppeteer from scratch, you take the English show and project it onto a screen behind a Spanish-speaking puppeteer.

If the English puppet waves a red flag for "anger," the Spanish puppeteer sees that flag and knows to wave their own red flag for ira (anger).
This is Semantic Projection: Taking a known meaning from one language and casting its "shadow" onto a word in another language.

2. The Problem: Bad Shadows

The problem with just projecting shadows is that sometimes the light gets distorted.

Maybe the English word "bank" (the side of a river) gets projected onto the Spanish word banco (a seat), because the computer thinks they are related.
Or maybe the computer gets confused by idioms. If an English speaker says, "It's raining cats and dogs," and the computer tries to project "cats" onto the Spanish word for cats (gatos), it fails because the Spanish idiom is totally different.

If we just blindly copy everything, our new library will be full of nonsense.

3. The Solution: The "Dictionary Detective" (The Filter)

This is where the authors' secret sauce comes in. They don't just trust the computer's guess; they bring in a Bilingual Dictionary as a strict detective.

Their algorithm, DBAlign, works in three passes, like a security checkpoint:

The VIP Pass (Intersection): If the computer's guess and the dictionary both agree that Word A translates to Word B, they let it through immediately. This is high-confidence.
The Dictionary Pass: If the computer was unsure, but the dictionary says "Yes, these are synonyms," they let it through.
The Base Pass: If the dictionary is silent, they check the computer's original guess one last time, but only if it doesn't conflict with the previous two steps.

The "Filter" Step:
After the alignment is done, they run a final check. If the computer tried to link the English word "garden" to a Spanish word, but the dictionary says "No, that's not a literal translation in this context," they delete the link.

Analogy: Imagine you are trying to match socks. You have a pile of left socks (English) and a pile of right socks (Spanish). The computer tries to pair them up. The Dictionary Detective walks by and says, "Wait, that red sock doesn't match that blue sock, even if they look similar." The pair is thrown out.

4. The Result: A New Library, Built Fast

By using this "Project-and-Filter" strategy, the team can take a text that is already tagged with meanings in English, translate it, and automatically generate a list of new "concept-word" pairs for a new language.

Why is this cool?
- It's cheap: You don't need armies of linguists. You just need a dictionary and a translator.
- It's smart: It ignores the "bad matches" that usually mess up these projects.
- It's scalable: It can work on languages that have very few resources (low-resource languages), helping to democratize AI technology.

Summary

Think of this paper as a construction crew that builds new language libraries by using the English blueprint. But instead of just copying the blueprints blindly, they use a Dictionary Detective to check every single brick. If a brick doesn't fit perfectly, they throw it away. The result is a sturdy, accurate library for a new language, built much faster and with fewer mistakes than before.

1. Problem Statement

The construction of WordNets (semantic resources organizing words into synonym sets or synsets) is a critical task for Natural Language Processing (NLP) applications such as Word Sense Disambiguation (WSD), semantic parsing, and machine translation evaluation. However, automating the construction of WordNets for new languages is challenging, particularly in multilingual settings.

Existing methods for sense generation (adding new lemmas to existing synsets) typically rely on corpus-level statistics derived from parallel data. These approaches suffer from two main limitations:

Scalability: They are sensitive to corpus size and distribution, making them less effective for low-resource languages.
Robustness: They often lack fine-grained control over alignment quality, leading to spurious sense projections.

The authors aim to solve this by developing a method that generates high-quality senses for target languages using sentence-level projection, reducing reliance on large-scale corpus statistics and improving robustness through dictionary-guided filtering.

2. Methodology

The proposed method, named ExpandNet, operates on a sense-tagged English source corpus and its translation into a target language. It utilizes a project-and-filter strategy consisting of two main components:

A. Sense Projection Pipeline

The core logic follows the semantic projection framework:

Input: A parallel bitext where the source sentences (English) are tagged with synset IDs (concepts).
Alignment: The system aligns words in the source sentence to words in the target sentence.
Projection: If a source word $x$ (tagged with concept $s$ ) is aligned with a target word $y$ , the system generates a new sense $(y, s)$ , effectively adding lemma $y$ to synset $s$ .

B. Dictionary-Based Alignment (DBAlign)

To ensure high-quality alignments, the authors introduce DBAlign, a novel alignment algorithm that augments a pre-trained base aligner with a bilingual dictionary ( $D$ ). The algorithm operates in three sequential passes to construct the set of alignment links ( $A$ ):

Pass 1 (Intersection): Accepts links present in both the base aligner's output ( $B$ ) and the dictionary ( $D$ ). These are treated as high-confidence alignments.
Pass 2 (Dictionary): Adds remaining links supported by the dictionary that do not conflict with Pass 1 links.
Pass 3 (Base): Adds remaining links from the base aligner that do not conflict with previously accepted links.

Conflict Resolution:

Non-conflict Constraint: Two links conflict if they share a source or target token. Once a token is linked, it cannot be linked again.
Diagonal Heuristic: If multiple valid candidates exist for a token within a pass, the algorithm selects the link where the source and target words are closest in relative position (monotonic word order), adapted from the fast_align system.
MWE Handling: Multi-word expressions (MWEs) are tokenized as single units if present in the dictionary.

C. Filtering and Validation

A crucial innovation is the dictionary-based filtering step. A sense projection is only accepted if the aligned pair $(x, y)$ is validated by the bilingual dictionary (i.e., $(x, y) \in D$ ).

Exception: If $x$ and $y$ share the same orthographic form (e.g., cognates or proper nouns), the sense is projected regardless of dictionary presence to handle rare words.
POS Filter: An optional filter discards projections where the source and target words do not share the same Part-of-Speech (Noun, Verb, Adjective, Adverb).

3. Key Contributions

Sentence-Level Projection: Unlike prior work relying on corpus-level statistics, this method operates at the sentence level, making it more robust across diverse data conditions and scalable to low-resource scenarios.
Dictionary-Guided Filtering: This is the first application of dictionary-based filtering specifically for sense generation. It significantly reduces spurious alignments by ensuring that only dictionary-verified translations receive sense tags.
DBAlign Algorithm: A novel three-pass alignment procedure that prioritizes high-confidence dictionary evidence while maintaining coverage through a base aligner, resolving conflicts via a diagonal heuristic.
Resource Efficiency: The method requires only a sense-tagged source corpus, a machine translation system (to generate the bitext), and a bilingual dictionary. It does not require pre-existing sense inventories in the target language.

4. Results and Evaluation

The authors evaluated ExpandNet on multiple languages, comparing it against:

Prior projection-based methods.
Dictionary-based baselines.
Large Language Model (LLM) baselines.

Key Findings:

Precision: The proposed "project-and-filter" strategy significantly improves precision compared to other methods. By filtering out non-literal translations, the system avoids generating incorrect sense assignments.
Interpretability: The method is highly interpretable because the logic (dictionary validation) is transparent, unlike "black box" LLM approaches.
Scalability: The approach is effective for languages where comprehensive WordNets do not yet exist, provided a bilingual dictionary is available.

5. Significance

This work addresses a critical bottleneck in multilingual NLP: the lack of structured lexical resources for low-resource languages. By demonstrating that a dictionary-guided, sentence-level approach can outperform corpus-statistics-based methods and complex LLMs in terms of precision and interpretability, the authors provide a practical, accessible pipeline for expanding WordNets.

The authors plan to open-source their code, documentation, and the generated sense inventories, facilitating further research and application in multilingual semantic processing. This approach shifts the paradigm from "statistical likelihood" to "lexical verification," offering a more reliable path for cross-lingual sense generation.