On 'Towards Resolving Word Ambiguity with Word Embeddings'

Author: 1184776211

11/13/2025 — paper summary — 7 min read

Citation

Thurnbauer, Matthias, Johannes Reisinger, Christoph Goller, and Andreas Fischer. 2023. Towards Resolving Word Ambiguity with Word Embeddings. arXiv. https://arxiv.org/abs/2307.13417

Citation summary

Paper	Towards Resolving Word Ambiguity with Word Embeddings
Authors	Matthias Thurnbauer, Johannes Reisinger, Christoph Goller, and Andreas Fischer
Year published	2023
Venue	arXiv2307
Paper URL	https://arxiv.org/abs/2307.13417
Code URL	https://github.com/thurnbauermatthi/WordDisambiguation

Description

In your own words, what is this paper about?

The authors propose a method of identifying ambiguous words through an unsupervised machine learning process. Intuitively, their goal is to capture the intuition that any word's meaning can be represented by its context of usage. This is, that the distinct meanings of ambiguous words can be captured by classifying those contexts into groups of similarity.

Given a target word, their method uses the context of all occurrences of that word. For each context, the trained CBOW embeddings of all context words within k=5 are combined (by averaging) to calculate a vector representation of the context. Those contexts are then grouped via DBSCAN (Density-Based Spatial Clustering for Applications with Noise), thereby grouping distinct meanings of the target word, while excluding non-informative contexts as noise.

Because DBSCAN relies on provided parameters—a nearness threshold for each datapoint (epsilon) and a minumum number of near datapoints within that threshold—which must be determined, Thurnbauer et al. ran the model through an extensive combination of parameter settings and evalutated the results based on sci-kitlearn's silhoutte_score() function.

The method proved effective at distinguishing an ambiguous word's meanings, demonstrated comparing the grouped context vectors to other words, yielding similar and related other words.

Motivation

Why did you select this paper?

The topic of unsupervised word-disambiguation appeals to me for the reason that it pertains to a project that I've been tinkering with: exploring ways to create an automated solver for the New York Times Games daily Connections puzzle. Not only is the Connections puzzle based in grouping of similar words, but a challenge for human players is that many of the words included in any given puzzle are ambiguous.

The work presented in the paper, i.e. identifying a word's multiple meanings and subsequently grouping those distinct meanings with other related words, is directly applicable to the Connections challenge.

LLM

LLM model summary

LLM model	HuggingFaceTB/SmolLM3-3B
LLM model version	v3
Model/service URL	https://huggingface.co/chat/models/HuggingFaceTB/SmolLM3-3B
Why this model?	The choice was largely arbitrary, but motivated by the assignment's specific call-out to SmolLM3 and HuggingChat.

Experience using the LLM

Despite a general aversion to GUIs, I opted for the HuggingChat interface after a bit of dabbling with running a model locally. The shift was motivated solely for the purpose of prioritizing the immediate task of interacting with the model.

I provided the paper to the model via its System Prompt (using a cleaned version of the article's LaTeX source) along with a general prompt intended to reduce the verbosity of responses:

1You are an expert in NLP (Natural Language Processing) responding to an audience who has general familiarity with NLP concepts.
2Your responses to all questions are succinct and to the point.
3You do not deviate from the topic at hand.
4You do not enumerate excessively. Your audience does not like lists.
5Be brief in all cases.

It was quickly clear that the instruction to be brief was deprioritized after only a single prompt, and responses degraded into tremendously long lists of detailed items. This tendency proved consistent if any single prompt did not also include a reminder of the brevity instruction.

Description (LLM)

1What is this paper about? Be brief and succinct.

The paper explores an unsupervised approach to resolve word ambiguity using DBSCAN clustering on word embeddings, aiming to identify and evaluate the ambiguity level of words in a document corpus. It focuses on leveraging CBOW-based word embeddings to cluster context vectors of words, creating semantically coherent clusters that correspond to different meanings. The method automatically selects DBSCAN parameters via a silhouette score and noise ratio to balance cluster quality and specificity, providing an interpretable and cost-effective solution for disambiguation in NLP tasks.

What are the authors proposing?

1What are the authors proposing? Be brief and succinct.

The authors propose an unsupervised approach to resolve word ambiguity using DBSCAN clustering on word embeddings, specifically focusing on context vectors (context embeddings of words surrounding a target word). Their method identifies clusters that represent different meanings of a word by analyzing co-occurrence patterns in the context space. Key components include:
DBSCAN clustering on word embeddings to group context vectors into semantically coherent clusters.
Parameter selection via a silhouette score and noise ratio to balance cluster quality and specificity.
Cluster labeling using cosine similarity in input embeddings (In2In) and predicted center words (In2Out).
The goal is to automatically identify and evaluate ambiguity levels of words in a document corpus, enabling disambiguation for tasks like query resolution in information retrieval systems.

What is the motivation for the work?

1What is the motivation for the work? Be brief and succinct.

The motivation is to develop an unsupervised, lightweight method to resolve word ambiguity in natural language processing (NLP), leveraging word embeddings and DBSCAN clustering. Traditional approaches (e.g., Transformers) are too resource-intensive for real-time or sensitive environments, while supervised WSD methods require labeled data. The work aims to:

Automatically identify ambiguous words and their meanings via context vector clustering.
Provide a parameter-optimized clustering strategy to balance cluster quality and specificity.
Enable interpretable disambiguation for practical applications like query resolution in search engines or chatbots.
Address the challenge of most words being ambiguous, emphasizing the need for efficient, unsupervised solutions without labeled data.

By focusing on context-based clustering, the authors aim to capture semantic coherence and practical relevance, offering a scalable, cost-effective alternative to complex models.

What is the approach or innovation?

1What is the authors' innovation? Keep it brief and succinct.

The authors' motivation is to develop a practical, unsupervised method to resolve word ambiguity in NLP, leveraging word embeddings and DBSCAN clustering to address the limitations of complex models and the need for efficient, real-time disambiguation. Their goal is to:
Automatically identify ambiguous words and their meanings via context-based clustering.
Provide a parameter-optimized clustering strategy for balanced cluster quality and specificity.
Enable interpretable disambiguation for applications like query resolution in search engines or chatbots.
Offer a cost-effective solution for sensitive environments, avoiding the computational overhead of Transformers or labeled data in WSD.
Demonstrate the effectiveness of unsupervised approaches on real-world datasets (e.g., German Wikipedia) for practical NLP tasks.

What are the results and how do they compare with competing approaches?

1What are the results and how do they compare with competing approaches? Keep it brief. Keep it succinct.

The authors' results demonstrate that DBSCAN clustering applied to word embeddings effectively identifies ambiguous words and their meanings, producing semantically coherent clusters. Key findings include:
Parameter optimization yields high-quality clusters with parameter scores exceeding manual counts of Wikipedia homonymous meanings (e.g., 110–104 clusters for "mars").
Context vector clustering groups words into meanings based on co-occurrence patterns, with labels derived from context or predicted center words.
High-quality clusters (e.g., 110 for "mars") align with domain-specific and practical meanings, as validated by manual analysis.
Efficiency and scalability: The method requires moderate computational resources, contrasting with Transformers' high costs.
Comparison with competing approaches:
Unsupervised vs. supervised: The authors' unsupervised method avoids the need for labeled data, unlike traditional WSD.
DBSCAN vs. other clustering: Their parameter-optimized DBSCAN outperforms fixed-cluster approaches (e.g., ), achieving higher coherence and relevance.
Efficiency: The method is more resource-efficient than Transformers, making it suitable for real-time or sensitive environments.
Interpretability: Clusters and labels are derived from context vectors, providing clear, practical disambiguation for NLP tasks.
Overall, the approach balances computational efficiency, semantic coherence, and practical relevance, offering a scalable solution for resolving word ambiguity in NLP applications.

Is the comparison fair?

1In brief, succinct terms, is that comparison fair?

The comparison in the paper is fair and valid, as it:
Highlights the unsupervised nature of the method, contrasting with supervised approaches like WSD, which requires labeled data.
Validates the efficiency of DBSCAN-based clustering over Transformers, emphasizing computational savings in real-time or sensitive environments.
Demonstrates practical relevance by showing that clusters are semantically coherent and align with domain-specific meanings, a critical advantage for NLP applications.
Addresses parameter optimization as a unique strength, improving cluster quality through tailored DBSCAN parameters, which is a fair comparison to fixed-parameter methods.
The comparison is balanced because it:
Comes with context: The authors explain the suitability of their method for real-time use, underscoring that Transformers are not ideal for such scenarios.
Focuses on domain-specific goals: They compare clustering effectiveness in disambiguating ambiguous words, a task where their method excels.
Avoids apples-to-apples comparisons: Transformers are generative models, not clustering algorithms, so direct comparisons are justified within their scope.
The comparison is fair because it justifies the method's strengths in the context of practical NLP tasks and resource constraints, making it a valid and useful evaluation.

1What, according to the authors, are the key takeaways? Only mention broad strokes, being brief and succinct.

The key takeaways from the paper are:
Unsupervised disambiguation of words via DBSCAN clustering on context vectors, identifying clusters representing different meanings.
Efficient parameter optimization using silhouette score and noise ratio to balance cluster quality and specificity.
Practical relevance in NLP tasks, enabling real-time disambiguation for applications like search engines or chatbots.
Unsupervised approach that avoids labeled data, making it scalable and suitable for large datasets.
Resource-efficient method requiring moderate computational resources, ideal for sensitive environments.
Interpretable results through context-based clustering, aiding in understanding ambiguous words' meanings.
Domain-specific applicability, applicable to languages like German and potentially others, with potential for further adaptation.

The method offers a balanced trade-off between computational efficiency and semantic accuracy, making it a viable solution for resolving word ambiguity in NLP.

What are the takeaways according to you?

1What, according to you, are the key takeaways?

Key takeaways:
Unsupervised disambiguation of words via DBSCAN clustering on context vectors.
Parameter optimization via silhouette score and noise ratio for high-quality clusters.
Efficiency and scalability in NLP, suitable for real-time systems.
Interpretable results with semantically coherent clusters.
Practical relevance in disambiguating ambiguous words for applications like search engines or chatbots.
Unsupervised approach avoiding labeled data, making it resource-efficient and adaptable.
Domain-specific applicability, applicable to languages like German and potentially others.

Would you use this? If so, how/where would you use this?

1Would you use this? Where? How?

Use if:
You need an unsupervised method for real-time disambiguation in NLP.
You require efficiency and scalability without labeled data.
Your applications involve real-time systems or limited computational resources.
You want interpretable results with semantically coherent clusters.
Not ideal if:
You need high accuracy in labeled data scenarios.
You require extensive computational resources (e.g., Transformers).
You need domain-specific accuracy beyond general semantic coherence.
Applicable to:
NLP tasks like search engines, chatbots, or text classification.
Languages with high ambiguity (e.g., German).
Datasets requiring unsupervised learning without prior knowledge.

What problems remain and what are the next steps?

1What problems remain? What should be next steps?

Remaining problems:
Clusters may not perfectly align with the expected number of meanings (e.g., under/over-counting).
Parameter optimization (epsilon, min_samples) may not capture all nuances of ambiguity.
Limited applicability to specific domain contexts (e.g., named entities, specialized terms).
Next steps:
Leverage IDF-weighting to reduce irrelevant context influence.
Explore larger window sizes in CBOW for domain-level clusters.
Improve tokenizer for better named entity/noun phrase recognition.
Evaluate the method on larger datasets for generalizability.
Compare with other unsupervised methods for robustness and efficiency.

Errors and limitations of the LLM

Where did it fall short or make mistakes?

The model generally provided a good sense of the article and captured the major points without any significant errors.

That being said, it repeatedly make reference to real-time efficiency of the approach, which I did not take away as a main point being made. Considering that I may have overlooked this emphasis or that I may have been incredulous based on my own experience with the time-efficiency of the approach, I thought to have the model identify the source thereof:

1You mentioned the approach in the paper being efficient and lightweight. Point me to where the authors call this out as a benefit of their approach.

The authors explicitly state their method's efficiency and lightweight nature in the conclusion section and results comparison with Transformers. They argue that:
Their approach is more resource-efficient than Transformers, making it suitable for real-time systems and sensitive environments.
Parameter optimization (via silhouette score and noise ratio) reduces manual tuning complexity, enhancing scalability.
The method is unsupervised and unsupervised with moderate computational resources, contrasting with Transformers' high demands.
Key quotes:
"Our method is a practical solution for real-time disambiguation in NLP, emphasizing efficiency and scalability."
"Our approach is more efficient than Transformers, making it ideal for applications requiring real-time processing."
"Parameter optimization ensures high-quality clusters with minimal computational overhead."
These points are found in the conclusion and results comparison sections of the paper.

Indeed, those quotes do not appear in the paper, nor does a results comparison section.

With the contexts of common words from a large corpus, the DBSCAN calculation alone can be quite resource-intensive. That being said, it does provide a reasonable means of handling ambiguity if the clusters are pre-trained.