Starting AI Similarity - Learning Word2Vec

Blog posts link automatically and apply the Jaccard algorithm, which lacks semantic similarity. Word2Vec identifies semantic similarity by using word vectors, comparing the closeness of words to determine relatedness. A lightweight application for document matching using Word2Vec is still cost-effective when compared with the more expensive computations of modern LLMs.

For simpler applications that don't require generative transformers, Word2Vec still has some uses with a working model. How can we apply this?

Here are some features we are looking to create using Word2Vec and NLP assessment similarities.

1) Blog Linking - Content Clusters

Jaccard matched words alone; this works well for direct matches, but some connections between blog articles only shared common words with little meaningful relationship.

By training on our blog posts, we can provide more relevant links between our blogs, creating content clusters where they occur naturally. Content clusters are a growing topic in the (SEO/AIO) online search and conversational LLM reach optimisation now. Enhancing our ability to produce clusters in the long term should increase the visibility of our website's content.

2) Similar Project Detection - Overlapping Projects

You can't know what is happening everywhere. Using similarity analysis, we can make similar projects more visible. Similar goals or requests can be clustered. Meaning that direct overlap with another team or team member is reduced, or teams struggling with the same issue can be joined.

The comments on projects are used for vector training, and if someone else raises a similar issue, then we can identify it and link the projects for review.

Projects are facing similar blockers, or solutions being developed that could be applied elsewhere.

3) Better Project Search

Projects have parent-child & dependency relations and are typically easy to find. When they are not, you search.

Close vectors for phrases go beyond word matching, finding similar phrases as you search. Presenting them in search results.

Or we could use Elasticsearch. " The most widely deployed vector database" - According to Elasticsearch.

Outside of Project Management Tools

4) Product Information Management

From typos to incorrect information, Word2Vec can spot irregularities in product titles, descriptions and attributes.

Using overlapping similarities on data sets, as few as ten known correct products allow the remaining titles to be checked against the sample data. Any titles falling outside a threshold are identified for further validation. For example, brands and sub-brands mixed accidentally. Irregular sizes in titles, 40mm vs 40m, would be separate words on separate vectors. This could also be applied to specific attribute values.

5) Sentiment analysis,

A widespread use case for word2Vec, where text is classified as positive, neutral or negative. This service is offered in many SAAS products used for brand analysis, social media comment reviews or product reviews.

6) Enhanced search and scrape.

Web scraping is a controversial topic, but there are some use cases where it is useful. Price match similar products if an exact match product isn't available, or if your data isn't 100% accurate and you need to find exact match possibilities for verification.

7) Keyword Expansion - PPC

Paid search in your industry can be quite niche, and a model trained using your appropriate text is more likely to be able to identify keywords, pairs, or triplicates that are closely related than general models.

Next Up?

Word2Vec isn't used in modern NLP. GloVe and or FastText are my next study items to understand static embeddings.

These, too, are not the end of the journey for NPL, but for me, it's essential to understand these concepts to implement lower-resource NPL work pipelines that are cost-effective and add real-world value to the work we do each day as project managers in various fields.

All the use cases that are listed here and others don't need full generative transforms to be effective.