Find Keyword Cannibalization Using OpenAI’s Text Embeddings With Examples via @sejournal, @vahandev

1 year ago 213

ARTICLE AD BOX

This caller bid of articles focuses connected moving with LLMs to standard your SEO tasks. We anticipation to assistance you integrate AI into SEO truthful you tin level up your skills.

We anticipation you enjoyed the previous article and recognize what vectors, vector distance, and text embeddings are.

Following this, it’s clip to flex your “AI cognition muscles” by learning however to usage substance embeddings to find keyword cannibalization.

We volition commencement with OpenAI’s substance embeddings and comparison them.

Model Dimensionality Pricing Notes

text-embedding-ada-002	1536	$0.10 per 1M tokens	Great for astir usage cases.
text-embedding-3-small	1536	$0.002 per 1M tokens	Faster and cheaper but little accurate
text-embedding-3-large	3072	$0.13 per 1M tokens	More close for analyzable agelong text-related tasks, slower

(*tokens tin beryllium considered arsenic words words.)

But earlier we start, you request to instal Python and Jupyter connected your computer.

Jupyter is simply a web-based instrumentality for professionals and researchers. It allows you to execute analyzable information investigation and instrumentality learning exemplary improvement utilizing immoderate programming language.

Don’t interest – it’s truly casual and takes small clip to decorativeness the installations. And remember, ChatGPT is your friend erstwhile it comes to programming.

In a nutshell:

Download and install Python.
Open your Windows bid enactment oregon terminal connected Mac.
Type this commands pip instal jupyterlab and pip instal notebook
Run Jupiter by this command: jupyter lab

We volition usage Jupyter to experimentation with substance embeddings; you’ll spot however amusive it is to enactment with!

But earlier we start, you indispensable motion up for OpenAI’s API and acceptable up billing by filling your balance.

Open AI Api Billing settings

Once you’ve done that, acceptable up email notifications to pass you erstwhile your spending exceeds a definite magnitude nether Usage limits.

Then, get API keys nether Dashboard > API keys, which you should support backstage and ne'er stock publicly.

OpenAI API keys

Now, you person each the indispensable tools to commencement playing with embeddings.

Open your machine bid terminal and benignant jupyter lab.
You should spot thing similar the beneath representation popular up successful your browser.
Click connected Python 3 nether Notebook.

jupyter lab

In the opened window, you volition constitute your code.

As a tiny task, let’s radical akin URLs from a CSV. The sample CSV has 2 columns: URL and Title. Our script’s task volition beryllium to radical URLs with akin semantic meanings based connected the rubric truthful we tin consolidate those pages into 1 and hole keyword cannibalization issues.

Here are the steps you request to do:

Install required Python libraries with the pursuing commands successful your PC’s terminal (or successful Jupyter notebook)

pip instal pandas openai scikit-learn numpy unidecode

The ‘openai’ room is required to interact with the OpenAI API to get embeddings, and ‘pandas’ is utilized for information manipulation and handling CSV record operations.

The ‘scikit-learn’ room is indispensable for calculating cosine similarity, and ‘numpy’ is indispensable for numerical operations and handling arrays. Lastly, unidecode is utilized to cleanable text.

Then, download the sample expanse arsenic a CSV, rename the record to pages.csv, and upload it to your Jupyter folder wherever your publication is located.

Set your OpenAI API cardinal to the cardinal you obtained successful the measurement above, and copy-paste the codification beneath into the notebook.

Run the codification by clicking the play triangle icon astatine the apical of the notebook.

import pandas arsenic pd import openai from sklearn.metrics.pairwise import cosine_similarity import numpy arsenic np import csv from unidecode import unidecode # Function to cleanable text def clean_text(text: str) -> str: # First, regenerate known problematic characters with their close equivalents replacements = { 'â€“': '–', # en dash 'â€™': '’', # close azygous quotation mark 'â€œ': '“', # near treble quotation mark 'â€': '”', # close treble quotation mark 'â€˜': '‘', # near azygous quotation mark 'â€': '—' # em dash } for old, caller successful replacements.items(): substance = text.replace(old, new) # Then, usage unidecode to transliterate immoderate remaining problematic Unicode characters substance = unidecode(text) instrumentality text # Load the CSV record with UTF-8 encoding from basal folder of Jupiter task folder df = pd.read_csv('pages.csv', encoding='utf-8') # Clean the 'Title' file to region unwanted symbols df['Title'] = df['Title'].apply(clean_text) # Set your OpenAI API key openai.api_key = 'your-api-key-goes-here' # Function to get embeddings def get_embedding(text): effect = openai.Embedding.create(input=[text], engine="text-embedding-ada-002") instrumentality response['data'][0]['embedding'] # Generate embeddings for each titles df['embedding'] = df['Title'].apply(get_embedding) # Create a matrix of embeddings embedding_matrix = np.vstack(df['embedding'].values) # Compute cosine similarity matrix similarity_matrix = cosine_similarity(embedding_matrix) # Define similarity threshold similarity_threshold = 0.9 # since threshold is 0.1 for dissimilarity # Create a database to store groups groups = [] # Keep way of visited indices visited = set() # Group akin titles based connected the similarity matrix for one successful range(len(similarity_matrix)): if one not successful visited: # Find each akin titles similar_indices = np.where(similarity_matrix[i] >= similarity_threshold)[0] # Log comparisons print(f"\nChecking similarity for '{df.iloc[i]['Title']}' (Index {i}):") print("-" * 50) for j successful range(len(similarity_matrix)): if one != j: # Ensure that a rubric is not compared with itself similarity_value = similarity_matrix[i, j] comparison_result = 'greater' if similarity_value >= similarity_threshold other 'less' print(f"Compared with '{df.iloc[j]['Title']}' (Index {j}): similarity = {similarity_value:.4f} ({comparison_result} than threshold)") # Add these indices to visited visited.update(similar_indices) # Add the radical to the list radical = df.iloc[similar_indices][['URL', 'Title']].to_dict('records') groups.append(group) print(f"\nFormed Group {len(groups)}:") for point successful group: print(f" - URL: {item['URL']}, Title: {item['Title']}") # Check if groups were created if not groups: print("No groups were created.") # Define the output CSV file output_file = 'grouped_pages.csv' # Write the results to the CSV record with UTF-8 encoding with open(output_file, 'w', newline='', encoding='utf-8') arsenic csvfile: fieldnames = ['Group', 'URL', 'Title'] writer = csv.DictWriter(csvfile, fieldnames=fieldnames) writer.writeheader() for group_index, radical successful enumerate(groups, start=1): for leafage successful group: cleaned_title = clean_text(page['Title']) # Ensure nary unwanted symbols successful the output writer.writerow({'Group': group_index, 'URL': page['URL'], 'Title': cleaned_title}) print(f"Writing Group {group_index}, URL: {page['URL']}, Title: {cleaned_title}") print(f"Output written to {output_file}")

This codification reads a CSV file, ‘pages.csv,’ containing titles and URLs, which you tin easy export from your CMS oregon get by crawling a lawsuit website utilizing Screaming Frog.

Then, it cleans the titles from non-UTF characters, generates embedding vectors for each rubric utilizing OpenAI’s API, calculates the similarity betwixt the titles, groups akin titles together, and writes the grouped results to a caller CSV file, ‘grouped_pages.csv.’

In the keyword cannibalization task, we usage a similarity threshold of 0.9, which means if cosine similarity is little than 0.9, we volition see articles arsenic different. To visualize this successful a simplified two-dimensional space, it volition look arsenic 2 vectors with an space of astir 25 degrees betwixt them.

<span class=

In your case, you whitethorn privation to usage a antithetic threshold, similar 0.85 (approximately 31 degrees betwixt them), and tally it connected a illustration of your information to measure the results and the wide prime of matches. If it is unsatisfactory, you tin summation the threshold to marque it much strict for amended precision.

You tin instal ‘matplotlib’ via terminal.

And usage the Python codification beneath successful a abstracted Jupyter notebook to visualize cosine similarities successful two-dimensional abstraction connected your own. Try it; it’s fun!

import matplotlib.pyplot arsenic plt import numpy arsenic np # Define the space for cosine similarity of 0.9. Change present to your desired value. theta = np.arccos(0.9) # Define the vectors u = np.array([1, 0]) v = np.array([np.cos(theta), np.sin(theta)]) # Define the 45 grade rotation matrix rotation_matrix = np.array([ [np.cos(np.pi/4), -np.sin(np.pi/4)], [np.sin(np.pi/4), np.cos(np.pi/4)] ]) # Apply the rotation to some vectors u_rotated = np.dot(rotation_matrix, u) v_rotated = np.dot(rotation_matrix, v) # Plotting the vectors plt.figure() plt.quiver(0, 0, u_rotated[0], u_rotated[1], angles='xy', scale_units='xy', scale=1, color='r') plt.quiver(0, 0, v_rotated[0], v_rotated[1], angles='xy', scale_units='xy', scale=1, color='b') # Setting the crippled limits to lone affirmative ranges plt.xlim(0, 1.5) plt.ylim(0, 1.5) # Adding labels and grid plt.xlabel('X-axis') plt.ylabel('Y-axis') plt.grid(True) plt.title('Visualization of Vectors with Cosine Similarity of 0.9') # Show the plot plt.show()

I usually usage 0.9 and higher for identifying keyword cannibalization issues, but you whitethorn request to acceptable it to 0.5 erstwhile dealing with aged nonfiction redirects, arsenic aged articles whitethorn not person astir identical articles that are fresher but partially close.

It whitethorn besides beryllium amended to person the meta statement concatenated with the rubric successful lawsuit of redirects, successful summation to the title.

So, it depends connected the task you are performing. We volition reappraisal however to instrumentality redirects successful a abstracted nonfiction aboriginal successful this series.

Now, let’s reappraisal the results with the 3 models mentioned supra and spot however they were capable to place adjacent articles from our information illustration from Search Engine Journal’s articles.

Data Sample

From the list, we already spot that the 2nd and 4th articles screen the aforesaid taxable connected ‘meta tags.’ The articles successful the 5th and 7th rows are beauteous overmuch the aforesaid – discussing the value of H1 tags successful SEO – and tin beryllium merged.

The nonfiction successful the 3rd enactment doesn’t person immoderate similarities with immoderate of the articles successful the database but has communal words similar “Tag” oregon “SEO.”

The nonfiction successful the 6th enactment is again astir H1, but not precisely the aforesaid arsenic H1’s value to SEO. Instead, it represents Google’s sentiment connected whether they should match.

Articles connected the 8th and 9th rows are rather adjacent but inactive different; they tin beryllium combined.

text-embedding-ada-002

By utilizing ‘text-embedding-ada-002,’ we precisely recovered the 2nd and 4th articles with a cosine similarity of 0.92 and the 5th and 7th articles with a similarity of 0.91.

Screenshot from Jupyter log showing cosine similarities

And it generated output with grouped URLs by utilizing the aforesaid radical fig for akin articles. (colors are applied manually for visualization purposes).

Output expanse with grouped URLs

For the 2nd and 3rd articles, which person communal words “Tag” and “SEO” but are unrelated, the cosine similarity was 0.86. This shows wherefore a precocious similarity threshold of 0.9 oregon greater is necessary. If we acceptable it to 0.85, it would beryllium afloat of mendacious positives and could suggest merging unrelated articles.

text-embedding-3-small

By utilizing ‘text-embedding-3-small,’ rather surprisingly, it didn’t find immoderate matches per our similarity threshold of 0.9 oregon higher.

For the 2nd and 4th articles, cosine similarity was 0.76, and for the 5th and 7th articles, with similarity 0.77.

To amended recognize this exemplary done experimentation, I’ve added a somewhat modified mentation of the 1st enactment with ’15’ vs. ’14’ to the sample.

“14 Most Important Meta And HTML Tags You Need To Know For SEO”
“15 Most Important Meta And HTML Tags You Need To Know For SEO”

Example which shows text-embedding-3-small results An illustration which shows text-embedding-3-small results

On the contrary, ‘text-embedding-ada-002’ gave 0.98 cosine similarity betwixt those versions.

Title 1	Title 2	Cosine Similarity
14 Most Important Meta And HTML Tags You Need To Know For SEO	15 Most Important Meta And HTML Tags You Need To Know For SEO	0.92
14 Most Important Meta And HTML Tags You Need To Know For SEO	Meta Tags: What You Need To Know For SEO	0.76

Here, we spot that this exemplary is not rather a bully acceptable for comparing titles.

text-embedding-3-large

This model’s dimensionality is 3072, which is 2 times higher than that of ‘text-embedding-3-small’ and ‘text-embedding-ada-002′, with 1536 dimensionality.

As it has much dimensions than the different models, we could expect it to seizure semantic meaning with higher precision.

However, it gave the 2nd and 4th articles cosine similarity of 0.70 and the 5th and 7th articles similarity of 0.75.

I’ve tested it again with somewhat modified versions of the archetypal nonfiction with ’15’ vs. ’14’ and without ‘Most Important’ successful the title.

“14 Most Important Meta And HTML Tags You Need To Know For SEO”
“15 Most Important Meta And HTML Tags You Need To Know For SEO”
“14 Meta And HTML Tags You Need To Know For SEO”

Title 1	Title 2	Cosine Similarity
14 Most Important Meta And HTML Tags You Need To Know For SEO	15 Most Important Meta And HTML Tags You Need To Know For SEO	0.95
14 Most Important Meta And HTML Tags You Need To Know For SEO	14 Most Important Meta And HTML Tags You Need To Know For SEO	0.93
14 Most Important Meta And HTML Tags You Need To Know For SEO	Meta Tags: What You Need To Know For SEO	0.70
15 Most Important Meta And HTML Tags You Need To Know For SEO	14 Most Important Meta And HTML Tags You Need To Know For SEO	0.86

So we tin spot that ‘text-embedding-3-large’ is underperforming compared to ‘text-embedding-ada-002’ erstwhile we cipher cosine similarities betwixt titles.

I privation to enactment that the accuracy of ‘text-embedding-3-large’ increases with the magnitude of the text, but ‘text-embedding-ada-002’ inactive performs amended overall.

Another attack could beryllium to portion distant halt words from the text. Removing these tin sometimes assistance absorption the embeddings connected much meaningful words, perchance improving the accuracy of tasks similar similarity calculations.

The champion mode to find whether removing halt words improves accuracy for your circumstantial task and dataset is to empirically trial some approaches and comparison the results.

Conclusion

With these examples, you person learned however to enactment with OpenAI’s embedding models and tin already execute a wide scope of tasks.

For similarity thresholds, you request to experimentation with your ain datasets and spot which thresholds marque consciousness for your circumstantial task by moving it connected smaller samples of information and performing a quality reappraisal of the output.

Please enactment that the codification we person successful this nonfiction is not optimal for ample datasets since you request to make substance embeddings of articles each clip determination is simply a alteration successful your dataset to measure against different rows.

To marque it efficient, we indispensable usage vector databases and store embedding accusation determination erstwhile generated. We volition screen however to usage vector databases precise soon and alteration the codification illustration present to usage a vector database.

More resources:

Featured Image: BestForBest/Shutterstock