Google Trends Dataset

[saurabh-khanna]

@Xiweeeen Just noting our next steps here. You can starting understanding the trending search results data from:

• any one day
• any one country
• type==organic search results for now

We will try to combine title and description into one variable, and then run BERTopic based topic modeling on this variable.

Feel free to post any notes/questions here!

[saurabh-khanna]

@Xiweeeen My code here uses Llama 3.2 model to predict categories. You can run this on the text column (text can be defined as search_query + title + description) in a small pandas dataframe first. Feel free to try different things.

I believe you will first need to:

• install aisuite: pip install 'aisuite[all]'
• set up ollama: https://ollama.com/download
• then run this in your terminal to download the Llama3.2 model: ollama pull llama3.2

This video might help if you are stuck. Happy coding!

[Xiweeeen]

@saurabh-khanna
I use the provided code to analyze the dataset. Well, the Llama model 3.2 is really cool. 3 things and opinions share with you:

1. Query + title + description – Hard to Distinct
   I combined the “query,” “title,” and “description” columns to create a new “text” column, but I found it difficult to distinguish unique entries even after using a “remove duplicates” function (drop duplicate function). The issue is that there are still more unique entries when considering these three columns together. However, if I focus only on the “query” column, it becomes much easier to count the number of unique queries, rank them, and distinguish them from each other. The downside is that focusing only on the “query” may result in losing some important context from the “title” and “description.” Therefore, I’m thinking about selecting only one “title” and “description” for each unique query to treat it as one topic. But I’m unsure if this approach would be effective in refining the data.

2. Using Different Characters or Half-Sentences Within Keywords – Same Results
   I experimented by using different characters or half sentences within the same keywords to see if the model would return different results. Interestingly, the LLaMA model still gave me the same answers, even with these variations. This suggests that the model isn’t sensitive to minor changes in keyword structure as long as the core terms are present.

3. Interesting Observation on Sentence Structure – Different Results
   II noticed an interesting behavior when using complete sentences as input. Despite using the same keywords with identical character and word counts, the model produced different results across attempts. Surprisingly, on my second try, the output aligned with my initial expectations, matching my intuitive impressions of what the result should be.

This makes me wonder if the model is influenced by latent factors beyond simple keyword matching. However, I also noticed that with other inputs, the model consistently returned the same results, even after multiple attempts. This inconsistency leaves me puzzled — why do some inputs produce variable outcomes while others remain consistent? Could the model be accounting for context or some hidden variables that aren’t immediately obvious?

[saurabh-khanna]

Next steps:

• Select a country to try on the full data: NL, US/UK, TW, HK
• Think carefully about improving the list of categories
• @saurabh-khanna to share starting code for generating embeddings

[Xiweeeen]

Hi @saurabh-khanna, I want to share three things with you.

1. The accuracy of my dataset, which focuses on Taiwanese individuals, is not as high as that of English datasets. I noticed that when I only use query titles and querry, the results differ from those obtained when using query + title + description. This shows that the more comprehensive the input information, the more accurate the results. However, even with more detailed inputs, the overall accuracy is still lower compared to English text datasets.

(capture the keywords not focus on understanding the text)

2. I think expanding categories can work really well by either combining existing themes or focusing on what’s trending. For example, the “Technology & Gadgets” category could include subcategories like “AI & Automation” or “Smartphones & Devices.” Similarly, “Entertainment” (like TV, movies, and music) could add areas such as “Anime & Manga,” “Streaming Services,” or even “Celebrity and Influencer News.”

In addition, It’s also super important to keep up with trending topics. For instance, health-related categories could feature things like “Mental Health,” “Fitness Trends,” or “Healthcare News.” And for social issues, we could include topics like “Climate Change,” “Economic Inequality,” or “Workplace Dynamics.” These additions would make the coverage broader and ensure it stays relevant and timely.

3. Lastly, I’m not sure why, but I can not apply the model to each text, yet I haven’t been able to make it work. It shows that I can’t access the API or something like that. I’ve already tried the three solutions provided by ChatGPT ((import httpx, import httpx), but none of them worked. 🥹 At first, I thought it was because I switched to Chinese, but I tried applying it to the notebook I previously used for validating English texts, and it still didn’t work. I’m not sure if I’m doing something wrong.

[saurabh-khanna]

Maybe we can try to do this in batches instead of apply:

texts = text_df['text'].tolist()
texts = texts[0]
responses = [classify_text(text) for text in texts]
text_df['category'] = responses

[saurabh-khanna]

Next steps:

• Check documentation of other models have better multilingual support than Llama3.2
• Check if accuracy for Chinese text is still low if we: i) translate it to english using easynmt, and ii) run the same model Llama3.2
• Check if we can ask the model in the prompt to create a new category
• Create embeddings using this approach

[Xiweeeen]

@saurabh-khanna

1. documentation

• A Survey on Large Language Models with Multilingualism: Recent Advances and New Frontiers
  https://arxiv.org/html/2405.10936v1?utm_source=chatgpt.com
  This article talks about the PaLM-2 model.

• Updated January 2025: a Comparative Analysis of Leading Large Language Models
  https://mindsdb.com/blog/navigating-the-llm-landscape-a-comparative-analysis-of-leading-large-language-models

• Top Large Language Models for Writers, Developers, and Marketers: A Comprehensive Comparison
  https://www.labellerr.com/blog/comparing-language-models-through-parameters-vs-real-life-experiments/

2. I'm still working on it... EasyNMT didn't work on my computer, but I'll keep trying later. 😭

3. The chatbox allows me to request new categories freely, with more extra options in categories and in an automatic way! It even suggests four potential category options in sequence and recommends the best one among them.

[Xiweeeen]

it works for easyNMT. I got the translation based on this model. But I find within the translation, the Ollama model can not really figure out the topic for the translating sentences...