Run sentiment analysis and reasons in bulk with LLMs

In our last tutorial, we demonstrated a straightforward example of conducting sentiment analysis using LLM chains.

In this tutorial, we’re going to enhance this example by configuring multiple outputs to our chain and adding them back to our original spreadsheet as two new columns - we’re extracting the reason behind the given sentiment as well. You’ll learn how you can have a single LLM prompt output multiple pieces of information.

Extracting Sentiment and Reason Simultaneously

In this example we’re going to layer on multiple pieces of functionality within the same chain to give us the desired output.

So, we’ll go back to our extract sentiment analysis chain that we created in the previous example.

Modifying the LLM Prompt

In our initial chain, we’ve already received the input for the message we want the AI chain to examine.

To obtain the explanation for the sentiment score, we could either introduce a separate LLM prompt or simply modify our original prompt and then transform the output into a JSON. We’ll proceed with the latter.

To accomplish this, we adjust our prompt:

MESSAGE:

"""
{{message}}
"""

Based on the above message, respond with a JSON object of the format "{ sentiment, reason }". The sentiment of the message as one of "neutral", "positive" or 
"negative" and the reason as to why that sentiment is assigned. Only respond with valid JSON.

And set the system prompt to:

You are an expert sentiment classifier. Only ever respond with JSON in the format "{ sentiment, reason }". Do not say anything else.

Configuring Your LLM Output

After we’ve adjusted our prompt, we can proceed to copy one of the messages from our dataset that we want the chain to examine, and then hit run.

We can now observe that this has provided us with a JSON-like output that includes a reason and the sentiment.

Adding Columns for Sentiment and Reason in our original dataset

We aim to add a ‘sentiment’ column and a ‘reason’ column to our original dataset. To do this, we need to adjust our output.

We must properly convert our string to JSON. This can be accomplished by adding the ‘Convert string to JSON’ step to our chain, specifying the text field we wish to convert.

In this instance, it’s the output of our LLM prompt called {{llm.answer}} that now includes the instruction to perform sentiment analysis and provide a rationale for it.

Configure Chain Output

Now instead of just the LLM answer, we actually want to reference the JSON output for ‘sentiment’, and we’ll add another one for ‘reason’.

Run in bulk

Now if we hit run all again, we should see we have a sentiment label and reason.

If we bulk run the chain on our spreadsheet, the chain is going to add a column ‘reason’ and a column ‘sentiment’ to it.

Export the results

Once finished, you’ll be able to export the CSV as a link. The CSV will now feature a new column sentiment, and reasonwith the output from our chain.