Sentiment analysis with few-shot prompting

Guillaume Laforge
Google Cloud - Community
4 min readJul 30, 2024

In a recent article, we talked about text classification using Gemini and LangChain4j.

A typical example of text classification is the case of sentiment analysis.

In my LangChain4j-powered Gemini workshop, I used this use case to illustrate the classification problem:

ChatLanguageModel model = VertexAiGeminiChatModel.builder()
.project(System.getenv("PROJECT_ID"))
.location(System.getenv("LOCATION"))
.modelName("gemini-1.5-flash-001")
.maxOutputTokens(10)
.maxRetries(3)
.build();

PromptTemplate promptTemplate = PromptTemplate.from("""
Analyze the sentiment of the text below.
Respond only with one word to describe the sentiment.

INPUT: This is fantastic news!
OUTPUT: POSITIVE

INPUT: Pi is roughly equal to 3.14
OUTPUT: NEUTRAL

INPUT: I really disliked the pizza. Who would use pineapples as a pizza topping?
OUTPUT: NEGATIVE

INPUT: {{text}}
OUTPUT:
""");

Prompt prompt = promptTemplate.apply(
Map.of("text", "I love strawberries!"));

Response<AiMessage> response = model.generate(prompt.toUserMessage());

System.out.println(response.content().text());

I used a PromptTemplate to craft the prompt, with a {{text}} placeholder value to analyze the sentiment of that particular text.

Notice that I used the few-shot prompting technique, with example inputs and outputs.

Few-shot prompting with a list of messages

Somehow, I had the impression that this INPUT/OUTPUT notation was a bit of a hack to encourage the LLM to believe this is an actual exchange between the user and the AI.

I believed it would be cleaner to use a real list of messages that alternate user and AI messages. So I implemented this alternative approach, but haven’t yet committed it to my workshop repository.

Meanwhile, as I was chatting with my colleague Dan Dobrin, he pointed me at this very recent blog post from the LangChain people, who were investigating few-shot prompting to improve tool-calling performance.

What’s interesting in their analysis was that overall, on this anecdata example, it seems LLMs do better with real user/AI messages than with a big string of inputs/outputs.

Let’s see how to implement the same approach, with a real exchange of messages:

List<ChatMessage> fewShotPrompts = List.of(
SystemMessage.from("""
Analyze the sentiment of the text below.
Respond only with one word to describe the sentiment.
"""),

UserMessage.from("This is fantastic news!"),
AiMessage.from("POSITIVE"),

UserMessage.from("Pi is roughly equal to 3.14"),
AiMessage.from("NEUTRAL"),

UserMessage.from("I really disliked the pizza. " +
"Who would use pineapples as a pizza topping?"),
AiMessage.from("NEGATIVE"),

UserMessage.from("I love strawberries!")
);

response = model.generate(fewShotPrompts);

System.out.println(response.content().text());

This is not much more verbose than the previous approach, as it’s still very readable. And when pulling the few-shot data from an external database, it feels cleaner than concatenating a big string.

More type-safe few-shot prompting with messages and AiServices

To further improve on the list of messages tactic, we can use LangChain4j’s AiServices concept, which is a higher-level abstraction than using the model and prompt templates directly.

enum Sentiment { POSITIVE, NEUTRAL, NEGATIVE }

interface SentimentAnalysis {
@SystemMessage("""
Analyze the sentiment of the text below.
Respond only with one word to describe the sentiment.
""")
Sentiment analyze(String text);
}

MessageWindowChatMemory memory =
MessageWindowChatMemory.withMaxMessages(10);

memory.add(UserMessage.from("This is fantastic news!"));
memory.add(AiMessage.from(Sentiment.POSITIVE.name()));

memory.add(UserMessage.from("Pi is roughly equal to 3.14"));
memory.add(AiMessage.from(Sentiment.NEUTRAL.name()));

memory.add(UserMessage.from("I really disliked the pizza. " +
"Who would use pineapples as a pizza topping?"));
memory.add(AiMessage.from(Sentiment.NEGATIVE.name()));

SentimentAnalysis analyzer =
AiServices.builder(SentimentAnalysis.class)
.chatLanguageModel(model)
.chatMemory(memory)
.build();

System.out.println(analyzer.analyze("I love strawberries!"));

This third and final approach may be a bit more verbose, and introduces a few more LangChain4j concepts like system messages, chat memory, and the AI service itself, but it has the advantages of being:

  • more type-safe, as we’re using a Sentiment enum, which is easier to manipulate from code,
  • cleaner, because we’re using system instructions to instruct the model about what its job is.

We created:

  • a Java enum to represent the possible values of the sentiment,
  • a SentimentAnalysis interface with a clear signature: a text in input, a Sentiment enum value in output,
  • a @SystemMessage instruction to describe the analysis task,
  • a ChatMemory (here a MessageWindowChatMemory) to hold the few-shot examples.

Then we bind everything together, thanks to AiServices: the analysis interface that LangChain4j will implement for us, the language model, and the chat memory.

Finally, users just have to call the analyze() method, passing the text to analyze.

I also like the fact that we are coding against an interface, and potentially later on, developers could swap the implementation of the sentiment analyzer, and use a different approach.

Conclusion

All three approaches are valid: a big string, a low-level list of messages, or an AiServices abstraction. But I have a slight preference for the approach that is more type-safe and less stringy.

Just like LangChain4j provides a TextClassification class that leverages vector embeddings for text similarity, we could investigate whether it would make sense to also add a few-shot prompting classification solution directly in the LangChain4j project.

Originally published at https://glaforge.dev on July 30, 2024.

--

--

Guillaume Laforge
Google Cloud - Community

Developer Advocate for Google Cloud, Apache Groovy programming language co-founder, Java Champion