Monthly Archives: April 2024

Organizations across industries are using automatic text summarization to more efficiently handle vast amounts of information and make better decisions. In the financial sector, investment banks condense earnings reports down to key takeaways to rapidly analyze quarterly performance. Media companies use summarization to monitor news and social media so journalists can quickly write stories on developing issues. Government agencies summarize lengthy policy documents and reports to help policymakers strategize and prioritize goals.

By creating condensed versions of long, complex documents, summarization technology enables users to focus on the most salient content. This leads to better comprehension and retention of critical information. The time savings allow stakeholders to review more material in less time, gaining a broader perspective. With enhanced understanding and more synthesized insights, organizations can make better informed strategic decisions, accelerate research, improve productivity, and increase their impact. The transformative power of advanced summarization capabilities will only continue growing as more industries adopt artificial intelligence (AI) to harness overflowing information streams.

In this post, we explore leading approaches for evaluating summarization accuracy objectively, including ROUGE metrics, METEOR, and BERTScore. Understanding the strengths and weaknesses of these techniques can help guide selection and improvement efforts. The overall goal of this post is to demystify summarization evaluation to help teams better benchmark performance on this critical capability as they seek to maximize value.

Types of summarization

Summarization can generally be divided into two main types: extractive summarization and abstractive summarization. Both approaches aim to condense long pieces of text into shorter forms, capturing the most critical information or essence of the original content, but they do so in fundamentally different ways.

Extractive summarization involves identifying and extracting key phrases, sentences, or segments from the original text without altering them. The system selects parts of the text deemed most informative or representative of the whole. Extractive summarization is useful if accuracy is critical and the summary needs to reflect the exact information from the original text. These could be use cases like highlighting specific legal terms, obligations, and rights outlined in the terms of use. The most common techniques used for extractive summarization are term frequency-inverse document frequency (TF-IDF), sentence scoring, text rank algorithm, and supervised machine learning (ML).

Abstractive summarization goes a step further by generating new phrases and sentences that were not in the original text, essentially paraphrasing and condensing the original content. This approach requires a deeper understanding of the text, because the AI needs to interpret the meaning and then express it in a new, concise form. Large language models (LLMs) are best suited for abstractive summarization because the transformer models use attention mechanisms to focus on relevant parts of the input text when generating summaries. The attention mechanism allows the model to assign different weights to different words or tokens in the input sequence, enabling it to capture long-range dependencies and contextually relevant information.

In addition to these two primary types, there are hybrid approaches that combine extractive and abstractive methods. These approaches might start with extractive summarization to identify the most important content and then use abstractive techniques to rewrite or condense that content into a fluent summary.

The challenge

Finding the optimal method to evaluate summary quality remains an open challenge. As organizations increasingly rely on automatic text summarization to distill key information from documents, the need grows for standardized techniques to measure summarization accuracy. Ideally, these evaluation metrics would quantify how well machine-generated summaries extract the most salient content from source texts and present coherent summaries reflecting the original meaning and context.

However, developing robust evaluation methodologies for text summarization presents difficulties:

• Human-authored reference summaries used for comparison often exhibit high variability based on subjective determinations of importance
• Nuanced aspects of summary quality like fluency, readability, and coherence prove difficult to quantify programmatically
• Wide variation exists across summarization methods from statistical algorithms to neural networks, complicating direct comparisons

Recall-Oriented Understudy for Gisting Evaluation (ROUGE)

ROUGE metrics, such as ROUGE-N and ROUGE-L, play a crucial role in evaluating the quality of machine-generated summaries compared to human-written reference summaries. These metrics focus on assessing the overlap between the content of machine-generated and human-crafted summaries by analyzing n-grams, which are groups of words or tokens. For instance, ROUGE-1 evaluates the match of individual words (unigrams), whereas ROUGE-2 considers pairs of words (bigrams). Additionally, ROUGE-N assesses the longest common subsequence of words between the two texts, allowing for flexibility in word order.

To illustrate this, consider the following examples:

• ROGUE-1 metric – ROUGE-1 evaluates the overlap of unigrams (single words) between a generated summary and a reference summary. For example, if a reference summary contains “The quick brown fox jumps,” and the generated summary is “The brown fox jumps quickly,” the ROUGE-1 metric would consider “brown,” “fox,” and “jumps” as overlapping unigrams. ROUGE-1 focuses on the presence of individual words in the summaries, measuring how well the generated summary captures the key words from the reference summary.
• ROGUE-2 metric – ROUGE-2 assesses the overlap of bigrams (pairs of adjacent words) between a generated summary and a reference summary. For instance, if the reference summary has “The cat is sleeping,” and the generated summary reads “A cat is sleeping,” ROUGE-2 would identify “cat is” and “is sleeping” as an overlapping bigram. ROUGE-2 provides insight into how well the generated summary maintains the sequence and context of word pairs compared to the reference summary.
• ROUGE-N metric – ROUGE-N is a generalized form where N represents any number, allowing evaluation based on n-grams (sequences of N words). Considering N=3, if the reference summary states “The sun is shining brightly,” and the generated summary is “Sun shining brightly,” ROUGE-3 would recognize “sun shining brightly” as a matching trigram. ROUGE-N offers flexibility to evaluate summaries based on different lengths of word sequences, providing a more comprehensive assessment of content overlap.

These examples illustrate how ROUGE-1, ROUGE-2, and ROUGE-N metrics function in evaluating automatic summarization or machine translation tasks by comparing generated summaries with reference summaries based on different levels of word sequences.

Calculate a ROUGE-N score

You can use the following steps to calculate a ROUGE-N score:

1. Tokenize the generated summary and the reference summary into individual words or tokens using basic tokenization methods like splitting by whitespace or natural language processing (NLP) libraries.
2. Generate n-grams (contiguous sequences of N words) from both the generated summary and the reference summary.
3. Count the number of overlapping n-grams between the generated summary and the reference summary.
4. Calculate precision, recall, and F1 score:
   • Precision – The number of overlapping n-grams divided by the total number of n-grams in the generated summary.
   • Recall – The number of overlapping n-grams divided by the total number of n-grams in the reference summary.
   • F1 score – The harmonic mean of precision and recall, calculated as (2 * precision * recall) / (precision + recall).
5. The aggregate F1 score obtained from calculating precision, recall, and F1 score for each row in the dataset is considered as the ROUGE-N score.

Limitations

ROGUE has the following limitations:

• Narrow focus on lexical overlap – The core idea behind ROUGE is to compare the system-generated summary to a set of reference or human-created summaries, and measure the lexical overlap between them. This means ROUGE has a very narrow focus on word-level similarity. It doesn’t actually evaluate semantic meaning, coherence, or readability of the summary. A system could achieve high ROUGE scores by simply extracting sentences word-for-word from the original text, without generating a coherent or concise summary.
• Insensitivity to paraphrasing – Because ROUGE relies on lexical matching, it can’t detect semantic equivalence between words and phrases. Therefore, paraphrasing and use of synonyms will often lead to lower ROUGE scores, even if the meaning is preserved. This disadvantages systems that paraphrase or summarize in an abstractive way.
• Lack of semantic understanding – ROUGE doesn’t evaluate whether the system truly understood the meanings and concepts in the original text. A summary could achieve high lexical overlap with references, while missing the main ideas or containing factual inconsistencies. ROUGE would not identify these issues.

When to use ROUGE

ROUGE is simple and fast to calculate. Use it as a baseline or benchmark for summary quality related to content selection. ROUGE metrics are most effectively employed in scenarios involving abstractive summarization tasks, automatic summarization evaluation, assessments of LLMs, and comparative analyses of different summarization approaches. By using ROUGE metrics in these contexts, stakeholders can quantitatively evaluate the quality and effectiveness of summary generation processes.

Metric for Evaluation of Translation with Explicit Ordering (METEOR)

One of the major challenges in evaluating summarization systems is assessing how well the generated summary flows logically, rather than just selecting relevant words and phrases from the source text. Simply extracting relevant keywords and sentences doesn’t necessarily produce a coherent and cohesive summary. The summary should flow smoothly and connect ideas logically, even if they aren’t presented in the same order as the original document.

The flexibility of matching by reducing words to their root or base form (For example, after stemming, words like “running,” “runs,” and “ran” all become “run”) and synonyms means METEOR correlates better with human judgements of summary quality. It can identify if important content is preserved, even if the wording differs. This is a key advantage over n-gram based metrics like ROUGE, which only look for exact token matches. METEOR also gives higher scores to summaries that focus on the most salient content from the reference. Lower scores are given to repetitive or irrelevant information. This aligns well with the goal of summarization to keep the most important content only. METEOR is a semantically meaningful metric that can overcome some of the limitations of n-gram matching for evaluating text summarization. The incorporation of stemming and synonyms allows for better assessment of information overlap and content accuracy.

To illustrate this, consider the following examples:

Reference Summary: Leaves fall during autumn.

Generated Summary 1: Leaves drop in fall.

Generated Summary 2: Leaves green in summer.

The words that match between the reference and generated summary 1 are highlighted:

Reference Summary: Leavesfall during autumn.

Generated Summary 1: Leaves drop in fall.

Even though “fall” and “autumn” are different tokens, METEOR recognizes them as synonyms through its synonym matching. “Drop” and “fall” are identified as a stemmed match. For generated summary 2, there are no matches with the reference summary besides “Leaves,” so this summary would receive a much lower METEOR score. The more semantically meaningful matches, the higher the METEOR score. This allows METEOR to better evaluate the content and accuracy of summaries compared to simple n-gram matching.

Calculate a METEOR score

Complete the following steps to calculate a METEOR score:

1. Tokenize the generated summary and the reference summary into individual words or tokens using basic tokenization methods like splitting by whitespace or NLP libraries.
2. Calculate the unigram precision, recall, and F-mean score, giving more weightage to recall than precision.
3. Apply a penalty for exact matches to avoid overemphasizing them. The penalty is chosen based on dataset characteristics, task requirements, and the balance between precision and recall. Subtract this penalty from the F-mean score calculated in Step 2.
4. Calculate the F-mean score for stemmed forms (reducing words to their base or root form) and synonyms for unigrams where applicable. Aggregate this with the earlier calculated F-mean score to obtain the final METEOR score. The METEOR score ranges from 0–1, where 0 indicates no similarity between the generated summary and reference summary, and 1 indicates perfect alignment. Typically, summarization scores fall between 0–0.6.

Limitations

When employing the METEOR metric for evaluating summarization tasks, several challenges may arise:

• Semantic complexity – METEOR’s emphasis on semantic similarity can struggle to capture the nuanced meanings and context in complex summarization tasks, potentially leading to inaccuracies in evaluation.
• Reference variability – Variability in human-generated reference summaries can impact METEOR scores, because differences in reference content may affect the evaluation of machine-generated summaries.
• Linguistic diversity – The effectiveness of METEOR may vary across languages due to linguistic variations, syntax differences, and semantic nuances, posing challenges in multilingual summarization evaluations.
• Length discrepancy – Evaluating summaries of varying lengths can be challenging for METEOR, because discrepancies in length compared to the reference summary may result in penalties or inaccuracies in assessment.
• Parameter tuning – Optimizing METEOR’s parameters for different datasets and summarization tasks can be time-consuming and require careful tuning to make sure the metric provides accurate evaluations.
• Evaluation bias – There is a risk of evaluation bias with METEOR if not properly adjusted or calibrated for specific summarization domains or tasks. This can potentially lead to skewed results and affect the reliability of the evaluation process.

By being aware of these challenges and considering them when using METEOR as a metric for summarization tasks, researchers and practitioners can navigate potential limitations and make more informed decisions in their evaluation processes.

When to use METEOR

METEOR is commonly used to automatically evaluate the quality of text summaries. It is preferable to use METEOR as an evaluation metric when the order of ideas, concepts, or entities in the summary matters. METEOR considers the order and matches n-grams between the generated summary and reference summaries. It rewards summaries that preserve sequential information. Unlike metrics like ROUGE, which rely on overlap of n-grams with reference summaries, METEOR matches stems, synonyms, and paraphrases. METEOR works better when there can be multiple correct ways of summarizing the original text. METEOR incorporates WordNet synonyms and stemmed tokens when matching n-grams. In short, summaries that are semantically similar but use different words or phrasing will still score well. METEOR has a built-in penalty for summaries with repetitive n-grams. Therefore, it discourages word-for-word extraction or lack of abstraction. METEOR is a good choice when semantic similarity, order of ideas, and fluent phrasing are important for judging summary quality. It is less appropriate for tasks where only lexical overlap with reference summaries matters.

BERTScore

Surface-level lexical measures like ROUGE and METEOR evaluate summarization systems by comparing the word overlap between a candidate summary and a reference summary. However, they rely heavily on exact string matching between words and phrases. This means they may miss semantic similarities between words and phrases that have different surface forms but similar underlying meanings. By relying only on surface matching, these metrics may underestimate the quality of system summaries that use synonymous words or paraphrase concepts differently from reference summaries. Two summaries could convey nearly identical information but receive low surface-level scores due to vocabulary differences.

BERTScore is a way to automatically evaluate how good a summary is by comparing it to a reference summary written by a human. It uses BERT, a popular NLP technique, to understand the meaning and context of words in the candidate summary and reference summary. Specifically, it looks at each word or token in the candidate summary and finds the most similar word in the reference summary based on the BERT embeddings, which are vector representations of the meaning and context of each word. It measures the similarity using cosine similarity, which tells how close the vectors are to each other. For each word in the candidate summary, it finds the most related word in the reference summary using BERT’s understanding of language. It compares all these word similarities across the whole summary to get an overall score of how semantically similar the candidate summary is to the reference summary. The more similar the words and meanings captured by BERT, the higher the BERTScore. This allows it to automatically evaluate the quality of a generated summary by comparing it to a human reference without needing human evaluation each time.

To illustrate this, imagine you have a machine-generated summary: “The quick brown fox jumps over the lazy dog.” Now, let’s consider a human-crafted reference summary: “A fast brown fox leaps over a sleeping canine.”

Calculate a BERTScore

Complete the following steps to calculate a BERTScore:

1. BERTScore uses contextual embeddings to represent each token in both the candidate (machine-generated) and reference (human-crafted) sentences. Contextual embeddings are a type of word representation in NLP that captures the meaning of a word based on its context within a sentence or text. Unlike traditional word embeddings that assign a fixed vector to each word regardless of its context, contextual embeddings consider the surrounding words to generate a unique representation for each word depending on how it is used in a specific sentence.
2. The metric then computes the similarity between each token in the candidate sentence with each token in the reference sentence using cosine similarity. Cosine similarity helps us quantify how closely related two sets of data are by focusing on the direction they point in a multi-dimensional space, making it a valuable tool for tasks like search algorithms, NLP, and recommendation systems.
3. By comparing the contextual embeddings and computing similarity scores for all tokens, BERTScore generates a comprehensive evaluation that captures the semantic relevance and context of the generated summary compared to the human-crafted reference.
4. The final BERTScore output provides a similarity score that reflects how well the machine-generated summary aligns with the reference summary in terms of meaning and context.

In essence, BERTScore goes beyond traditional metrics by considering the semantic nuances and context of sentences, offering a more sophisticated evaluation that closely mirrors human judgment. This advanced approach enhances the accuracy and reliability of evaluating summarization tasks, making BERTScore a valuable tool in assessing text generation systems.

Limitations:

Although BERTScore offers significant advantages in evaluating summarization tasks, it also comes with certain limitations that need to be considered:

• Computational intensity – BERTScore can be computationally intensive due to its reliance on pre-trained language models like BERT. This can lead to longer evaluation times, especially when processing large volumes of text data.
• Dependency on pre-trained models – The effectiveness of BERTScore is highly dependent on the quality and relevance of the pre-trained language model used. In scenarios where the pre-trained model may not adequately capture the nuances of the text, the evaluation results may be affected.
• Scalability – Scaling BERTScore for large datasets or real-time applications can be challenging due to its computational demands. Implementing BERTScore in production environments may require optimization strategies to provide efficient performance.
• Domain specificity – BERTScore’s performance may vary across different domains or specialized text types. Adapting the metric to specific domains or tasks may require fine-tuning or adjustments to produce accurate evaluations.
• Interpretability – Although BERTScore provides a comprehensive evaluation based on contextual embeddings, interpreting the specific reasons behind the similarity scores generated for each token can be complex and may require additional analysis.
• Reference-free evaluation – Although BERTScore reduces the reliance on reference summaries for evaluation, this reference-free approach may not fully capture all aspects of summarization quality, particularly in scenarios where human-crafted references are essential for assessing content relevance and coherence.

Acknowledging these limitations can help you make informed decisions when using BERTScore as a metric for evaluating summarization tasks, providing a balanced understanding of its strengths and constraints.

When to use BERTScore

BERTScore can evaluate the quality of text summarization by comparing a generated summary to a reference summary. It uses neural networks like BERT to measure semantic similarity beyond just exact word or phrase matching. This makes BERTScore very useful when semantic fidelity preserving the full meaning and content is critical for your summarization task. BERTScore will give higher scores to summaries that convey the same information as the reference summary, even if they use different words and sentence structures. The bottom line is that BERTScore is ideal for summarization tasks where retaining the full semantic meaning not just keywords or topics is vital. Its advanced neural scoring allows it to compare meaning beyond surface-level word matching. This makes it suitable for cases where subtle differences in wording can substantially alter overall meaning and implications. BERTScore, in particular, excels in capturing semantic similarity, which is crucial for assessing the quality of abstractive summaries like those produced by Retrieval Augmented Generation (RAG) models.

Model evaluation frameworks

Model evaluation frameworks are essential for accurately gauging the performance of various summarization models. These frameworks are instrumental in comparing models, providing coherence between generated summaries and source content, and pinpointing deficiencies in evaluation methods. By conducting thorough assessments and consistent benchmarking, these frameworks propel text summarization research by advocating standardized evaluation practices and enabling multifaceted model comparisons.

In AWS, the FMEval library within Amazon SageMaker Clarify streamlines the evaluation and selection of foundation models (FMs) for tasks like text summarization, question answering, and classification. It empowers you to evaluate FMs based on metrics such as accuracy, robustness, creativity, bias, and toxicity, supporting both automated and human-in-the-loop evaluations for LLMs. With UI-based or programmatic evaluations, FMEval generates detailed reports with visualizations to quantify model risks like inaccuracies, toxicity, or bias, helping organizations align with their responsible generative AI guidelines. In this section, we demonstrate how to use the FMEval library.

Evaluate Claude v2 on summarization accuracy using Amazon Bedrock

The following code snippet is an example of how to interact with the Anthropic Claude model using Python code:

import json
# We use Claude v2 in this example.
# See https://docs.anthropic.com/claude/reference/claude-on-amazon-bedrock#list-available-models
# for instructions on how to list the model IDs for all available Claude model variants.
model_id = 'anthropic.claude-v2'
accept = "application/json"
contentType = "application/json"
# `prompt_data` is structured in the format that the Claude model expects, as documented here:
# https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-claude.html#model-parameters-claude-request-body
prompt_data = """Human: Who is Barack Obama?
Assistant:
"""
# For more details on parameters that can be included in `body` (such as "max_tokens_to_sample"),
# see https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-claude.html#model-parameters-claude-request-body
body = json.dumps({"prompt": prompt_data, "max_tokens_to_sample": 500})
# Invoke the model
response = bedrock_runtime.invoke_model(
body=body, modelId=model_id, accept=accept, contentType=contentType
)
# Parse the invocation response
response_body = json.loads(response.get("body").read())
print(response_body.get("completion"))

In simple terms, this code performs the following actions:

1. Import the necessary libraries, including json, to work with JSON data.
2. Define the model ID as anthropic.claude-v2 and set the content type for the request.
3. Create a prompt_data variable that structures the input data for the Claude model. In this case, it asks the question “Who is Barack Obama?” and expects a response from the model.
4. Construct a JSON object named body that includes the prompt data, and specify additional parameters like the maximum number of tokens to generate.
5. Invoke the Claude model using bedrock_runtime.invoke_model with the defined parameters.
6. Parse the response from the model, extract the completion (generated text), and print it out.

Make sure the AWS Identity and Access Management (IAM) role associated with the Amazon SageMaker Studio user profile has access to the Amazon Bedrock models being invoked. Refer to Identity-based policy examples for Amazon Bedrock for guidance on best practices and examples of identity-based policies for Amazon Bedrock.

Using the FMEval library to evaluate the summarized output from Claude

We use the following code to evaluate the summarized output:

from fmeval.data_loaders.data_config import DataConfig
from fmeval.model_runners.bedrock_model_runner import BedrockModelRunner
from fmeval.constants import MIME_TYPE_JSONLINES
from fmeval.eval_algorithms.summarization_accuracy import SummarizationAccuracy
config = DataConfig(
    dataset_name="gigaword_sample",
    dataset_uri="gigaword_sample.jsonl",
    dataset_mime_type=MIME_TYPE_JSONLINES,
    model_input_location="document",
    target_output_location="summary"
)
bedrock_model_runner = BedrockModelRunner(
    model_id=model_id,
    output='completion',
    content_template='{"prompt": $prompt, "max_tokens_to_sample": 500}'
)
eval_algo = SummarizationAccuracy()
eval_output = eval_algo.evaluate(model=bedrock_model_runner, dataset_config=config,
prompt_template="Human: Summarise the following text in one sentence: $featurennAssistant:n", save=True)

In the preceding code snippet, to evaluate text summarization using the FMEval library, we complete the following steps:

1. Create a ModelRunner to perform invocation on your LLM. The FMEval library provides built-in support for Amazon SageMaker endpoints and Amazon SageMaker JumpStart LLMs. You can also extend the ModelRunner interface for any LLMs hosted anywhere.
2. Use supported eval_algorithms like toxicity, summarization, accuracy, semantic, and robustness, based on your evaluation needs.
3. Customize the evaluation configuration parameters for your specific use case.
4. Use the evaluation algorithm with either built-in or custom datasets to evaluate your LLM model. The dataset used in this case is sourced from the following GitHub repo.

Refer to the developer guide and examples for detailed usage of evaluation algorithms.

The following table summarizes the results of the evaluation.

Check out the sample notebook for more details about the summarization evaluation that we discussed in this post.

Conclusion

ROUGE, METEOR, and BERTScore all measure the quality of machine-generated summaries, but focus on different aspects like lexical overlap, fluency, or semantic similarity. Make sure to select the metric that aligns with what defines “good” for your specific summarization use case. You can also use a combination of metrics. This provides a more well-rounded evaluation and guards against potential weaknesses of any individual metric. With the right measurements, you can iteratively improve your summarizers to meet whichever notion of accuracy matters most.

Additionally, FM and LLM evaluation is necessary to be able to productionize these models at scale. With FMEval, you get a vast set of built-in algorithms across many NLP tasks, but also a scalable and flexible tool for large-scale evaluations of your own models, datasets, and algorithms. To scale up, you can use this package in your LLMOps pipelines to evaluate multiple models. To learn more about FMEval in AWS and how to use it effectively, refer to Use SageMaker Clarify to evaluate large language models. For further understanding and insights into the capabilities of SageMaker Clarify in evaluating FMs, see Amazon SageMaker Clarify Makes It Easier to Evaluate and Select Foundation Models.

About the Authors

Dinesh Kumar Subramani is a Senior Solutions Architect based in Edinburgh, Scotland. He specializes in artificial intelligence and machine learning, and is member of technical field community with in Amazon. Dinesh works closely with UK Central Government customers to solve their problems using AWS services. Outside of work, Dinesh enjoys spending quality time with his family, playing chess, and exploring a diverse range of music.

Pranav Sharma is an AWS leader driving technology and business transformation initiatives across Europe, the Middle East, and Africa. He has experience in designing and running artificial intelligence platforms in production that support millions of customers and deliver business outcomes. He has played technology and people leadership roles for Global Financial Services organizations. Outside of work, he likes to read, play tennis with his son, and watch movies.

Read More

Conversational artificial intelligence (AI) assistants are engineered to provide precise, real-time responses through intelligent routing of queries to the most suitable AI functions. With AWS generative AI services like Amazon Bedrock, developers can create systems that expertly manage and respond to user requests. Amazon Bedrock is a fully managed service that offers a choice of high-performing foundation models (FMs) from leading AI companies like AI21 Labs, Anthropic, Cohere, Meta, Stability AI, and Amazon using a single API, along with a broad set of capabilities you need to build generative AI applications with security, privacy, and responsible AI.

This post assesses two primary approaches for developing AI assistants: using managed services such as Agents for Amazon Bedrock, and employing open source technologies like LangChain. We explore the advantages and challenges of each, so you can choose the most suitable path for your needs.

What is an AI assistant?

An AI assistant is an intelligent system that understands natural language queries and interacts with various tools, data sources, and APIs to perform tasks or retrieve information on behalf of the user. Effective AI assistants possess the following key capabilities:

• Natural language processing (NLP) and conversational flow
• Knowledge base integration and semantic searches to understand and retrieve relevant information based on the nuances of conversation context
• Running tasks, such as database queries and custom AWS Lambda functions
• Handling specialized conversations and user requests

We demonstrate the benefits of AI assistants using Internet of Things (IoT) device management as an example. In this use case, AI can help technicians manage machinery efficiently with commands that fetch data or automate tasks, streamlining operations in manufacturing.

Agents for Amazon Bedrock approach

Agents for Amazon Bedrock allows you to build generative AI applications that can run multi-step tasks across a company’s systems and data sources. It offers the following key capabilities:

• Automatic prompt creation from instructions, API details, and data source information, saving weeks of prompt engineering effort
• Retrieval Augmented Generation (RAG) to securely connect agents to a company’s data sources and provide relevant responses
• Orchestration and running of multi-step tasks by breaking down requests into logical sequences and calling necessary APIs
• Visibility into the agent’s reasoning through a chain-of-thought (CoT) trace, allowing troubleshooting and steering of model behavior
• Prompt engineering abilities to modify the automatically generated prompt template for enhanced control over agents

You can use Agents for Amazon Bedrock and Knowledge Bases for Amazon Bedrock to build and deploy AI assistants for complex routing use cases. They provide a strategic advantage for developers and organizations by simplifying infrastructure management, enhancing scalability, improving security, and reducing undifferentiated heavy lifting. They also allow for simpler application layer code because the routing logic, vectorization, and memory is fully managed.

Solution overview

This solution introduces a conversational AI assistant tailored for IoT device management and operations when using Anthropic’s Claude v2.1 on Amazon Bedrock. The AI assistant’s core functionality is governed by a comprehensive set of instructions, known as a system prompt, which delineates its capabilities and areas of expertise. This guidance makes sure the AI assistant can handle a wide range of tasks, from managing device information to running operational commands.

"""The following is the system prompt that outlines the full scope of the AI assistant's capabilities:
You are an IoT Ops agent that handles the following activities:
- Looking up IoT device information
- Checking IoT operating metrics (historical data)
- Performing actions on a device-by-device ID
- Answering general questions
You can check device information (Device ID, Features, Technical Specifications, Installation Guide, Maintenance and Troubleshooting, Safety Guidelines, Warranty, and Support) from the "IotDeviceSpecs" knowledge base.
Additionally, you can access device historical data or device metrics. The device metrics are stored in an Athena DB named "iot_ops_glue_db" in a table named "iot_device_metrics". 
The table schema includes fields for oil level, temperature, pressure, received_at timestamp, and device_id.
The available actions you can perform on the devices include start, shutdown, and reboot."""

Equipped with these capabilities, as detailed in the system prompt, the AI assistant follows a structured workflow to address user questions. The following figure provides a visual representation of this workflow, illustrating each step from initial user interaction to the final response.

The workflow is composed of the following steps:

1. The process begins when a user requests the assistant to perform a task; for example, asking for the maximum data points for a specific IoT device device_xxx. This text input is captured and sent to the AI assistant.
2. The AI assistant interprets the user’s text input. It uses the provided conversation history, action groups, and knowledge bases to understand the context and determine the necessary tasks.
3. After the user’s intent is parsed and understood, the AI assistant defines tasks. This is based on the instructions that are interpreted by the assistant as per the system prompt and user’s input.
4. The tasks are then run through a series of API calls. This is done using ReAct prompting, which breaks down the task into a series of steps that are processed sequentially:
   1. For device metrics checks, we use the check-device-metrics action group, which involves an API call to Lambda functions that then query Amazon Athena for the requested data.
   2. For direct device actions like start, stop, or reboot, we use the action-on-device action group, which invokes a Lambda function. This function initiates a process that sends commands to the IoT device. For this post, the Lambda function sends notifications using Amazon Simple Email Service (Amazon SES).
   3. We use Knowledge Bases for Amazon Bedrock to fetch from historical data stored as embeddings in the Amazon OpenSearch Service vector database.
5. After the tasks are complete, the final response is generated by the Amazon Bedrock FM and conveyed back to the user.
6. Agents for Amazon Bedrock automatically stores information using a stateful session to maintain the same conversation. The state is deleted after a configurable idle timeout elapses.

Technical overview

The following diagram illustrates the architecture to deploy an AI assistant with Agents for Amazon Bedrock.

It consists of the following key components:

• Conversational interface – The conversational interface uses Streamlit, an open source Python library that simplifies the creation of custom, visually appealing web apps for machine learning (ML) and data science. It is hosted on Amazon Elastic Container Service (Amazon ECS) with AWS Fargate, and it is accessed using an Application Load Balancer. You can use Fargate with Amazon ECS to run containers without having to manage servers, clusters, or virtual machines.
• Agents for Amazon Bedrock – Agents for Amazon Bedrock completes the user queries through a series of reasoning steps and corresponding actions based on ReAct prompting:
  • Knowledge Bases for Amazon Bedrock – Knowledge Bases for Amazon Bedrock provides fully managed RAG to supply the AI assistant with access to your data. In our use case, we uploaded device specifications into an Amazon Simple Storage Service (Amazon S3) bucket. It serves as the data source to the knowledge base.
  • Action groups – These are defined API schemas that invoke specific Lambda functions to interact with IoT devices and other AWS services.
  • Anthropic Claude v2.1 on Amazon Bedrock – This model interprets user queries and orchestrates the flow of tasks.
  • Amazon Titan Embeddings – This model serves as a text embeddings model, transforming natural language text—from single words to complex documents—into numerical vectors. This enables vector search capabilities, allowing the system to semantically match user queries with the most relevant knowledge base entries for effective search.

The solution is integrated with AWS services such as Lambda for running code in response to API calls, Athena for querying datasets, OpenSearch Service for searching through knowledge bases, and Amazon S3 for storage. These services work together to provide a seamless experience for IoT device operations management through natural language commands.

Benefits

This solution offers the following benefits:

• Implementation complexity:
  • Fewer lines of code are required, because Agents for Amazon Bedrock abstracts away much of the underlying complexity, reducing development effort
  • Managing vector databases like OpenSearch Service is simplified, because Knowledge Bases for Amazon Bedrock handles vectorization and storage
  • Integration with various AWS services is more streamlined through pre-defined action groups
• Developer experience:
  • The Amazon Bedrock console provides a user-friendly interface for prompt development, testing, and root cause analysis (RCA), enhancing the overall developer experience
• Agility and flexibility:
  • Agents for Amazon Bedrock allows for seamless upgrades to newer FMs (such as Claude 3.0) when they become available, so your solution stays up to date with the latest advancements
  • Service quotas and limitations are managed by AWS, reducing the overhead of monitoring and scaling infrastructure
• Security:
  • Amazon Bedrock is a fully managed service, adhering to AWS’s stringent security and compliance standards, potentially simplifying organizational security reviews

Although Agents for Amazon Bedrock offers a streamlined and managed solution for building conversational AI applications, some organizations may prefer an open source approach. In such cases, you can use frameworks like LangChain, which we discuss in the next section.

LangChain dynamic routing approach

LangChain is an open source framework that simplifies building conversational AI by allowing the integration of large language models (LLMs) and dynamic routing capabilities. With LangChain Expression Language (LCEL), developers can define the routing, which allows you to create non-deterministic chains where the output of a previous step defines the next step. Routing helps provide structure and consistency in interactions with LLMs.

For this post, we use the same example as the AI assistant for IoT device management. However, the main difference is that we need to handle the system prompts separately and treat each chain as a separate entity. The routing chain decides the destination chain based on the user’s input. The decision is made with the support of an LLM by passing the system prompt, chat history, and user’s question.

Solution overview

The following diagram illustrates the dynamic routing solution workflow.

The workflow consists of the following steps:

1. The user presents a question to the AI assistant. For example, “What are the max metrics for device 1009?”
2. An LLM evaluates each question along with the chat history from the same session to determine its nature and which subject area it falls under (such as SQL, action, search, or SME). The LLM classifies the input and the LCEL routing chain takes that input.
3. The router chain selects the destination chain based on the input, and the LLM is provided with the following system prompt:

"""Given the user question below, classify it as one of the candidate prompts. You may want to modify the input considering the chat history and the context of the question. 
Sometimes the user may just assume that you have the context of the conversation and may not provide a clear input. Hence, you are being provided with the chat history for more context. 
Respond with only a Markdown code snippet containing a JSON object formatted EXACTLY as specified below. 
Do not provide an explanation to your classification beside the Markdown, I just need to know your decision on which destination and next_inputs
<candidate prompt>
physics: Good for answering questions about physics
sql: sql: Good for querying sql from AWS Athena. User input may look like: get me max or min for device x?
lambdachain: Good to execute actions with Amazon Lambda like shutting down a device or turning off an engine User input can be like, shutdown device x, or terminate process y, etc.
rag: Good to search knowledgebase and retrieve information about devices and other related information. User question can be like: what do you know about device x?
default: if the input is not well suited for any of the candidate prompts above. this could be used to carry on the conversation and respond to queries like provide a summary of the conversation
</candidate prompt>"""

The LLM evaluates the user’s question along with the chat history to determine the nature of the query and which subject area it falls under. The LLM then classifies the input and outputs a JSON response in the following format:

<Markdown>
```json
{{
"destination": string  name of the prompt to use
"next_inputs": string  a potentially modified version of the original input
}}
```

The router chain uses this JSON response to invoke the corresponding destination chain. There are four subject-specific destination chains, each with its own system prompt:

1. SQL-related queries are sent to the SQL destination chain for database interactions. You can use LCEL to build the SQL chain.
2. Action-oriented questions invoke the custom Lambda destination chain for running operations. With LCEL, you can define your own custom function; in our case, it’s a function to run a predefined Lambda function to send an email with a device ID parsed. Example user input might be “Shut down device 1009.”
3. Search-focused inquiries proceed to the RAG destination chain for information retrieval.
4. SME-related questions go to the SME/expert destination chain for specialized insights.
5. Each destination chain takes the input and runs the necessary models or functions:
   1. The SQL chain uses Athena for running queries.
   2. The RAG chain uses OpenSearch Service for semantic search.
   3. The custom Lambda chain runs Lambda functions for actions.
   4. The SME/expert chain provides insights using the Amazon Bedrock model.
6. Responses from each destination chain are formulated into coherent insights by the LLM. These insights are then delivered to the user, completing the query cycle.
7. User input and responses are stored in Amazon DynamoDB to provide context to the LLM for the current session and from past interactions. The duration of persisted information in DynamoDB is controlled by the application.

Technical overview

The following diagram illustrates the architecture of the LangChain dynamic routing solution.

The web application is built on Streamlit hosted on Amazon ECS with Fargate, and it is accessed using an Application Load Balancer. We use Anthropic’s Claude v2.1 on Amazon Bedrock as our LLM. The web application interacts with the model using LangChain libraries. It also interacts with variety of other AWS services, such as OpenSearch Service, Athena, and DynamoDB to fulfill end-users’ needs.

Benefits

This solution offers the following benefits:

• Implementation complexity:
  • Although it requires more code and custom development, LangChain provides greater flexibility and control over the routing logic and integration with various components.
  • Managing vector databases like OpenSearch Service requires additional setup and configuration efforts. The vectorization process is implemented in code.
  • Integrating with AWS services may involve more custom code and configuration.
• Developer experience:
  • LangChain’s Python-based approach and extensive documentation can be appealing to developers already familiar with Python and open source tools.
  • Prompt development and debugging may require more manual effort compared to using the Amazon Bedrock console.
• Agility and flexibility:
  • LangChain supports a wide range of LLMs, allowing you to switch between different models or providers, fostering flexibility.
  • The open source nature of LangChain enables community-driven improvements and customizations.
• Security:
  • As an open source framework, LangChain may require more rigorous security reviews and vetting within organizations, potentially adding overhead.

Conclusion

Conversational AI assistants are transformative tools for streamlining operations and enhancing user experiences. This post explored two powerful approaches using AWS services: the managed Agents for Amazon Bedrock and the flexible, open source LangChain dynamic routing. The choice between these approaches hinges on your organization’s requirements, development preferences, and desired level of customization. Regardless of the path taken, AWS empowers you to create intelligent AI assistants that revolutionize business and customer interactions

Find the solution code and deployment assets in our GitHub repository, where you can follow the detailed steps for each conversational AI approach.

About the Authors

Ameer Hakme is an AWS Solutions Architect based in Pennsylvania. He collaborates with Independent Software Vendors (ISVs) in the Northeast region, assisting them in designing and building scalable and modern platforms on the AWS Cloud. An expert in AI/ML and generative AI, Ameer helps customers unlock the potential of these cutting-edge technologies. In his leisure time, he enjoys riding his motorcycle and spending quality time with his family.

Sharon Li is an AI/ML Solutions Architect at Amazon Web Services based in Boston, with a passion for designing and building Generative AI applications on AWS. She collaborates with customers to leverage AWS AI/ML services for innovative solutions.

Kawsar Kamal is a senior solutions architect at Amazon Web Services with over 15 years of experience in the infrastructure automation and security space. He helps clients design and build scalable DevSecOps and AI/ML solutions in the Cloud.

Read More

The Amazon EU Design and Construction (Amazon D&C) team is the engineering team designing and constructing Amazon warehouses. The team navigates a large volume of documents and locates the right information to make sure the warehouse design meets the highest standards. In the post A generative AI-powered solution on Amazon SageMaker to help Amazon EU Design and Construction, we presented a question answering bot solution using a Retrieval Augmented Generation (RAG) pipeline with a fine-tuned large language model (LLM) for Amazon D&C to efficiently retrieve accurate information from a large volume of unorganized documents, and provide timely and high-quality services in their construction projects. The Amazon D&C team implemented the solution in a pilot for Amazon engineers and collected user feedback.

In this post, we share how we analyzed the feedback data and identified limitations of accuracy and hallucinations RAG provided, and used the human evaluation score to train the model through reinforcement learning. To increase training samples for better learning, we also used another LLM to generate feedback scores. This method addressed the RAG limitation and further improved the bot response quality. We present the reinforcement learning process and the benchmarking results to demonstrate the LLM performance improvement. The solution uses Amazon SageMaker JumpStart as the core service for model deployment, fine-tuning, and reinforcement learning.

Collect feedback from Amazon engineers in a pilot project

After developing the solution described in A generative AI-powered solution on Amazon SageMaker to help Amazon EU Design and Construction, the Amazon D&C team deployed the solution and ran a pilot project with Amazon engineers. The engineers accessed the pilot system through a web application developed by Streamlit, connected with the RAG pipeline. In the pipeline, we used Amazon OpenSearch Service for the vector database, and deployed a fine-tuned Mistral-7B-Instruct model on Amazon SageMaker.

One of the key objectives of the pilot is to collect feedback from Amazon engineers and use the feedback to further reduce LLM hallucinations. To achieve this, we developed a feedback collection module in the UI, as shown in the following figure, and stored the web session information and user feedback in Amazon DynamoDB. Through the feedback collection UI, Amazon engineers can select from five satisfaction levels: strongly disagree, disagree, neutral, agree, and strongly agree, corresponding to the feedback scores from 1–5. They can also provide a better answer to the question or comment on why the LLM response is not satisfactory.

During the pilot, users provided 118 feedback responses. The following chart shows the distribution from negative to positive. We observed the incorrect answers marked by the negative feedback (disagree and strongly disagree) are 45% among the total feedback (53 out of 118). Some of the errors (about 10% of negative feedback and 7.5% of overall responses) can be addressed by user education and prompt engineering. Additionally, we can address the issue with the solution of LLM fine-tuning and reinforcement learning, described in the next section.

Some users simply left a note, such as “Great!” for the strongly agree answers, or “Doesn’t answer the question,” for the strongly disagree answers. Other users provided scores and explained how they justify the LLM answers in their notes. In the following example, the user specifically provided the adequate document and content to correct the LLM hallucination.

Improve bot response with supervised fine-tuning and reinforcement learning

The solution consists of three steps of fine-tuning:

1. Conduct supervised fine-tuning using labeled data. This method was described in A generative AI-powered solution on Amazon SageMaker to help Amazon EU Design and Construction.
2. Collect user feedback to label the question-answer pairs for further LLM tuning.
3. When the training data is ready, further tune the model using reinforcement learning from human feedback (RLHF).

RLHF is widely used throughout generative artificial intelligence (AI) and LLM applications. It incorporates human feedback in the rewards function and trains the model with a reinforcement learning algorithm to maximize rewards, which makes the model perform tasks more aligned with human goals. The following diagram shows the pipeline of the steps.

We tested the methodology using the Amazon D&C documents with a Mistral-7B model on SageMaker JumpStart.

Supervised fine-tuning

In the previous post, we demonstrated how the fine-tuned Falcon-7B model outperforms the RAG pipeline and improves the quality and accuracy of QA bot response. For this post, we performed supervised fine-tuning on the Mistral-7B model. The supervised fine-tuning used the PEFT/LoRA technique (LoRA_r = 512, LoRA_alpha = 1024) on 436,207,616 parameters (5.68% of the total 7,677,964,288 parameters). The training was conducted on a p3.8x node with 137 samples synthetically generated by LLM and validated by humans; the process is well converged after 20 epochs, as shown in the following figure.

The fine-tuned model was validated by 274 samples, and the inference results were compared with the reference answers by the semantic similarity score. The score is 0.8100, which is higher than the score of 0.6419 from the traditional RAG.

Collect human and AI feedback for reinforcement learning

For RLHF, a sufficient amount of high-quality training samples labeled by subject matter experts (SMEs) are essential. However, poor-quality human labels will likely cause worse model performance than the original model after RLHF training. SMEs’ time is a scarce resource in any organization; reviewing hundreds or thousands of LLM responses and providing feedback requires a significant time investment from SMEs that may not have a clear return on investment.

To address this challenge, we adopted the reinforcement learning from AI feedback (RLAIF) method. RLAIF employs an AI assistant (another LLM) to provide evaluation scores, rather than from humans. In this hybrid learning approach, the learning agent refines the actions not only based on the interaction with a human but also from feedback provided by another AI model. It is much more scalable to provide sufficient training data for reinforcement learning, and much less subjective because the training is not dependent on particular preferences from a small group of SMEs. The SMEs may still be involved, but only to supervise the process and examine the AI feedback quality. This tremendously reduces the SMEs’ workload, because the AI-generated scores and labels are much less complicated for SMEs to analyze, filter, and sort.

In this use case, we provide the following prompt to an Anthropic Claude 2 model to generate the evaluation scores as the training samples for the reinforcement learning step:

"""

Human:

I am going to provide you with a question, a reference answer, and an machine generated answer to the question.

question: {question}

reference answer: {reference_answer}

machine_answer: {machine_answer}

Please analyze the question, the reference answer, and the machine generated answer to determine if the machine generated answer semantically matches the reference answer.

First think through your answer inside of <thinking> tags, then assign a score between 0.0 and 5.0 based on how well the machine answer matches the reference answer. Score 5.0 means a completely match, score 0.0 means a completely mismatch. Answer the score inside of <scoring> tags.

Also provide the reason to give this score inside of <reason> tags.

Assistant: <thinking>

"""

The Claude 2 model generates an evaluation score and a comment on why the score is given, as shown in the following example.

Out of the 274 validation questions, the supervised fine-tuned model generated 159 responses that have AI scores greater than 4. We observed 60 answers with scores lower than 3; there is space to improve the overall response quality.

The Amazon Engineering SMEs validated this AI feedback and acknowledged the benefits of using AI scores. Without AI feedback, the SMEs would need some time to review and analyze each LLM response to identify the cut-off answers and hallucinations, and to judge whether the LLM is returning correct contents and key concepts. AI feedback provides AI scores automatically and enables the SMEs to use filtering, sorting, and grouping to validate the scores and identify trends in the responses. This reduces the average SME’s review time by 80%.

Reinforcement learning from human and AI feedback

When the training samples are ready, we use the proximal policy optimization (PPO) algorithm to perform reinforcement learning. PPO uses a policy gradient method, which takes small steps to update the policy in the learning process, so that the learning agents can reliably reach the optimal policy network. This makes the training process more stable and reduces the possibility of divergence.

During the training, first we use the human- and AI-labeled data to build a reward model, which will be used guide the weights update in the learning process. For this use case, we select a distilroberta-base reward model and train it by samples in the following format:

[Instruction, Chosen_response, Rejected_response]

The following is an example of a training record.

The reward model is trained with the learning rate 1e-5. As shown in the following chart, the training well converges after 10 epochs.

Then the reward model is used for reinforcement learning. For this use case, we use PEFT/LoRA to accelerate the learning. The LoRA_r is 32 and LoRA_alpha is 64. The PPO-based learning is on a rate of 1.41e-5 in three epochs. Reinforcement training is highly volatile and very sensitive to the label data and feedback data. We use the reward_mean metric, the average reward obtained from the environment by specifically monitoring the reward model, to observe the learning process.

After RLAIF, we can see the number of bot responses with positive scores (above 3) increase from 78.1% to 85.5%, and those with negative scores (below 3) decrease from 21.9% to 14.5%. In total, the average AI feedback score increases up to 8%, from 3.9 to 4.2.

You can see the improvement in the following example.

Conclusion

This post is a continuation of our work with the Amazon EU Design and Construction team as discussed in A generative AI-powered solution on Amazon SageMaker to help Amazon EU Design and Construction. In this post, we showed how we generated human and AI feedback data to fine-tune the Mistral-7B model with reinforcement learning. The model after RLAIF provided better performance for Amazon Engineering’s question answering bot, improved the AI feedback score by 8%. In the Amazon D&C team’s pilot project, using RLAIF reduced the validation workload for SMEs by an estimated 80%. As the next step, we will scale up this solution by connecting with Amazon Engineering’s data infrastructure, and design a framework to automate the continuous learning process with a human in the loop. We will also further improve the AI feedback quality by tuning the prompt template.

Through this process, we learned how to further improve the quality and performance of question answering tasks through RLHF and RLAIF.

• Human validation and augmentation are essential to provide accurate and responsible outputs from LLM. The human feedback can be used in RLHF to further improve the model response.
• RLAIF automates the evaluation and learning cycle. The AI-generated feedback is less subjective because it doesn’t depend on a particular preference from a small pool of SMEs.
• RLAIF is more scalable to improve the bot quality through continued reinforcement learning while minimizing the efforts required from SMEs. It is especially useful for developing domain-specific generative AI solutions within large organizations.
• This process should be done on a regular basis, especially when new domain data is available to be covered by the solution.

In this use case, we used SageMaker JumpStart to test multiple LLMs and experiment with multiple LLM training approaches. It significantly accelerates the AI feedback and learning cycle with maximized efficiency and quality. For your own project, you can introduce the human-in-the-loop approach to collect your users’ feedback, or generate AI feedback using another LLM. Then you can follow the three-step process defined in this post to fine-tune your models using RLHF and RLAIF. We recommend experimenting with the methods using SageMaker JumpStart to speed up the process.

About the Author

Yunfei Bai is a Senior Solutions Architect at AWS. With a background in AI/ML, data science, and analytics, Yunfei helps customers adopt AWS services to deliver business results. He designs AI/ML and data analytics solutions that overcome complex technical challenges and drive strategic objectives. Yunfei has a PhD in Electronic and Electrical Engineering. Outside of work, Yunfei enjoys reading and music.

Elad Dwek is a Construction Technology Manager at Amazon. With a background in construction and project management, Elad helps teams adopt new technologies and data-based processes to deliver construction projects. He identifies needs and solutions, and facilitates the development of the bespoke attributes. Elad has an MBA and a BSc in Structural Engineering. Outside of work, Elad enjoys yoga, woodworking, and traveling with his family.

Luca Cerabone is a Business Intelligence Engineer at Amazon. Drawing from his background in data science and analytics, Luca crafts tailored technical solutions to meet the unique needs of his customers, driving them towards more sustainable and scalable processes. Armed with an MSc in Data Science, Luca enjoys engaging in DIY projects, gardening and experimenting with culinary delights in his leisure moments.

Read More

In various industries, such as financial services, telecommunications, and healthcare, customers use a digital identity process, which usually involves several steps to verify end-users during online onboarding or step-up authentication. An example of one step that can be used is face search, which can help determine whether a new end-user’s face matches those associated with an existing account.

Building an accurate face search system involves several steps. The system must be able to detect human faces in images, extract the faces into vector representations, store face vectors in a database, and compare new faces against existing entries. Amazon Rekognition makes this effortless by giving you pre-trained models that are invoked via simple API calls.

Amazon Rekognition enables you to achieve very high face search accuracy with a single face image. In some cases, you can use multiple images of the same person’s face to create user vectors and improve accuracy even further. This is especially helpful when images have variations in lighting, poses, and appearances.

In this post, we demonstrate how to use the Amazon Rekognition Face Search APIs with user vectors to increase the similarity score for true matches and decrease the similarity score for true non-matches.

We compare the results of performing face matching with and without user vectors.

Amazon Rekognition face matching

Amazon Rekognition face matching enables measuring the similarity of a face vector extracted from one image to a face vector extracted from another image. A pair of face images is said to be a true match if both images contain the face of the same person, and a true non-match otherwise. Amazon Rekognition returns a score for the similarity of the source and target faces. The minimum similarity score is 0, implying very little similarity, and the maximum is 100.

For comparing a source face with a collection of target faces (1:N matching), Amazon Rekognition allows you to create a Collection object and populate it with faces from images using API calls.

When adding a face to a collection, Amazon Rekognition doesn’t store the actual image of the face but rather the face vector, a mathematical representation of the face. With the SearchFaces API, you can compare a source face with one or several collections of target faces.

In June 2023, AWS launched user vectors, a new capability that significantly improves face search accuracy by using multiple face images of a user. Now, you can create user vectors, which aggregate multiple face vectors of the same user. User vectors offer higher face search accuracy with more robust depictions, because they contain varying degrees of lighting, sharpness, pose, appearance, and more. This improves the accuracy compared to searching against individual face vectors.

In the following sections, we outline the process of using Amazon Rekognition user vectors. We guide you through creating a collection, storing face vectors in that collection, aggregating those face vectors into user vectors, and then comparing the results of searching against those individual face vectors and user vectors.

Solution overview

For this solution, we use an Amazon Rekognition collection of users, each with its associated indexed face vectors from a number of different images of faces for each user.

Let’s look at the workflow to build a collection with users and faces:

1. Create an Amazon Rekognition collection.
2. For each user, create a user in the collection.
3. For each image of the user, add the face to the collection (IndexFaces, which returns face ID corresponding to each face vector).
4. Associate all indexed face IDs with the user (this is necessary for user vectors).

Then, we will compare the following workflows:

Searching with a new given input image against individual face vectors in our collection:

1. Get all faces from an image (DetectFaces).
2. For each face, compare against individual faces in our collection (SearchFacesByImage).

Searching with a new given input image against user vectors in our collection:

1. Get all faces from an image (DetectFaces).
2. For each face, compare to the user vector (SearchUsersByImage).

Now let’s describe the solution in details.

Prerequisites

Add the following policy to your AWS Identity and Access Management (IAM) user or role. The policy grants you permission to the relevant Amazon Rekognition APIs and allows access to an Amazon Simple Storage Service (Amazon S3) bucket to store the images:

{
    "Version": "2012-10-17",
    "Statement": [
        {
            "Sid": "RekognitionPermissions",
            "Effect": "Allow",
            "Action": [
                "rekognition:CreateCollection",
                "rekognition:DeleteCollection",
                "rekognition:CreateUser",
                "rekognition:IndexFaces",
                "rekognition:DetectFaces",
                "rekognition:AssociateFaces",
                "rekognition:SearchUsersByImage",
                "rekognition:SearchFacesByImage"
            ],
            "Resource": "*"
        },
        {
            "Sid": "S3BucketPermissions",
            "Effect": "Allow",
            "Action": [
                "s3:GetObject",
                "s3:PutObject",
                "s3:ListBucket"
            ],
            "Resource": [
                "arn:aws:s3:::<replace_with_your_bucket>/*",
                "arn:aws:s3:::<replace_with_your_bucket>"
            ]
        }
    ]
}

Create an Amazon Rekognition collection and add users and faces

First, we create an S3 bucket to store users’ images. We organize the bucket by creating a folder for each user that contains their personal images. Our images folder looks like the following structure:

── images
│   ├── photo.jpeg
│   ├── Swami
│   │   ├── Swami1.jpeg
│   │   └── Swami2.jpeg
│   └── Werner
│       ├── Werner1.jpeg
│       ├── Werner2.jpeg
│       └── Werner3.jpeg

Our S3 bucket has a directory for each user that stores their images. There are currently two folders, and each contains several images. You can add more folders for your users, each containing one or more images to be indexed.

Next, we create our Amazon Rekognition collection. We have supplied helpers.py, which contains different methods that we use:

• create_collection – Create a new collection
• delete_collection – Delete a collection
• create_user – Create a new user in a collection
• add_faces_to_collection – Add faces to collection
• associate_faces – Associate face_ids to a user in a collection
• get_subdirs – Get all subdirectories under the S3 prefix
• get_files – Get all files under the S3 prefix

The following is an example method for creating an Amazon Rekognition collection:

import boto3
session = boto3.Session()
client = session.client('rekognition')

def create_collection(collection_id):
    try:
        # Create a collection
        print('Creating collection:' + collection_id)
        response = client.create_collection(CollectionId=collection_id)
        print('Collection ARN: ' + response['CollectionArn'])
        print('Status code: ' + str(response['StatusCode']))
        print('Done...')
    except client.exceptions.ResourceAlreadyExistsException:
        print('Resource already exits...')

Create the collection with the following code:

import helpers
collection_id = "faces-collection"
helpers.create_collection(collection_id)

Next, let’s add the face vectors into our collection and aggregate them into user vectors.

For each user in the S3 directory, we create a user vector in the collection. Then we index the face images for each user into the collection as individual face vectors, which generates face IDs. Lastly, we associate the face IDs to the appropriate user vector.

This creates two types of vectors in our collection:

• Individual face vectors
• User vectors, which are built based on the face vector IDs supplied using the method associate_faces

See the following code:

bucket = '<replace_with_your_bucket>'
prefix = 'images/'

# Get all the users directories from s3 containing the images
folder_list = helpers.get_subdirs(bucket, prefix)
print(f"Found users folders: {folder_list}")
print()

for user_id in folder_list:
    face_ids = []
    helpers.create_user(collection_id, user_id)
    # Get all files per user under the s3 user directory
    images = helpers.get_files(bucket, prefix + user_id + "/")
    print (f"Found images={images} for {user_id}")
    for image in images:
        face_id = helpers.add_faces_to_collection(bucket, image, collection_id)
        face_ids.append(face_id)
    helpers.associate_faces(collection_id, user_id, face_ids)
    print()

We use the following methods:

• get_subdirs – Returns a list of all the users’ directories. In our example, the value is [Swami,Werner].
• get_files – Returns all the images files under the S3 prefix for the user.
• face_ids – This is a list containing all the face IDs belonging to a user. We use this list when calling the AssociateFaces API.

As explained earlier, you can add more users by adding folders for them (the folder dictates the user ID) and add your images in that folder (no ordering is required for the files).

Now that our environment is set up and we have both individual face vectors and user vectors, let’s compare our search quality against each of them. To do that, we use a new photo with multiple people and attempt to match their faces against our collection, first against the individual face vectors and then against the user vectors.

Face search of image against a collection of individual face vectors

To search against our individual face vectors, we use the Amazon Rekognition SearchFacesByImage API. This function uses a source face image to search against individual face vectors in our collection and returns faces that match our defined similarity score threshold.

An important consideration is that the SearchFacesByImage API will only operate on the largest face detected in the image. If multiple faces are present, you need to crop each individual face and pass it separately to the method for identification.

For extracting faces details from an image (such as their location on the image), we use the Amazon Rekognition DetectFaces API.

The following detect_faces_in_image method detects faces in an image. For each face, it performs the following actions:

• Print its bounding box location
• Crop the face from the image and check if such face exists in the collection and print the user or ‘Unknown’
• Print the similarity score

The example Python code uses the Pillow library for doing the image manipulations (such as printing, drawing, and cropping).

We use a similarity score threshold of 99%, which is a common setting for identity verification use cases.

Run the following code:

import detect_users
from PIL import Image

# The image we would like to match faces against our collection.
file_key= "images/photo.jpeg"

img = detect_users.detect_faces_in_image(
    bucket, 
    file_key, 
    collection_id, 
    threshold=99
)
img.show() # or in Jupyter use display(img)

file_key is the S3 object key we want to match against our collection. We have supplied an example image (photo.jpeg) under the images folder.

The following image shows our results.

Using a threshold of 99%, only one person was identified. Dr. Werner Vogels was flagged as Unknown. If we run the same code using a lower threshold of 90 (set threshold=90), we get the following results.

Now we see Dr. Werner Vogel’s face has a similarity score of 96.86%. Next, let’s check if we can get the similarity score above our defined threshold by using user vectors.

Face search of image against a collection of user vectors

To search against our user vectors, we use the Amazon Rekognition SearchUsersByImage API. This function uses a source face image to search against user vectors in our collection and returns users that match our defined similarity score threshold.

The same consideration is relevant here – the SearchUsersByImage API will only operate on the largest face detected in the image. If there are multiple faces present, you need to crop each individual face and pass it separately to the method for identification.

For extracting faces details from an image (such as their location on the image), we use the Amazon Rekognition DetectFaces API.

The following detect_users_in_image method detects faces in an image. For each face, it performs the following actions:

• Print its bounding box location
• Crop the face from the image and check if such user face exists in our collection and print the user or ‘Unknown’
• Print the similarity score

See the following code:

import boto3
import io
import math
from PIL import Image, ImageDraw, ImageFont

def detect_users_in_image(bucket, key, collection_id, threshold=80):

    session = boto3.Session()
    client = session.client('rekognition')

    # Load image from S3 bucket
    s3_connection = boto3.resource('s3')
    s3_object = s3_connection.Object(bucket, key)
    s3_response = s3_object.get()

    stream = io.BytesIO(s3_response['Body'].read())
    image = Image.open(stream)

    # Call DetectFaces to find faces in image
    response = client.detect_faces(
        Image={'S3Object': {'Bucket': bucket, 'Name': key}},
        Attributes=['ALL']
    )

    imgWidth, imgHeight = image.size
    draw = ImageDraw.Draw(image)

    # Calculate and display bounding boxes for each detected face
    for faceDetail in response['FaceDetails']:
        print('The detected face is between ' + str(faceDetail['AgeRange']['Low'])
              + ' and ' + str(faceDetail['AgeRange']['High']) + ' years old')

        box = faceDetail['BoundingBox']
        left = imgWidth * box['Left']
        top = imgHeight * box['Top']
        width = imgWidth * box['Width']
        height = imgHeight * box['Height']

        print('Left: ' + '{0:.0f}'.format(left))
        print('Top: ' + '{0:.0f}'.format(top))
        print('Face Width: ' + "{0:.0f}".format(width))
        print('Face Height: ' + "{0:.0f}".format(height))

        points = (
            (left, top),
            (left + width, top),
            (left + width, top + height),
            (left, top + height),
            (left, top)
        )

        # Crop the face box and convert it to byte array
        face = image.crop((left, top, left + width, top + height))
        imgByteArr = image_to_byte_array(face, image.format)

        # Search for a user in our collection using the cropped image
        user_response = client.search_users_by_image(
            CollectionId=collection_id,
            Image={'Bytes': imgByteArr},
            UserMatchThreshold=threshold
        )
        # print (user_response)

        # Extract user id and the similarity from the response
        if (user_response['UserMatches']):
            similarity = user_response['UserMatches'][0]['Similarity']
            similarity = (math.trunc(similarity * 100) / 100) if isinstance(similarity, float) else similarity
            user_id = user_response['UserMatches'][0]['User']['UserId']
            print(f"User {user_id} was found, similarity of {similarity}%")
            print("")
        else:
            user_id = "Unknown"
            similarity = 0

        draw.line(points, fill='#00d400', width=4)
        font = ImageFont.load_default(size=25)
        draw.text((left, top - 30), user_id, fill='#00d400', font=font)
        if similarity > 0:
            draw.text((left, top + 1), str(similarity), fill='#00d400', font=font)

    return image

The function returns a modified image with the results that can be saved to Amazon S3 or printed. The function also outputs statistics about the estimated ages of the faces to the terminal.

Run the following code:

import detect_users
from PIL import Image

# The image we would like to match faces against our collection.
file_key= "images/photo.jpeg"

img = detect_users.detect_users_in_image(
    bucket, 
    file_key, 
    collection_id, 
    threshold=99
)
img.show() # or in Jupyter use display(img)

The following image shows our results.

The users that exist in our collection were identified correctly with high similarity (over 99%).

We were able to increase the similarity score by using three face vectors per user vector. As we increase the number of face vectors used, we expect the similarity score for true matches to also increase. You can use up to 100 face vectors per user vector.

An end-to-end example code can be found in the GitHub repository. It includes a detailed Jupyter notebook that you can run on Amazon SageMaker Studio (or other alternatives).

Clean up

To delete the collection, use the following code:

helpers.delete_collection(collection_id)

Conclusion

In this post, we presented how to use Amazon Rekognition user vectors to implement face search against a collection of users’ faces. We demonstrated how to improve face search accuracy by using multiple face images per user and compared it against individual face vectors. Additionally, we described how you can use the different Amazon Rekognition APIs to detect faces. The provided example code serves as a solid foundation for constructing a functional face search system.

For more information about Amazon Rekognition user vectors, refer to Searching faces in a collection. If you’re new to Amazon Rekognition, you can use our Free Tier, which lasts 12 months and includes processing 5,000 images per month and storing 1,000 user vector objects per month.

About the Authors

Arik Porat is a Senior Startups Solutions Architect at Amazon Web Services. He works with startups to help them build and design their solutions in the cloud, and is passionate about machine learning and container-based solutions. In his spare time, Arik likes to play chess and video games.

Eliran Efron is a Startups Solutions Architect at Amazon Web Services. Eliran is a data and compute enthusiast, assisting startups designing their system architectures. In his spare time, Eliran likes to build and race cars in Touring races and build IoT devices.

Read More