Connect with us


Build accurate ML training datasets using point-in-time queries with Amazon SageMaker Feature Store and Apache Spark

This post is co-written with Raphey Holmes, Software Engineering Manager, and Jason Mackay, Principal Software Development Engineer, at GoDaddy. GoDaddy is the world’s largest services platform for entrepreneurs around the globe, empowering their worldwide community of over 20 million customers—and entrepreneurs everywhere—by giving them all the help and tools they need to grow online. GoDaddy…



This post is co-written with Raphey Holmes, Software Engineering Manager, and Jason Mackay, Principal Software Development Engineer, at GoDaddy.

GoDaddy is the world’s largest services platform for entrepreneurs around the globe, empowering their worldwide community of over 20 million customers—and entrepreneurs everywhere—by giving them all the help and tools they need to grow online. GoDaddy needs a robust, validated, and customizable ML feature management solution, and has chosen Amazon SageMaker Feature Store to manage thousands of features across dozens of feature groups with unique data pipelines and update schedules. Feature Store lets GoDaddy use point-in-time queries to support accurate training and deployment of machine learning (ML) models, covering everything from personalizing content, to preventing fraud, to helping customers find the perfect domain name.

Feature Store lets you define groups of features, use batch ingestion and streaming ingestion, retrieve features with as low as single-digit millisecond latency for highly accurate online predictions, and extract point-in-time correct datasets for training. Instead of building and maintaining these infrastructure capabilities, you get a fully managed service that scales as your data grows, enables feature sharing across teams, and lets data scientists focus on building great ML-driven products for game-changing business use cases. Teams can now deliver robust features and reuse them in a variety of models that may be built by different teams.

In this post, we (the joint team of GoDaddy and AWS architects), explain how to use Feature Store and the processing power of Apache Spark to create accurate training datasets using point-in-time queries against reusable feature groups in a scalable fashion.

Avoid data leakage by using point-in-time correct queries

In ML, data leakage or target leakage is accidentally using data in model training that wouldn’t be available at the time of prediction. Leakage can be subtle and difficult to detect, yet the business impact can be significant. Models with leakage perform unrealistically well in development, but they deliver poor model accuracy in production without the benefit of future data.

Leakage with time-dependent features can occur in a wide range of use cases. For example, a model predicting lung disease might use features about a patient’s use of medications or surgical procedures. A recommendation model on a website may use customer orders to predict what offers would be most attractive to that customer. These features are valid when used correctly, but data scientists must ensure that the feature values are built only using data that could be known before the target was observed. For example, if a patient was diagnosed at time t1, any data about medications or hospital visits at times beyond t1 must be excluded when creating a training dataset.

So how do data science teams provide a rich set of ML features, while ensuring they don’t leak future data into their trained models? Companies are increasingly adopting the use of a feature store to help solve this model training challenge. A robust feature store provides an offline store with a complete history of feature values. Feature records include timestamps to support point-in-time correct queries. Data scientists can query for the exact set of feature values that would have been available at a specific time, without the chance of including data from beyond that time.

Let’s use a diagram to explain the concept of a point-in-time feature query. Imagine we’re training a fraud detection model on a set of historical transactions. Each transaction has features associated with various entities involved in the transaction, such as the consumer, merchant, and credit card. Feature values for these entities change over time, and they’re updated on different schedules. To avoid leaking future feature values, a point-in-time query retrieves the state of each feature that was available at each transaction time, and no later. For example, the transaction at time t2 can only use features available before time t2, and the transaction at t1 can’t use features from timestamps greater than t1.

The resulting training dataset in the following diagram shows that a point-in-time query returns an accurate set of feature values for each transaction, avoiding values that would have only been known in the future. Reliably retrieving the right set of values from history ensures that model performance won’t suffer when it faces real-world transactions.

To solidify the concept one step further, let’s consider two other types of queries that don’t protect you from data leakage:

  • Get latest feature values – Consider a query that simply returns the latest feature values available for each feature. Although such a query works well for creating a batch scoring dataset, the query mistakenly leaks data that wasn’t available for the transactions at t1 and t2, providing a very poor training dataset.
  • Get features as of a specific timestamp – Likewise, a query that returns all feature data as of a single timestamp produces an inappropriate training dataset, because it treats all records uniformly instead of using a distinct timestamp for each training dataset entry. So-called time travel capabilities are great for auditing and reproducing experiments. However, time travel doesn’t solve for accurate training dataset extraction, because it doesn’t account for event timestamps that vary for each training record.

Point-in-time queries as part of the overall ML lifecycle

The following diagram shows how point-in-time queries fit into the overall ML lifecycle. The diagram starts with a set of automated feature pipelines that perform feature transformations and ingest feature records into Feature Store. Pipelines for individual feature groups are independent and can run on varying schedules. One feature group may be refreshed nightly, whereas another is updated hourly, and a third may have updates that are triggered as source data arrives on an input stream, such as an Apache Kafka topic or via Amazon Kinesis Data Streams.

Depending on the configuration of your feature groups, the resulting features are made available in an online store, an offline store, or both with automatic synchronization. The offline store provides a secure and scalable repository of features, letting data scientists create training and validation datasets or batch scoring datasets from a set of feature groups that can be managed centrally with a full history of feature values.

The result of a point-in-time query is a training dataset, which can then be used directly with an ML algorithm, or as input to a SageMaker training job. A point-in-time query can be run interactively in an Apache Spark environment such as a SageMaker notebook or Amazon EMR notebook. For large-scale training datasets, you can run these queries in a SageMaker Processing job, which lets you scale the job to a multi-instance cluster without having to manage any infrastructure.

Although we show an Apache Spark implementation of point in time queries in this post, Amazon Athena provides another alternative. With Athena, you can create a temporary table containing your selection criteria, and then use SQL to join that criteria to your offline store feature groups. The SQL query can select the most recent feature values that are older than the targeted event time for each training dataset row.

To learn how to create automated batch pipelines, see Automate feature engineering pipelines with Amazon SageMaker. For more information about streaming feature pipelines, see Using streaming ingestion with Amazon SageMaker Feature Store to make ML-backed decisions in near-real time.

Feature Store timestamps

Before we walk through how to perform point-in-time correct queries against the offline store, it’s important to understand the definition and purpose of the three different timestamps that SageMaker provides in the offline store schema for every feature group:

  • Event time – The customer-defined timestamp associated with the feature record, such as the transaction time, the time a customer placed an order, or the time a new insurance claim was created. The specific name of this feature is specified when the feature group is created, and the customer’s ingestion code is responsible for populating this timestamp.
  • API invocation time – The time when SageMaker receives the API call to write a feature record to the feature store. This timestamp is automatically populated by SageMaker as the api_invocation_time feature.
  • Write time – The time when the feature record is persisted to the offline store. This timestamp is always greater than API invocation time, and is automatically populated by SageMaker as the write_time feature.

Depending on your use case, you can use a combination of the timestamp fields in SageMaker to choose accurate feature values. Each instance in a training dataset has a customer-defined event time and a record identifier. When you join a list of training events against feature history in a point-in-time query, you can ignore all records that happened after the instance-specific event timestamp, and select the most recent of the remaining records. The event time field is the key to this join. In a standard use case, choosing the most recent remaining record is sufficient. However, if a feature value has been corrected or revised as part of a feature backfill, the offline store contains multiple records with the same event time and record identifier. Both the original record and the new record are available, and they each have a distinct write time. For these situations, a point-in-time query can use the write_time feature or the api_invocation_time feature as a tie-breaker, ensuring the corrected feature value is returned.

Implement a point-in-time correct query

Now that we have explained the concepts, let’s dive into the implementation details of how to efficiently perform point-in-time queries using Feature Store and Apache Spark. You can try this out in your own account using the following GitHub repo. This repository contains three Jupyter notebooks plus some schema files used to create our feature groups. In this section, we show you the inner workings of a query implementation. In the notebook, we also provide a reusable function that makes this as simple as passing a few parameters:

def get_historical_feature_values ( fg_name: str, entity_df: DataFrame, spark: SparkSession, allowed_staleness_days: int = 14, remove_extra_columns: bool = True) -> DataFrame:

The implementation of our point-in-time query uses SageMaker, Jupyter notebooks, and Apache Spark (PySpark). Most of the intermediate data is stored in Spark DataFrames, which gives us powerful built-in methods to manipulate, filter, and reduce that dataset so that the query runs efficiently. To enable Spark to run properly in our environment, we configure the Spark session to allocate extra driver memory and executor cores:

spark = (SparkSession .builder .config(“spark.driver.extraClassPath”, classpath) .config(“spark.executor.memory”, ‘1g’) .config(‘spark.executor.cores’, ’16’) .config(“spark.driver.memory”,’8g’) .getOrCreate())

Next, we load the historical transaction dataset that contains the raw credit card transactions along with the target, meaning the fraud label that we want to predict. This data contains primarily the attributes that are part of the transaction itself, and includes the following columns:

|– tid: string (nullable = true) |– event_time: string (nullable = true) |– cc_num: long (nullable = true) |– consumer_id: string (nullable = true) |– amount: double (nullable = true) |– fraud_label: string (nullable = true)

The need to run point-in-time queries originates from not having perfect information at all times pertaining to a given transaction. For example, if we had a complete set of aggregate features for every transaction event, we could use this data directly. Most organizations can’t build out this type of data for every event, but instead run periodic jobs that calculate these aggregates, perhaps on an hourly or daily basis.

For this post, we simulate these daily snapshots by running an aggregation function for each day in our timeframe that spans 1 month. The aggregation code creates two dataframes, one indexed by credit card number, and the other indexed by consumer ID. Each dataframe contains data for lookback periods of 1–7 days. These datasets simulate the periodic job runs that create snapshots of our aggregate features, and they’re written to the offline store. The following screenshot is a sample of the aggregated consumer features that are generated periodically.

To prepare the input criteria for our point-in-time query, we begin by creating an entity dataframe, which contains one row for each desired training dataset row. The entity dataframe identifies the consumer IDs of interest, each paired with an event time that represents our cutoff time for that training row. The consumer ID is used to join with feature data from the consumer feature group, and the transaction event time helps us filter out newer feature values. For our example, we look up a subset of historical transactions from one specific week:

last_1w_df = spark.sql(‘select * from trans where event_time >= “2021-03-25T00:00:00Z” and event_time <= "2021-03-31T23:59:59Z"') cid_ts_tuples = r: (r.consumer_id, r.cc_num, r.event_time, r.amount, int(r.fraud_label))).collect() entity_df = spark.createDataFrame(cid_ts_tuples, entity_df_schema)

This produces the following entity dataframe that drives our point-in-time query.

To query against the offline store, we need to know the Amazon Simple Storage Service (Amazon S3) location of our feature group. We use the describe_feature_group method to look up that location:

feature_group_info = sagemaker_client.describe_feature_group(FeatureGroupName=CONS_FEATURE_GROUP) resolved_offline_store_s3_location = feature_group_info[‘OfflineStoreConfig’][‘S3StorageConfig’][‘ResolvedOutputS3Uri’] # Spark’s Parquet file reader requires replacement of ‘s3’ with ‘s3a’ offline_store_s3a_uri = resolved_offline_store_s3_location.replace(“s3:”, “s3a:”)

In the preceding code, we use the S3A filesystem client. This client ensures we have the latest patches and performance enhancements for accessing S3 objects.

Now we use Spark to read data from the offline store, which is stored in Parquet format in the S3 location from the preceding code:

feature_store_df =

The following output shows the schema of the data read from the offline store. It contains several additional fields automatically populated by SageMaker: timestamp fields, as defined earlier in this post (write_time, api_invocation_time), a soft delete flag (is_deleted), and date-time partitioning fields (year, month, day, and hour).

|– consumer_id: string (nullable = true) |– num_trans_last_7d: long (nullable = true) |– avg_amt_last_7d: double (nullable = true) |– num_trans_last_1d: long (nullable = true) |– avg_amt_last_1d: double (nullable = true) |– event_time: string (nullable = true) |– write_time: timestamp (nullable = true) |– api_invocation_time: timestamp (nullable = true) |– is_deleted: boolean (nullable = true) |– year: integer (nullable = true) |– month: integer (nullable = true) |– day: integer (nullable = true) |– hour: integer (nullable = true)

The is_deleted attribute is a Boolean soft delete indicator for the referenced record identifier. If the DeleteRecord method is called, a new record is inserted with the is_deleted flag set to True in the offline store. The date-time partitioning fields are used to segregate the individual data files written to the offline store, and are useful when navigating to a desired timeframe or reading a subset of data.

To optimize the performance of the point-in-time query, we can immediately filter out all records that don’t meet our overall criteria. In many cases, this optimization drastically reduces the number of records we carry forward, and therefore makes the subsequent joins more efficient. We use a min/max time window to drop all data that doesn’t meet our timeframe boundary. We also include a staleness window to ensure that we don’t include records that are too old to be useful. The appropriate length of the staleness window is specific to your use case. See the following code:

# NOTE: This filter is simply a performance optimization # Filter out records from after query max_time and before staleness # window prior to the min_time. # Doing this prior to individual {consumer_id, joindate} filtering # speeds up subsequent filters for large scale queries. # Choose a “staleness” window of time before which we want # to ignore records allowed_staleness_days = 14 # Eliminate history that is outside of our time window # This window represents the {max_time – min_time} delta, # plus our staleness window # entity_df used to define bounded time window minmax_time = entity_df.agg(sql_min(“query_date”), sql_max(“query_date”)).collect() min_time, max_time = minmax_time[0][“min(query_date)”], minmax_time[0][“max(query_date)”] # Via the staleness check, we are actually removing items when # event_time is MORE than N days before min_time # Usage: datediff ( enddate, startdate ) – returns days filtered = feature_store_active_df.filter( (feature_store_active_df.event_time <= max_time) & (datediff(lit(min_time), feature_store_active_df.event_time) <= allowed_staleness_days) )

Now we’re ready to join the filtered dataset with the entity dataframe to reduce the results to only those consumer IDs (our entities) that are part of our desired training dataset. This inner join uses consumer_id as a join key, thereby removing transactions for other consumers:

t_joined = (filtered.join(entity_df, filtered.consumer_id == entity_df.consumer_id, ‘inner’) .drop(entity_df.consumer_id)

This results in an enhanced dataframe with all the aggregate attributes from our consumer feature group, for each targeted training row. We still need to remove transactions that are outside of our selected time window. This window is defined as the time no later than the event time of interest, and no earlier than our selected staleness allowance. This time window filtering is run against each item that is part of our chosen list of training rows. See the following filter code, with the results named drop_future_and_stale_df:

# Filter out data from after query time to remove future data leakage. # Also filter out data that is older than our allowed staleness # window (days before each query time) drop_future_and_stale_df = t_joined.filter( (t_joined.event_time <= entity_df.query_date) & (datediff(entity_df.query_date, t_joined.event_time) <= allowed_staleness_days))

In our final training dataset, we want to allow for multiple aggregate records per entity ID (think multiple credit card transactions by a single consumer), but only keep exactly one record per transaction. Therefore, we assemble a composite key made from the consumer ID and the query timestamp: {x.consumer_id}-{x.query_date}. This step ensures that only the latest aggregate record for each composite key remains. Doing this naively (using a real sort operation) would be expensive. Instead, we can implement this using a custom reduction passed to Spark RDD reduceByKey(), which scales very well for large datasets. See the following code:

# Group by record id and query timestamp, select only the latest # remaining record by event time, # using write time as a tie breaker to account for any more # recent backfills or data corrections. latest = x: (f'{x.consumer_id}-{x.query_date}’, x)) .reduceByKey( lambda x, y: x if (x.event_time, x.write_time) > (y.event_time, y.write_time) else y).values() latest_df = latest.toDF(drop_future_and_stale_df.schema)

To view our final results, we can select specific columns for display and reference a test_consumer_id taken from our original dataframe:‘consumer_id’, ‘query_date’, ‘avg_amt_last_7d’, ‘event_time’, ‘write_time’) .where(latest_df.consumer_id == test_consumer_id) .orderBy(col(‘query_date’).desc(),col(‘event_time’).desc(), col(‘write_time’).desc()) .show(15,False)

We can also drop columns that we don’t need for training:

cols_to_drop = (‘api_invocation_time’,’write_time’,’is_deleted’,’cc_num’,’year’,’month’,’day’,’hour’) latest_df = latest_df.drop(*cols_to_drop)

The following screenshot is a sample of the final results from our point-in-time query. These results demonstrate clearly that we’re only choosing features from the past, and not leaking any future values. The event time for each record is earlier than the query timestamp, ensuring we have the latest features, without using features that would have only been known in the future.

This completes the historical query, and we now have an accurate training dataset that represents a point-in-time query for each individual training transaction.


In this post, we described the concept of point-in-time correct queries and explained the importance of these queries in training effective ML models. We showed an efficient and reproducible way to use historical feature data using Feature Store and Apache Spark. We hope you experiment with the code we’ve provided, and try it out on your own datasets. We’re always looking forward to your feedback, either through your usual AWS Support contacts or on the Amazon SageMaker Discussion Forum.

About the Authors

Paul Hargis has focused his efforts on Machine Learning at several companies, including AWS, Amazon, and Hortonworks. He enjoys building technology solutions and also teaching people how to make the most of it. Prior to his role at AWS, he was lead architect for Amazon Exports and Expansions helping improve experience for international shoppers. Paul likes to help customers expand their machine learning initiatives to solve real-world problems.


  Raphey Holmes is an engineering manager on GoDaddy’s Machine Learning platform team. Prior to changing careers, he worked for a decade as a high school physics teacher, and he still loves all things related to teaching and learning. See picture attached.



Jason Mackay is a Principal SDE at GoDaddy on the GoDaddy’s Machine Learning Team. He has been in the software industry for 25 years spanning operating systems, parallel/concurrent/distributed systems, formal languages, high performance cryptography, big data, and machine learning.



Mark Roy is a Principal Machine Learning Architect for AWS, helping customers design and build AI/ML solutions. Mark’s work covers a wide range of ML use cases, with a primary interest in computer vision, deep learning, and scaling ML across the enterprise. He has helped companies in many industries, including insurance, financial services, media and entertainment, healthcare, utilities, and manufacturing. Mark holds six AWS certifications, including the ML Specialty Certification. Prior to joining AWS, Mark was an architect, developer, and technology leader for over 25 years, including 19 years in financial services.


Continue Reading
Click to comment

Leave a Reply

Your email address will not be published.


Customize pronunciation using lexicons in Amazon Polly

Amazon Polly is a text-to-speech service that uses advanced deep learning technologies to synthesize natural-sounding human speech. It is used in a variety of use cases, such as contact center systems, delivering conversational user experiences with human-like voices for automated real-time status check, automated account and billing inquiries, and by news agencies like The Washington…




Amazon Polly is a text-to-speech service that uses advanced deep learning technologies to synthesize natural-sounding human speech. It is used in a variety of use cases, such as contact center systems, delivering conversational user experiences with human-like voices for automated real-time status check, automated account and billing inquiries, and by news agencies like The Washington Post to allow readers to listen to news articles.

As of today, Amazon Polly provides over 60 voices in 30+ language variants. Amazon Polly also uses context to pronounce certain words differently based upon the verb tense and other contextual information. For example, “read” in “I read a book” (present tense) and “I will read a book” (future tense) is pronounced differently.

However, in some situations you may want to customize the way Amazon Polly pronounces a word. For example, you may need to match the pronunciation with local dialect or vernacular. Names of things (e.g., Tomato can be pronounced as tom-ah-to or tom-ay-to), people, streets, or places are often pronounced in many different ways.

In this post, we demonstrate how you can leverage lexicons for creating custom pronunciations. You can apply lexicons for use cases such as publishing, education, or call centers.

Customize pronunciation using SSML tag

Let’s say you stream a popular podcast from Australia and you use the Amazon Polly Australian English (Olivia) voice to convert your script into human-like speech. In one of your scripts, you want to use words that are unknown to Amazon Polly voice. For example, you want to send Mātariki (Māori New Year) greetings to your New Zealand listeners. For such scenarios, Amazon Polly supports phonetic pronunciation, which you can use to achieve a pronunciation that is close to the correct pronunciation in the foreign language.

You can use the Speech Synthesis Markup Language (SSML) tag to suggest a phonetic pronunciation in the ph attribute. Let me show you how you can use SSML tag.

First, login into your AWS console and search for Amazon Polly in the search bar at the top. Select Amazon Polly and then choose Try Polly button.

In the Amazon Polly console, select Australian English from the language dropdown and enter following text in the Input text box and then click on Listen to test the pronunciation.

I’m wishing you all a very Happy Mātariki.

Sample speech without applying phonetic pronunciation:

If you hear the sample speech above, you can notice that the pronunciation of Mātariki – a word which is not part of Australian English – isn’t quite spot-on. Now, let’s look at how in such scenarios we can use phonetic pronunciation using SSML tag to customize the speech produced by Amazon Polly.

To use SSML tags, turn ON the SSML option in Amazon Polly console. Then copy and paste following SSML script containing phonetic pronunciation for Mātariki specified inside the ph attribute of the tag.

I’m wishing you all a very Happy Mātariki.

With the tag, Amazon Polly uses the pronunciation specified by the ph attribute instead of the standard pronunciation associated by default with the language used by the selected voice.

Sample speech after applying phonetic pronunciation:

If you hear the sample sound, you’ll notice that we opted for a different pronunciation for some of vowels (e.g., ā) to make Amazon Polly synthesize the sounds that are closer to the correct pronunciation. Now you might have a question, how do I generate the phonetic transcription “” for the word Mātariki?

You can create phonetic transcriptions by referring to the Phoneme and Viseme tables for the supported languages. In the example above we have used the phonemes for Australian English.

Amazon Polly offers support in two phonetic alphabets: IPA and X-Sampa. Benefit of X-Sampa is that they are standard ASCII characters, so it is easier to type the phonetic transcription with a normal keyboard. You can use either of IPA or X-Sampa to generate your transcriptions, but make sure to stay consistent with your choice, especially when you use a lexicon file which we’ll cover in the next section.

Each phoneme in the phoneme table represents a speech sound. The bolded letters in the “Example” column of the Phoneme/Viseme table in the Australian English page linked above represent the part of the word the “Phoneme” corresponds to. For example, the phoneme /j/ represents the sound that an Australian English speaker makes when pronouncing the letter “y” in “yes.”

Customize pronunciation using lexicons

Phoneme tags are suitable for one-off situations to customize isolated cases, but these are not scalable. If you process huge volume of text, managed by different editors and reviewers, we recommend using lexicons. Using lexicons, you can achieve consistency in adding custom pronunciations and simultaneously reduce manual effort of inserting phoneme tags into the script.

A good practice is that after you test the custom pronunciation on the Amazon Polly console using the tag, you create a library of customized pronunciations using lexicons. Once lexicons file is uploaded, Amazon Polly will automatically apply phonetic pronunciations specified in the lexicons file and eliminate the need to manually provide a tag.

Create a lexicon file

A lexicon file contains the mapping between words and their phonetic pronunciations. Pronunciation Lexicon Specification (PLS) is a W3C recommendation for specifying interoperable pronunciation information. The following is an example PLS document:

Matariki Mātariki NZ New Zealand

Make sure that you use correct value for the xml:lang field. Use en-AU if you’re uploading the lexicon file to use with the Amazon Polly Australian English voice. For a complete list of supported languages, refer to Languages Supported by Amazon Polly.

To specify a custom pronunciation, you need to add a element which is a container for a lexical entry with one or more element and one or more pronunciation information provided inside element.

The element contains the text describing the orthography of the element. You can use a element to specify the word whose pronunciation you want to customize. You can add multiple elements to specify all word variations, for example with or without macrons. The element is case-sensitive, and during speech synthesis Amazon Polly string matches the words inside your script that you’re converting to speech. If a match is found, it uses the element, which describes how the is pronounced to generate phonetic transcription.

You can also use for commonly used abbreviations. In the preceding example of a lexicon file, NZ is used as an alias for New Zealand. This means that whenever Amazon Polly comes across “NZ” (with matching case) in the body of the text, it’ll read those two letters as “New Zealand”.

For more information on lexicon file format, see Pronunciation Lexicon Specification (PLS) Version 1.0 on the W3C website.

You can save a lexicon file with as a .pls or .xml file before uploading it to Amazon Polly.

Upload and apply the lexicon file

Upload your lexicon file to Amazon Polly using the following instructions:

  1. On the Amazon Polly console, choose Lexicons in the navigation pane.
  2. Choose Upload lexicon.
  3. Enter a name for the lexicon and then choose a lexicon file.
  4. Choose the file to upload.
  5. Choose Upload lexicon.

If a lexicon by the same name (whether a .pls or .xml file) already exists, uploading the lexicon overwrites the existing lexicon.

Now you can apply the lexicon to customize pronunciation.

  1. Choose Text-to-Speech in the navigation pane.
  2. Expand Additional settings.
  3. Turn on Customize pronunciation.
  4. Choose the lexicon on the drop-down menu.

You can also choose Upload lexicon to upload a new lexicon file (or a new version).

It’s a good practice to version control the lexicon file in a source code repository. Keeping the custom pronunciations in a lexicon file ensures that you can consistently refer to phonetic pronunciations for certain words across the organization. Also, keep in mind the pronunciation lexicon limits mentioned on Quotas in Amazon Polly page.

Test the pronunciation after applying the lexicon

Let’s perform quick test using “Wishing all my listeners in NZ, a very Happy Mātariki” as the input text.

We can compare the audio files before and after applying the lexicon.

Before applying the lexicon:

After applying the lexicon:


In this post, we discussed how you can customize pronunciations of commonly used acronyms or words not found in the selected language in Amazon Polly. You can use SSML tag which is great for inserting one-off customizations or testing purposes. We recommend using Lexicon to create a consistent set of pronunciations for frequently used words across your organization. This enables your content writers to spend time on writing instead of the tedious task of adding phonetic pronunciations in the script repetitively. You can try this in your AWS account on the Amazon Polly console.

Summary of resources

About the Authors

Ratan Kumar is a Solutions Architect based out of Auckland, New Zealand. He works with large enterprise customers helping them design and build secure, cost-effective, and reliable internet scale applications using the AWS cloud. He is passionate about technology and likes sharing knowledge through blog posts and twitch sessions.

Maciek Tegi is a Principal Audio Designer and a Product Manager for Polly Brand Voices. He has worked in professional capacity in the tech industry, movies, commercials and game localization. In 2013, he was the first audio engineer hired to the Alexa Text-To- Speech team. Maciek was involved in releasing 12 Alexa TTS voices across different countries, over 20 Polly voices, and 4 Alexa celebrity voices. Maciek is a triathlete, and an avid acoustic guitar player.


Continue Reading


AWS Week in Review – May 16, 2022

This post is part of our Week in Review series. Check back each week for a quick roundup of interesting news and announcements from AWS! I had been on the road for the last five weeks and attended many of the AWS Summits in Europe. It was great to talk to so many of you…




This post is part of our Week in Review series. Check back each week for a quick roundup of interesting news and announcements from AWS!

I had been on the road for the last five weeks and attended many of the AWS Summits in Europe. It was great to talk to so many of you in person. The Serverless Developer Advocates are going around many of the AWS Summits with the Serverlesspresso booth. If you attend an event that has the booth, say “Hi ” to my colleagues, and have a coffee while asking all your serverless questions. You can find all the upcoming AWS Summits in the events section at the end of this post.

Last week’s launches
Here are some launches that got my attention during the previous week.

AWS Step Functions announced a new console experience to debug your state machine executions – Now you can opt-in to the new console experience of Step Functions, which makes it easier to analyze, debug, and optimize Standard Workflows. The new page allows you to inspect executions using three different views: graph, table, and event view, and add many new features to enhance the navigation and analysis of the executions. To learn about all the features and how to use them, read Ben’s blog post.

Example on how the Graph View looks

Example on how the Graph View looks

AWS Lambda now supports Node.js 16.x runtime – Now you can start using the Node.js 16 runtime when you create a new function or update your existing functions to use it. You can also use the new container image base that supports this runtime. To learn more about this launch, check Dan’s blog post.

AWS Amplify announces its Android library designed for Kotlin – The Amplify Android library has been rewritten for Kotlin, and now it is available in preview. This new library provides better debugging capacities and visibility into underlying state management. And it is also using the new AWS SDK for Kotlin that was released last year in preview. Read the What’s New post for more information.

Three new APIs for batch data retrieval in AWS IoT SiteWise – With this new launch AWS IoT SiteWise now supports batch data retrieval from multiple asset properties. The new APIs allow you to retrieve current values, historical values, and aggregated values. Read the What’s New post to learn how you can start using the new APIs.

AWS Secrets Manager now publishes secret usage metrics to Amazon CloudWatch – This launch is very useful to see the number of secrets in your account and set alarms for any unexpected increase or decrease in the number of secrets. Read the documentation on Monitoring Secrets Manager with Amazon CloudWatch for more information.

For a full list of AWS announcements, be sure to keep an eye on the What’s New at AWS page.

Other AWS News
Some other launches and news that you may have missed:

IBM signed a deal with AWS to offer its software portfolio as a service on AWS. This allows customers using AWS to access IBM software for automation, data and artificial intelligence, and security that is built on Red Hat OpenShift Service on AWS.

Podcast Charlas Técnicas de AWS – If you understand Spanish, this podcast is for you. Podcast Charlas Técnicas is one of the official AWS podcasts in Spanish. This week’s episode introduces you to Amazon DynamoDB and shares stories on how different customers use this database service. You can listen to all the episodes directly from your favorite podcast app or the podcast web page.

AWS Open Source News and Updates – Ricardo Sueiras, my colleague from the AWS Developer Relation team, runs this newsletter. It brings you all the latest open-source projects, posts, and more. Read edition #112 here.

Upcoming AWS Events
It’s AWS Summits season and here are some virtual and in-person events that might be close to you:

You can register for re:MARS to get fresh ideas on topics such as machine learning, automation, robotics, and space. The conference will be in person in Las Vegas, June 21–24.

That’s all for this week. Check back next Monday for another Week in Review!

— Marcia


Continue Reading


Personalize your machine translation results by using fuzzy matching with Amazon Translate

A person’s vernacular is part of the characteristics that make them unique. There are often countless different ways to express one specific idea. When a firm communicates with their customers, it’s critical that the message is delivered in a way that best represents the information they’re trying to convey. This becomes even more important when…




A person’s vernacular is part of the characteristics that make them unique. There are often countless different ways to express one specific idea. When a firm communicates with their customers, it’s critical that the message is delivered in a way that best represents the information they’re trying to convey. This becomes even more important when it comes to professional language translation. Customers of translation systems and services expect accurate and highly customized outputs. To achieve this, they often reuse previous translation outputs—called translation memory (TM)—and compare them to new input text. In computer-assisted translation, this technique is known as fuzzy matching. The primary function of fuzzy matching is to assist the translator by speeding up the translation process. When an exact match can’t be found in the TM database for the text being translated, translation management systems (TMSs) often have the option to search for a match that is less than exact. Potential matches are provided to the translator as additional input for final translation. Translators who enhance their workflow with machine translation capabilities such as Amazon Translate often expect fuzzy matching data to be used as part of the automated translation solution.

In this post, you learn how to customize output from Amazon Translate according to translation memory fuzzy match quality scores.

Translation Quality Match

The XML Localization Interchange File Format (XLIFF) standard is often used as a data exchange format between TMSs and Amazon Translate. XLIFF files produced by TMSs include source and target text data along with match quality scores based on the available TM. These scores—usually expressed as a percentage—indicate how close the translation memory is to the text being translated.

Some customers with very strict requirements only want machine translation to be used when match quality scores are below a certain threshold. Beyond this threshold, they expect their own translation memory to take precedence. Translators often need to apply these preferences manually either within their TMS or by altering the text data. This flow is illustrated in the following diagram. The machine translation system processes the translation data—text and fuzzy match scores— which is then reviewed and manually edited by translators, based on their desired quality thresholds. Applying thresholds as part of the machine translation step allows you to remove these manual steps, which improves efficiency and optimizes cost.

Machine Translation Review Flow

Figure 1: Machine Translation Review Flow

The solution presented in this post allows you to enforce rules based on match quality score thresholds to drive whether a given input text should be machine translated by Amazon Translate or not. When not machine translated, the resulting text is left to the discretion of the translators reviewing the final output.

Solution Architecture

The solution architecture illustrated in Figure 2 leverages the following services:

  • Amazon Simple Storage Service – Amazon S3 buckets contain the following content:
    • Fuzzy match threshold configuration files
    • Source text to be translated
    • Amazon Translate input and output data locations
  • AWS Systems Manager – We use Parameter Store parameters to store match quality threshold configuration values
  • AWS Lambda – We use two Lambda functions:
    • One function preprocesses the quality match threshold configuration files and persists the data into Parameter Store
    • One function automatically creates the asynchronous translation jobs
  • Amazon Simple Queue Service – An Amazon SQS queue triggers the translation flow as a result of new files coming into the source bucket

Solution Architecture Diagram

Figure 2: Solution Architecture

You first set up quality thresholds for your translation jobs by editing a configuration file and uploading it into the fuzzy match threshold configuration S3 bucket. The following is a sample configuration in CSV format. We chose CSV for simplicity, although you can use any format. Each line represents a threshold to be applied to either a specific translation job or as a default value to any job.

default, 75 SourceMT-Test, 80

The specifications of the configuration file are as follows:

  • Column 1 should be populated with the name of the XLIFF file—without extension—provided to the Amazon Translate job as input data.
  • Column 2 should be populated with the quality match percentage threshold. For any score below this value, machine translation is used.
  • For all XLIFF files whose name doesn’t match any name listed in the configuration file, the default threshold is used—the line with the keyword default set in Column 1.

Auto-generated parameter in Systems Manager Parameter Store

Figure 3: Auto-generated parameter in Systems Manager Parameter Store

When a new file is uploaded, Amazon S3 triggers the Lambda function in charge of processing the parameters. This function reads and stores the threshold parameters into Parameter Store for future usage. Using Parameter Store avoids performing redundant Amazon S3 GET requests each time a new translation job is initiated. The sample configuration file produces the parameter tags shown in the following screenshot.

The job initialization Lambda function uses these parameters to preprocess the data prior to invoking Amazon Translate. We use an English-to-Spanish translation XLIFF input file, as shown in the following code. It contains the initial text to be translated, broken down into what is referred to as segments, represented in the source tags.

Consent Form CONSENT FORM FORMULARIO DE CONSENTIMIENTO Screening Visit: Screening Visit Selección

The source text has been pre-matched with the translation memory beforehand. The data contains potential translation alternatives—represented as tags—alongside a match quality attribute, expressed as a percentage. The business rule is as follows:

  • Segments received with alternative translations and a match quality below the threshold are untouched or empty. This signals to Amazon Translate that they must be translated.
  • Segments received with alternative translations with a match quality above the threshold are pre-populated with the suggested target text. Amazon Translate skips those segments.

Let’s assume the quality match threshold configured for this job is 80%. The first segment with 99% match quality isn’t machine translated, whereas the second segment is, because its match quality is below the defined threshold. In this configuration, Amazon Translate produces the following output:

Consent Form FORMULARIO DE CONSENTIMIENTO CONSENT FORM FORMULARIO DE CONSENTIMIENTO Screening Visit: Visita de selección Screening Visit Selección

In the second segment, Amazon Translate overwrites the target text initially suggested (Selección) with a higher quality translation: Visita de selección.

One possible extension to this use case could be to reuse the translated output and create our own translation memory. Amazon Translate supports customization of machine translation using translation memory thanks to the parallel data feature. Text segments previously machine translated due to their initial low-quality score could then be reused in new translation projects.

In the following sections, we walk you through the process of deploying and testing this solution. You use AWS CloudFormation scripts and data samples to launch an asynchronous translation job personalized with a configurable quality match threshold.


For this walkthrough, you must have an AWS account. If you don’t have an account yet, you can create and activate one.

Launch AWS CloudFormation stack

  1. Choose Launch Stack:
  2. For Stack name, enter a name.
  3. For ConfigBucketName, enter the S3 bucket containing the threshold configuration files.
  4. For ParameterStoreRoot, enter the root path of the parameters created by the parameters processing Lambda function.
  5. For QueueName, enter the SQS queue that you create to post new file notifications from the source bucket to the job initialization Lambda function. This is the function that reads the configuration file.
  6. For SourceBucketName, enter the S3 bucket containing the XLIFF files to be translated. If you prefer to use a preexisting bucket, you need to change the value of the CreateSourceBucket parameter to No.
  7. For WorkingBucketName, enter the S3 bucket Amazon Translate uses for input and output data.
  8. Choose Next.

    Figure 4: CloudFormation stack details

  9. Optionally on the Stack Options page, add key names and values for the tags you may want to assign to the resources about to be created.
  10. Choose Next.
  11. On the Review page, select I acknowledge that this template might cause AWS CloudFormation to create IAM resources.
  12. Review the other settings, then choose Create stack.

AWS CloudFormation takes several minutes to create the resources on your behalf. You can watch the progress on the Events tab on the AWS CloudFormation console. When the stack has been created, you can see a CREATE_COMPLETE message in the Status column on the Overview tab.

Test the solution

Let’s go through a simple example.

  1. Download the following sample data.
  2. Unzip the content.

There should be two files: an .xlf file in XLIFF format, and a threshold configuration file with .cfg as the extension. The following is an excerpt of the XLIFF file.

English to French sample file extract

Figure 5: English to French sample file extract

  1. On the Amazon S3 console, upload the quality threshold configuration file into the configuration bucket you specified earlier.

The value set for test_En_to_Fr is 75%. You should be able to see the parameters on the Systems Manager console in the Parameter Store section.

  1. Still on the Amazon S3 console, upload the .xlf file into the S3 bucket you configured as source. Make sure the file is under a folder named translate (for example, /translate/test_En_to_Fr.xlf).

This starts the translation flow.

  1. Open the Amazon Translate console.

A new job should appear with a status of In Progress.

Auto-generated parameter in Systems Manager Parameter Store

Figure 6: In progress translation jobs on Amazon Translate console

  1. Once the job is complete, click into the job’s link and consult the output. All segments should have been translated.

All segments should have been translated. In the translated XLIFF file, look for segments with additional attributes named lscustom:match-quality, as shown in the following screenshot. These custom attributes identify segments where suggested translation was retained based on score.

Custom attributes identifying segments where suggested translation was retained based on score

Figure 7: Custom attributes identifying segments where suggested translation was retained based on score

These were derived from the translation memory according to the quality threshold. All other segments were machine translated.

You have now deployed and tested an automated asynchronous translation job assistant that enforces configurable translation memory match quality thresholds. Great job!


If you deployed the solution into your account, don’t forget to delete the CloudFormation stack to avoid any unexpected cost. You need to empty the S3 buckets manually beforehand.


In this post, you learned how to customize your Amazon Translate translation jobs based on standard XLIFF fuzzy matching quality metrics. With this solution, you can greatly reduce the manual labor involved in reviewing machine translated text while also optimizing your usage of Amazon Translate. You can also extend the solution with data ingestion automation and workflow orchestration capabilities, as described in Speed Up Translation Jobs with a Fully Automated Translation System Assistant.

About the Authors

Narcisse Zekpa is a Solutions Architect based in Boston. He helps customers in the Northeast U.S. accelerate their adoption of the AWS Cloud, by providing architectural guidelines, design innovative, and scalable solutions. When Narcisse is not building, he enjoys spending time with his family, traveling, cooking, and playing basketball.

Dimitri Restaino is a Solutions Architect at AWS, based out of Brooklyn, New York. He works primarily with Healthcare and Financial Services companies in the North East, helping to design innovative and creative solutions to best serve their customers. Coming from a software development background, he is excited by the new possibilities that serverless technology can bring to the world. Outside of work, he loves to hike and explore the NYC food scene.


Continue Reading


Copyright © 2021 Today's Digital.