Connect with us

AI

Task-Oriented Dialog Agents: Recent Advances and Challenges

Hotel bookings, weather queries, restaurant reservations – chatbots and virtual assistants seem to be the perfect solution for these kinds of routine inquiries. Not surprisingly, task-oriented dialog systems that support these chatbots have become a hot topic in the AI and NLP research community. In this article, we will review the recent research breakthroughs in […]

The post Task-Oriented Dialog Agents: Recent Advances and Challenges appeared first on TOPBOTS.

Published

on

pipeline dialog systems

Hotel bookings, weather queries, restaurant reservations – chatbots and virtual assistants seem to be the perfect solution for these kinds of routine inquiries. Not surprisingly, task-oriented dialog systems that support these chatbots have become a hot topic in the AI and NLP research community.

In this article, we will review the recent research breakthroughs in task-oriented (or goal-oriented) dialog systems, see the open challenges for building a good task-oriented chatbot, and list some of the most appealing directions for further research. We’ll draw on the key ideas from two recent research papers:

Subscribe to our AI Research mailing list at the bottom of this article to be alerted when we release new summaries.

Recent advances

Existing studies on task-oriented dialog systems are usually classified into two broad categories: (1) pipeline methods and (2) end-to-end methods. 

In the first case, the whole system is divided into several components, such as natural language understanding (NLU), dialog state tracking (DST), dialog policy learning, and natural language generation (NLG). Some other combination modes are also possible. In contrast to pipeline approaches, end-to-end methods build a dialog system using a single model, where natural language context is taken as input and natural language response is generated as an output.

Obviously, pipeline systems with their modular structure are more interpretable and stable, and thus a better fit for real-world commercial applications. At the same time, end-to-end systems require fewer annotations, making them another promising alternative for business applications. So, let’s review the recent research advances in both pipeline approaches and end-to-end models.

Pipeline methods

Natural language understanding (NLU)

This component parses a user’s utterance into a structured semantic representation, usually consisting of intent and slot-value pairs. Intent here indicates the function of the utterance, e.g. querying or providing information. Slot–value pairs are semantic elements mentioned in the utterance. For example, in the utterance “Can you recommend a Chinese restaurant in Manhattan?”, the slot–value pairs can be “cuisine” – “Chinese” and “location” – “Manhattan”.

Intent detection and slot–value extraction can be addressed using RNNs, CNNs, recursive neural networks, recurrent conditional random fields, or a BERT model. Zhang et al. (2020) also mention some more recent approaches to strengthening the connection between intent classification and slot tagging, e.g. Liu and Lane (2016) applied attention mechanisms to allow interaction between word and sentence representations, while Goo et al. (2018) used an intent gate to direct the slot tagging process. 

Dialog state tracking (DST) 

At this step of the pipeline, the dialog state tracker estimates the user’s goal by taking the entire dialog context as input. In most recent studies, the user’s goal is represented by slot–value pairs.

In 2013, Henderson et al. from the University of Cambridge introduced deep learning for dialog state tracking, where they used a sliding window to output a sequence of probability distributions over an arbitrary number of possible values. Fixed vocabulary for the slot allows applying classification for value prediction. For free-form slots, the modern approaches suggest either generating values directly or predicting the span of the value in the utterance.

Dialog state tracking can be also performed using RNNs and neural belief trackers (NBT), among other approaches.

Dialog policy learning

After the dialog state representation is defined, the dialog policy generates the next system action. Policy learning is usually optimized either with supervised learning or reinforcement learning. Sometimes a rule-based approach is first employed to warm-start the system.

Most of the modern approaches to dialog policy learning rely on reinforcement learning (RL). Cuayáhuitl et al. (2015) applied deep RL technique, where feature representations and dialog policy were learned simultaneously. However, training an RL policy requires lots of interactions with human users, which is time-consuming and costly. Thus, many approaches have been suggested to address this problem, including the use of user simulators. More recently, model-based RL approaches were suggested (Peng et al., 2018; Wu et al., 2019; Su et al., 2018), where the environment is modeled to simulate the dynamics of the conversation. The dialog policy is then trained alternately through learning from real users and planning with the environment model.

Natural language generation (NLG)

This component maps the dialog act generated by the dialog policy to a natural language utterance. The response should correspond to a dialog act to ensure task completion, and also be natural, specific, and informative.

This can be approached as a conditional language generation task. Recently, Peng et al. (2020) proposed to start by pre-training a GPT with a large-scale corpus, and then fine-tuning the model on target NLG tasks with a small number of training samples.

End-to-end methods

Chen et al. (2020) mention two key limitations of pipeline approaches:

  • The credit assignment problem, where the user’s feedback is hard to assign to a specific module of a pipeline
  • Process interdependence, where any changes to or retraining of one component require all the other components to be adapted accordingly

These issues can be addressed by building an end-to-end neural generative model for task-oriented dialog systems. Most of these end-to-end methods utilize seq2seq models.

However, the traditional approach requires thousands of dialogs to learn basic behaviors. To improve data efficiency, Hybrid Code Networks (HCNs) were introduced at ACL 2017. They combine an RNN with domain-specific knowledge encoded as software and system action templates. Furthermore, HCNs can be trained with supervised learning, reinforcement learning, or a mixture of both. Zhao and Eskenazi (2016) also proposed deploying deep reinforcement learning to train end-to-end task-oriented dialog systems. In particular, they used a variant of Deep Recurrent Q-Networks (DRQN).

Challenges

Despite the significant progress of recently introduced task-oriented dialog systems, there are still many issues that remain unsolved.

Data scarcity

Dialog is a complex NLP task, where a system needs to learn grammar, syntax, dialog reasoning, and language generation. Mastering all these skills requires lots of training data, and getting domain-specific data is quite expensive and time-consuming. Some of the recently introduced studies (Peng et al., 2020Budzianowski and Vulic, 2019) address this data scarcity problem by using pre-training models for their dialog agents to capture the implicit language knowledge from large-scale corpora. Then, the system is fine-tuned using the data for a specific dialog task.

Multi-turn dynamics

While open-domain dialog research primarily focuses on generating meaningful and consistent responses, task-oriented dialog agents should also be able to complete a specific task. Thus, the dialog management component of these dialog systems should aim at accurate user goal estimation and efficient dialog planning.

Knowledge base integration

Integrating external knowledge into end-to-end neural models is not a trivial task, since these methods don’t include explicit structured dialog state representation. The recent approaches have been to use memory networks for knowledge integration (Wu et al., 2019), to store knowledge base tuples in separate context-free memory (Lin et al., 2019), or to apply a two-step knowledge base retrieval technique (Qin et al., 2019). 

Robustness

Even the most sophisticated dialog systems are vulnerable to simple input perturbations. Models are usually trained with very little noise in the training data, while in real applications, users’ requests and responses can be out of domain and unexpected given the training data distribution. One of the possible solutions to this problem can be to combine rule-based methods with neural models (Liang et al., 2016).

Privacy protection

A single dialog agent often serves a large number of people. Moreover, it learns from interactions with these people, and thus can inadvertently store some sensitive information. It’s important to consider this issue when building task-oriented dialog systems.

Future research directions

Given the above mentioned challenges, some of the most appealing research directions include:

  • Addressing data scarcity by using pre-training models or other approaches
  • Increasing sample efficiency in policy learning via better dialog planning
  • Aiming at zero-shot domain adaptation, where a dialog agent can be applied in a new domain without any domain-specific training data
  • Improving the robustness of task-oriented dialog systems
  • Building end-to-end task-oriented dialog models without intermediate supervision

Many interesting research ideas addressing the key challenges of building task-oriented dialog agents have been presented at the top AI and NLP academic conferences this year. In our premium research summaries, we’ve curated and featured the most interesting research papers that introduce new approaches to policy learning, knowledge base integration, modeling long context, and more. Click here to purchase our Conversational AI Research summaries.

Enjoy this article? Sign up for more AI research updates.

We’ll let you know when we release more summary articles like this one.

AI

5 Work From Home Office Essentials

Working remotely from home had been increasing in popularity, but it’s now become a necessity for many professionals due to the pandemic. “Some companies are eager to reopen their doors and return to the office, but a large number of employer and employees are making the transitional work environment a permanent change.”  They can’t guarantee […]

The post 5 Work From Home Office Essentials appeared first on Aiiot Talk – Artificial Intelligence | Internet of Things | Technology.

Published

on

Working remotely from home had been increasing in popularity, but it’s now become a necessity for many professionals due to the pandemic.

“Some companies are eager to reopen their doors and return to the office, but a large number of employer and employees are making the transitional work environment a permanent change.” 

They can’t guarantee their health and safety in a socially-crowded space, plus, companies are able to save tons of money they would have spent on their commercial lease or mortgage payments.

That’s not the say that working from home doesn’t come at its own costs, however. It can lead to a huge hit in productivity without the right equipment in place. To maximize your performance and efficiency in a remote setting, be sure to purchase these five office essentials.

1. Powerful PC

This one probably feels like an obvious pointer, but let’s knock it off our list. You won’t be able to get by with a make-shift work station and in today’s digital domain, your computer will be at the core of everything you do.

Never-ending loading wheels, delayed downloads, and slow rendering will add seconds to every task you do, so if your company didn’t provide you with a workhorse computer tower, you might look into investing in one yourself, then deduct the cost in your tax return.

Depending on your line of work, it might make more sense to go for a laptop vs a desktop computer. Unless your tasks demand super sophisticated software and large storage space, you can probably get by with a portable PC. That way, when coffee shops begin to reopen and allow patrons to sit inside, you can work on-the-go without feeling tethered to your desk.

2. Ergonomic Office Chair

If you’re looking at a long-term remote situation, it’s worth spending the big bucks on an ergonomic office chair. You should feel comfortably locked into your seat for eight hours a day—at least if you want to concentrate on your workflow, rather than the cramp in your back.

Shop around for an office chair that’s sophisticated in design and specifically built to hold the human body. Some stand-out features you should look out for include:

  • Targeted support around the lumbar spine
  • Adjustable height so you can adjust the seat as necessary for your arms to rest naturally on the keyboard
  • Swivel base to effortlessly turn your body, preventing neck strain
  • Cushioned seat to comfort your tailbone
  • Ventilated fabric that promotes airflow so you don’t feel overheated when sitting in the chair for several hours

You might have to pay a couple of hundred dollars for the best-of-the-line features, but there is another item that might qualify as an eligible tax deduction—just be sure to keep all your receipts organized with a document scanner in case the IRS raises their eyebrows and issues an audit.

3. Wireless Keyboard

If you want to type faster and feel better while you’re at it, then a wireless keyboard is clutch. They enable you to bring the keys closer, decreasing the extension length of your arms and accompanying shoulder strain.

“It also helps reduce the strain on your eyes by moving the bright screen farther away from your direct line of sight.” 

And, last but not least, the keys are placed in an ergonomic position for a more natural finger splay, with ample cushioning wrist cushioning that helps prevent overuse injuries such as a carpal tunnel.

4. Noise-cancelling Headphones

To truly get in the zone, you should block out distractions with headphones the cancel noise in your environment—especially if your work station is set up in a common area. Other tips to stay focused include installing a website blocker and leaving your cellphone on the other side of the room.

5. House plant or flowers

People are scientifically proven to be more productive when working near fresh flowers or lush greenery. The good news is that you don’t need to have a green thumb or natural lighting to achieve this effect—even artificial foliage can brighten your mood and improve your performance.

Working from home sometimes can feel like you’re locked inside all day, so bringing the outside world inside your space can help ward off burnout.

Take these tips with you into 2021 and set yourself up for success in your new home office setting.

Source: https://www.aiiottalk.com/business/work-from-home-office-essentials/

Continue Reading

AI

zomato digitizes menus using Amazon Textract and Amazon SageMaker

This post is co-written by Chiranjeev Ghai, ML Engineer at zomato. zomato is a global food-tech company based in India. Are you the kind of person who has very specific cravings? Maybe when the mood hits, you don’t want just any kind of Indian food—you want Chicken Chettinad with a side of paratha, and nothing […]

Published

on

This post is co-written by Chiranjeev Ghai, ML Engineer at zomato. zomato is a global food-tech company based in India.

Are you the kind of person who has very specific cravings? Maybe when the mood hits, you don’t want just any kind of Indian food—you want Chicken Chettinad with a side of paratha, and nothing else will hit the spot! To help picky eaters satisfy their cravings, we at zomato have recently added enhanced search engine capabilities to our restaurant aggregation and food delivery platform. These capabilities enable us to recommend restaurants to zomato users based on searches for specific dishes.

We power this functionality with machine learning (ML), using it to extract and structure text data from menu images. To develop this menu digitization technology, we partnered with Amazon ML Solutions Lab to explore the capabilities of the AWS ML Stack. This post summarizes how we used Amazon Textract and Amazon SageMaker to develop a customized menu digitization solution.

Extracting raw text from menus with Amazon Textract

The first component of this solution was to accurately extract all the text in the menu image. This process is known as optical character recognition (OCR). For our use case, we experimented with both in-house and commercial OCR solutions.

We first created an in-house OCR solution by stacking a pre-trained text detection model and a pre-trained text recognition model. The challenge with these models was that they were trained on a standard text dataset that didn’t match the eclectic fonts found in restaurant menus. To improve system performance, we fine-tuned these models by generating a dataset of 1.5 million synthetic text images that were more representative of text in menus.

After evaluating our in-house solution and several commercial OCR solutions, we found that Amazon Textract offers the best text recognition precision and recall. Restaurants often get creative when designing their menus, so OCR robustness was crucial for this use case. Amazon Textract particularly differentiated itself when processing menus with unique fonts, background images, and low image resolutions. Using it is as simple as making an API call:

#Python 3.6
import boto3
textract_client = boto3.client( 'textract', region_name = '' #insert the AWS region you're working in
)
textract_response = textract_client.detect_document_text( Document={ 'S3Object': { 'Bucket': '', #insert the name of the S3 bucket containing your image 'Name': '' #insert the S3 key of your image } }
) print(textract_response)

The following code is the Amazon Textract output for a sample image:

{'DocumentMetadata': {'Pages': 1}, 'Blocks': [{'BlockType': 'PAGE', 'Geometry': {'BoundingBox': {'Width': 1.0, 'Height': 1.0, 'Left': 0.0, 'Top': 0.0}, ... {'BlockType': 'WORD', 'Text': 'Dim', 'Geometry': {'BoundingBox': {'Width': 0.10242128372192383, 'Height': 0. 048968635499477386, 'Left': 0. 24052166938781738, 'Top': 0. 02556285448372364},
... 

The raw outputs are visualized by overlaying them on top of the image. The following image visualizes the preceding raw output. The black boxes are the text-detection bounding boxes provided by Amazon Textract. Extracted text is displayed on the right. Note the unconventional fonts, colors, and images on this menu.

The following image visualizes Amazon Textract outputs for a menu with a different design. Black boxes are the text-detection bounding boxes provided by Amazon Textract. Extracted text is displayed on the right. Again, this menu has unconventional fonts, colors, and images.

Using Amazon SageMaker to build a menu structure detector

The next component of this solution was to group the detections from Amazon Textract by menu section. This enabled our search engine to distinguish between entrees, desserts, beverages, and so on. We framed this as a computer vision problem—object detection, to be precise—and used Amazon SageMaker Ground Truth to collect training data. Ground Truth accelerated this process by providing a fully managed annotation tool that we customized to ask human annotators to draw bounding boxes around every menu section in the image. We used an annotation workforce from AWS Marketplace because this was a niche labeling task, and public labelers from Amazon Mechanical Turk didn’t perform well. With Ground Truth, it took just a few days and approximately $1,400 to label 4,086 images with triplicate redundancy.

With labeled data in hand, we faced a paradox of choice when selecting model-building approaches because object detection is such a thoroughly studied problem. Our choices included:

  • Removing low-confidence labels from the labeled dataset – Because even human annotators can make mistakes, Ground Truth calculates confidence scores for labels by having multiple annotators (for this use case, three) label the same image. Setting a higher confidence threshold for labels can decrease the noise in the training data at the expense of having less training data.
  • Data augmentation – Techniques for image data augmentation include horizontal flipping, cropping, shearing, and rotation. Data augmentation can make models more robust by increasing the amount of training data. However, excessive data augmentation may result in poor model convergence.
  • Feature engineering – From our experience in applying computer vision to processing menus, we had a variety of techniques in mind to emphasize or de-emphasize various aspects of the input images. For example, see the following images.

The following is the original image of a menu.

The following image shows the redacted image (overlay white boxes on a black background where text detections were found).

The following is a text cropped image. On a black background, the image has overlay crops from the original image where text detections were found.

The following is a single channel and text cropped image. The image is encoded as a single RGB channel (for this image, green). You can apply this with other transformations, in this case text cropping.

 

We also had the following additional model-building methods to choose from:

  • Model architectures like YOLO, SSD, and RCNN, with VGG or ResNet backbones – Each architecture has different trade-offs of model accuracy, inference time, model size, and more. For this use case, model accuracy was the most important metric because menu images were batch processed.
  • Using a model pre-trained on a general object detection task or starting from scratch – Transfer learning can be helpful when training complex models on small datasets. However, the task of detecting menu sections is very different from a general object detection task (for example, PASCAL VOC), so the pre-training may not be relevant.
  • Optimizer parameters – These include learning rate, momentum, regularization coefficients, and early stopping configuration.

With so many hyperparameters to consider, we turned to the automatic tuning feature of Amazon SageMaker to coordinate a massive tuning job across all these variables. The following code is an example of tuning a single model architecture and input data configuration:

import sagemaker
import boto3
from sagemaker.amazon.amazon_estimator import get_image_uri
from sagemaker.estimator import Estimator
from sagemaker.tuner import HyperparameterTuner, IntegerParameter, CategoricalParameter, ContinuousParameter
import itertools
from time import sleep #set to the region you're working in
REGION_NAME = ''
#set a S3 path for SageMaker to store the outputs of the training jobs S3_OUTPUT_PATH = ''
#set a S3 location for your training dataset, #assumed to be an augmented manifest file
#see: https://docs.aws.amazon.com/sagemaker/latest/dg/augmented-manifest.html
TRAIN_DATA_LOCATION = ''
#set a S3 location for your validation data, #assumed to be an augmented manifest file
VAL_DATA_LOCATION = ''
#specify which fields in the augmented manifest file are relevant for training
DATA_ATTRIBUTE_NAMES = [,]
#specify image shape
IMAGE_SHAPE = #specify label width
LABEL_WIDTH = #specify number of samples in the training dataset
NUM_TRAINING_SAMPLES = sgm_role = sagemaker.get_execution_role()
boto_session = boto3.session.Session( region_name = REGION_NAME
)
sgm_session = sagemaker.Session( boto_session = boto_session
)
training_image = get_image_uri( region_name = REGION_NAME, repo_name = 'object-detection', repo_version = 'latest'
) #set training job configuration
object_detection_estimator = Estimator( image_name = training_image, role = sgm_role, train_instance_count = 1, train_instance_type = 'ml.p3.2xlarge', train_volume_size = 50, train_max_run = 360000, input_mode = 'Pipe', output_path = S3_OUTPUT_PATH, sagemaker_session = sgm_session
) #set input data configuration
train_data = sagemaker.session.s3_input( s3_data = TRAIN_DATA_LOCATION, distribution = 'FullyReplicated', record_wrapping = 'RecordIO', s3_data_type = 'AugmentedManifestFile', attribute_names = DATA_ATTRIBUTE_NAMES
) val_data = sagemaker.session.s3_input( s3_data = VAL_DATA_LOCATION, distribution = 'FullyReplicated', record_wrapping = 'RecordIO', s3_data_type = 'AugmentedManifestFile', attribute_names = DATA_ATTRIBUTE_NAMES
) data_channels = { 'train': train_data, 'validation' : val_data
} #set static hyperparameters
#see: https://docs.aws.amazon.com/sagemaker/latest/dg/object-detection-api-config.html
static_hyperparameters = { 'num_classes' : 1, 'epochs' : 100, 'lr_scheduler_step' : '15,30', 'lr_scheduler_factor' : 0.1, 'overlap_threshold' : 0.5, 'nms_threshold' : 0.45, 'image_shape' : IMAGE_SHAPE, 'label_width' : LABEL_WIDTH, 'num_training_samples' : NUM_TRAINING_SAMPLES, 'early_stopping' : True, 'early_stopping_min_epochs' : 5, 'early_stopping_patience' : 1, 'early_stopping_tolerance' : 0.05,
} #set ranges for tunable hyperparameters
hyperparameter_ranges = { 'learning_rate': ContinuousParameter( min_value = 1e-5, max_value = 1e-2, scaling_type = 'Auto' ), 'mini_batch_size': IntegerParameter( min_value = 8, max_value = 64, scaling_type = 'Auto' )
} #Not all hyperparameters are feasible to tune directly
#see: https://docs.aws.amazon.com/sagemaker/latest/dg/object-detection-tuning.html
#For these we run model tuning jobs in parallel using a for loop
#We take this approach for tuning over different model architectures #and different feature engineering configurations
use_pretrained_options = [0, 1]
base_network_options = ['resnet-50', 'vgg-16'] for use_pretrained, base_network in itertools.product(use_pretrained_options, base_network_options): static_hyperparameter_configuration = { **static_hyperparameters, 'use_pretrained_model' : use_pretrained, 'base_network' : base_network } object_detection_estimator.set_hyperparameters( **static_hyperparameter_configuration ) tuner = HyperparameterTuner( estimator = object_detection_estimator, objective_metric_name = 'validation:mAP', strategy = 'Bayesian', hyperparameter_ranges = hyperparameter_ranges, max_jobs = 24, max_parallel_jobs = 2, early_stopping_type = 'Auto', ) tuner.fit( inputs = data_channels ) print(f'Started tuning job: {tuner.latest_tuning_job.name}') #wait a bit before starting next job so auto generated names don't conflict sleep(60)

This code uses version 1.72.0 of the Amazon SageMaker Python SDK, which is the default version installed in Amazon SageMaker notebook instances. Version 2.X introduces breaking changes. For more information, see Use Version 2.x of the SageMaker Python SDK.

We used powerful GPU hardware (p3.2xlarge instances), and it took us just 1 week and approximately $1,500 to explore 455 unique parameter configurations. Of these configurations, Amazon SageMaker found that a fine-tuned Faster R-CNN model with text cropping performed the best, with a mean average precision score of 0.93. This aligned with results from our prior work in this space, which found that two-stage detectors generally outperform single-stage detectors in processing menus.

The following is an example of how the object detection model processed a menu. In this image, the purple boxes are the predicted bounding boxes from the menu section detection model. Black boxes are the text detection bounding boxes provided by Amazon Textract.

Using Amazon SageMaker to build rule- and ML-based text classifiers

The final component in the solution was a layer of text classification. To enable our enhanced search functionality, we had to know if each detection within a menu section was the menu section title, name of a dish, price of a dish, or something else (such as a description of a dish or the name of the restaurant). To this end, we developed a hybrid rule- and ML-based text classification system.

The first step of the classification was to use a rule to determine if a detection was a price or not. This rule simply calculated the proportion of numeric characters in the detection. If the proportion was greater than 40%, the detection was classified as a price. Although simple, this classifier worked well in practice. We used Amazon SageMaker notebook instances as a convenient interactive environment to develop this and other rules.

After the prices were filtered out, the remaining detections were classified as dish or not dish. From our experience in processing menus, we intuitively knew that in many cases, the location of prices was sufficient to do this classification. For these menus, dishes and prices are listed side by side, so simply classifying detections located to the left of prices as dishes worked well.

The following example shows how the rules-based text classification system processed a menu. Green boxes are detections classified as dishes (by the price location rule). Red boxes are detections classified as not dishes (by the price location rule). Blue boxes are detections classified as prices. Final dish detections are on the right.

Some menus might include lengthy dish descriptions or may not list prices next to individual dishes. These menus violate the assumptions of the price location rules, so we turned to model-based text classification. We used Amazon SageMaker training jobs to experiment with many modeling approaches in parallel, including an XGBoost model trained on hashed word count vectors. In the end, we found that a fine-tuned BERT model from GluonNLP achieved the best performance with an AUROC score of 0.86.

The following image is an example of how the model-based text classification system processed a menu. Green boxes are detections classified as dishes (by the BERT model). Red boxes are detections classified as not dishes (by the BERT model). Blue boxes are detections classified as prices. The final dish detections are on the right.

Of the remaining detections (those not classified as prices or dishes), a final round of classification identified menu section titles. We created features that captured the font size of the detection, the location of the detection on the menu, and the length of the words within the detection. We used these features as inputs to a logistic regression model that predicted if a detection is a menu section title or not.

Key features of Amazon SageMaker

In the end, we found that doing OCR was as simple as making an API call to Amazon Textract. However, our use case required additional customization. We selected Amazon SageMaker as an ML platform to develop this customization because it offered several key features:

  • Amazon SageMaker Notebooks made it easy to spin up Jupyter notebook environments for prototyping and testing rules and models.
  • Ground Truth helped us build and deploy a custom image annotation tool with no front-end experience required.
  • Amazon SageMaker automatic tuning enabled us to run massive hyperparameter tuning jobs on powerful hardware, and included an intuitive interface for tracking the results of hundreds of experiments. You can implement tuning jobs with early stopping conditions, which makes experimentation cost-effective.

Amazon SageMaker offers additional integration benefits from including all the preceding features in a single platform:

  • Amazon SageMaker Notebooks come pre-installed with all the dependencies needed to build models that can be optimized with automatic tuning.
  • Ground Truth offers easy access to labelers from Mechanical Turk or AWS Marketplace.
  • Automatic tuning can directly ingest the manifest files created by Amazon SageMaker Ground Truth.

Putting it all together

Our menu digitization system can extract text from images of menus, group it by menu section, extract the title of the section, extract the dishes within each section, and pair each dish with its price. The following is a visualization of the end-to-end solution.

The workflow contains the following steps:

  1. The input is an image of a menu.
  2. Amazon Textract performs OCR on the input image.
  3. An ML-based computer vision model predicts bounding boxes for menu sections in the menu image.
  4. A rules-based classifier classifies Amazon Textract detections as price or not price.
  5. A rules-based classifier (5a) attempts to use the location of price detections to classify the not price detections as dish or not dish. If this rule doesn’t successfully classify most of the detections on the page, an ML-based classifier is used instead (5b).
  6. The ML-based classifier uses hand-crafted features to classify not dish detections as menu section title or not menu section title.
  7.  The menu text is structured by combining the menu section detections and the text classification results.

The following image visualizes a sample output of the system. Green boxes are detections classified as dishes. Blue boxes are detections classified as prices. Yellow boxes are detections classified as menu section titles. Purple boxes are predicted menu section bounding boxes.

The following code is the structured output:

[ { "title":{ "text":"Shrimp Dishes" }, "dishes":[ { "text":"Shrimp Masala", "price":{ "text":"140" } }, { "text":"Shrimp Biryani", "price":{ "text":"170" } }, { "text":"Shrimp Pulav", "price":{ "text":"160" } } ] }, ...
]

Conclusion

We built a system that uses ML to digitize menus without any human input required. This system will improve user experience by powering new features such as advanced dish search and review highlight verification. Our content team will also use it to accelerate creating menus for online ordering.

To explore these capabilities of Amazon Textract and Amazon SageMaker in more depth, see Automatically extract text and structured data from documents with Amazon Textract and Amazon SageMaker Automatic Model Tuning: Using Machine Learning for Machine Learning.

The Amazon ML Solutions Lab helped us accelerate our use of ML by pairing our team with ML experts. The ML Solutions Lab brings to every customer engagement learnings from more than 20 years of Amazon’s ML innovations in areas such as fulfillment and logistics, personalization and recommendations, computer vision and translation, fraud prevention, forecasting, and supply chain optimization. To learn more about the AWS ML Solutions Lab, contact your account manager or visit Amazon Machine Learning Solutions Lab.


About the Authors

Chiranjeev Ghai is a Machine Learning Engineer. In his current role, he has been aiding automation at zomato by leveraging a wide variety of ML optimisations ranging from Image Classification, Product Recommendation, and Text Detection. When not building models, he likes to spend his time playing video games at home.

Ryan Cheng is a Deep Learning Architect in the Amazon ML Solutions Lab. He has worked on a wide range of ML use cases from sports analytics to optical character recognition. In his spare time, Ryan enjoys cooking.

Andrew Ang is a Deep Learning Architect at the Amazon ML Solutions Lab, where he helps AWS customers identify and build AI/ML solutions to address their business problems.

Vinayak Arannil is a Data Scientist at the Amazon Machine Learning Solutions Lab. He has worked on various domains of data science like computer vision, natural language processing, recommendation systems, etc.

Source: https://aws.amazon.com/blogs/machine-learning/zomato-digitizes-menus-using-amazon-textract-and-amazon-sagemaker/

Continue Reading

AI

zomato digitizes menus using Amazon Textract and Amazon SageMaker

This post is co-written by Chiranjeev Ghai, ML Engineer at zomato. zomato is a global food-tech company based in India. Are you the kind of person who has very specific cravings? Maybe when the mood hits, you don’t want just any kind of Indian food—you want Chicken Chettinad with a side of paratha, and nothing […]

Published

on

This post is co-written by Chiranjeev Ghai, ML Engineer at zomato. zomato is a global food-tech company based in India.

Are you the kind of person who has very specific cravings? Maybe when the mood hits, you don’t want just any kind of Indian food—you want Chicken Chettinad with a side of paratha, and nothing else will hit the spot! To help picky eaters satisfy their cravings, we at zomato have recently added enhanced search engine capabilities to our restaurant aggregation and food delivery platform. These capabilities enable us to recommend restaurants to zomato users based on searches for specific dishes.

We power this functionality with machine learning (ML), using it to extract and structure text data from menu images. To develop this menu digitization technology, we partnered with Amazon ML Solutions Lab to explore the capabilities of the AWS ML Stack. This post summarizes how we used Amazon Textract and Amazon SageMaker to develop a customized menu digitization solution.

Extracting raw text from menus with Amazon Textract

The first component of this solution was to accurately extract all the text in the menu image. This process is known as optical character recognition (OCR). For our use case, we experimented with both in-house and commercial OCR solutions.

We first created an in-house OCR solution by stacking a pre-trained text detection model and a pre-trained text recognition model. The challenge with these models was that they were trained on a standard text dataset that didn’t match the eclectic fonts found in restaurant menus. To improve system performance, we fine-tuned these models by generating a dataset of 1.5 million synthetic text images that were more representative of text in menus.

After evaluating our in-house solution and several commercial OCR solutions, we found that Amazon Textract offers the best text recognition precision and recall. Restaurants often get creative when designing their menus, so OCR robustness was crucial for this use case. Amazon Textract particularly differentiated itself when processing menus with unique fonts, background images, and low image resolutions. Using it is as simple as making an API call:

#Python 3.6
import boto3
textract_client = boto3.client( 'textract', region_name = '' #insert the AWS region you're working in
)
textract_response = textract_client.detect_document_text( Document={ 'S3Object': { 'Bucket': '', #insert the name of the S3 bucket containing your image 'Name': '' #insert the S3 key of your image } }
) print(textract_response)

The following code is the Amazon Textract output for a sample image:

{'DocumentMetadata': {'Pages': 1}, 'Blocks': [{'BlockType': 'PAGE', 'Geometry': {'BoundingBox': {'Width': 1.0, 'Height': 1.0, 'Left': 0.0, 'Top': 0.0}, ... {'BlockType': 'WORD', 'Text': 'Dim', 'Geometry': {'BoundingBox': {'Width': 0.10242128372192383, 'Height': 0. 048968635499477386, 'Left': 0. 24052166938781738, 'Top': 0. 02556285448372364},
... 

The raw outputs are visualized by overlaying them on top of the image. The following image visualizes the preceding raw output. The black boxes are the text-detection bounding boxes provided by Amazon Textract. Extracted text is displayed on the right. Note the unconventional fonts, colors, and images on this menu.

The following image visualizes Amazon Textract outputs for a menu with a different design. Black boxes are the text-detection bounding boxes provided by Amazon Textract. Extracted text is displayed on the right. Again, this menu has unconventional fonts, colors, and images.

Using Amazon SageMaker to build a menu structure detector

The next component of this solution was to group the detections from Amazon Textract by menu section. This enabled our search engine to distinguish between entrees, desserts, beverages, and so on. We framed this as a computer vision problem—object detection, to be precise—and used Amazon SageMaker Ground Truth to collect training data. Ground Truth accelerated this process by providing a fully managed annotation tool that we customized to ask human annotators to draw bounding boxes around every menu section in the image. We used an annotation workforce from AWS Marketplace because this was a niche labeling task, and public labelers from Amazon Mechanical Turk didn’t perform well. With Ground Truth, it took just a few days and approximately $1,400 to label 4,086 images with triplicate redundancy.

With labeled data in hand, we faced a paradox of choice when selecting model-building approaches because object detection is such a thoroughly studied problem. Our choices included:

  • Removing low-confidence labels from the labeled dataset – Because even human annotators can make mistakes, Ground Truth calculates confidence scores for labels by having multiple annotators (for this use case, three) label the same image. Setting a higher confidence threshold for labels can decrease the noise in the training data at the expense of having less training data.
  • Data augmentation – Techniques for image data augmentation include horizontal flipping, cropping, shearing, and rotation. Data augmentation can make models more robust by increasing the amount of training data. However, excessive data augmentation may result in poor model convergence.
  • Feature engineering – From our experience in applying computer vision to processing menus, we had a variety of techniques in mind to emphasize or de-emphasize various aspects of the input images. For example, see the following images.

The following is the original image of a menu.

The following image shows the redacted image (overlay white boxes on a black background where text detections were found).

The following is a text cropped image. On a black background, the image has overlay crops from the original image where text detections were found.

The following is a single channel and text cropped image. The image is encoded as a single RGB channel (for this image, green). You can apply this with other transformations, in this case text cropping.

 

We also had the following additional model-building methods to choose from:

  • Model architectures like YOLO, SSD, and RCNN, with VGG or ResNet backbones – Each architecture has different trade-offs of model accuracy, inference time, model size, and more. For this use case, model accuracy was the most important metric because menu images were batch processed.
  • Using a model pre-trained on a general object detection task or starting from scratch – Transfer learning can be helpful when training complex models on small datasets. However, the task of detecting menu sections is very different from a general object detection task (for example, PASCAL VOC), so the pre-training may not be relevant.
  • Optimizer parameters – These include learning rate, momentum, regularization coefficients, and early stopping configuration.

With so many hyperparameters to consider, we turned to the automatic tuning feature of Amazon SageMaker to coordinate a massive tuning job across all these variables. The following code is an example of tuning a single model architecture and input data configuration:

import sagemaker
import boto3
from sagemaker.amazon.amazon_estimator import get_image_uri
from sagemaker.estimator import Estimator
from sagemaker.tuner import HyperparameterTuner, IntegerParameter, CategoricalParameter, ContinuousParameter
import itertools
from time import sleep #set to the region you're working in
REGION_NAME = ''
#set a S3 path for SageMaker to store the outputs of the training jobs S3_OUTPUT_PATH = ''
#set a S3 location for your training dataset, #assumed to be an augmented manifest file
#see: https://docs.aws.amazon.com/sagemaker/latest/dg/augmented-manifest.html
TRAIN_DATA_LOCATION = ''
#set a S3 location for your validation data, #assumed to be an augmented manifest file
VAL_DATA_LOCATION = ''
#specify which fields in the augmented manifest file are relevant for training
DATA_ATTRIBUTE_NAMES = [,]
#specify image shape
IMAGE_SHAPE = #specify label width
LABEL_WIDTH = #specify number of samples in the training dataset
NUM_TRAINING_SAMPLES = sgm_role = sagemaker.get_execution_role()
boto_session = boto3.session.Session( region_name = REGION_NAME
)
sgm_session = sagemaker.Session( boto_session = boto_session
)
training_image = get_image_uri( region_name = REGION_NAME, repo_name = 'object-detection', repo_version = 'latest'
) #set training job configuration
object_detection_estimator = Estimator( image_name = training_image, role = sgm_role, train_instance_count = 1, train_instance_type = 'ml.p3.2xlarge', train_volume_size = 50, train_max_run = 360000, input_mode = 'Pipe', output_path = S3_OUTPUT_PATH, sagemaker_session = sgm_session
) #set input data configuration
train_data = sagemaker.session.s3_input( s3_data = TRAIN_DATA_LOCATION, distribution = 'FullyReplicated', record_wrapping = 'RecordIO', s3_data_type = 'AugmentedManifestFile', attribute_names = DATA_ATTRIBUTE_NAMES
) val_data = sagemaker.session.s3_input( s3_data = VAL_DATA_LOCATION, distribution = 'FullyReplicated', record_wrapping = 'RecordIO', s3_data_type = 'AugmentedManifestFile', attribute_names = DATA_ATTRIBUTE_NAMES
) data_channels = { 'train': train_data, 'validation' : val_data
} #set static hyperparameters
#see: https://docs.aws.amazon.com/sagemaker/latest/dg/object-detection-api-config.html
static_hyperparameters = { 'num_classes' : 1, 'epochs' : 100, 'lr_scheduler_step' : '15,30', 'lr_scheduler_factor' : 0.1, 'overlap_threshold' : 0.5, 'nms_threshold' : 0.45, 'image_shape' : IMAGE_SHAPE, 'label_width' : LABEL_WIDTH, 'num_training_samples' : NUM_TRAINING_SAMPLES, 'early_stopping' : True, 'early_stopping_min_epochs' : 5, 'early_stopping_patience' : 1, 'early_stopping_tolerance' : 0.05,
} #set ranges for tunable hyperparameters
hyperparameter_ranges = { 'learning_rate': ContinuousParameter( min_value = 1e-5, max_value = 1e-2, scaling_type = 'Auto' ), 'mini_batch_size': IntegerParameter( min_value = 8, max_value = 64, scaling_type = 'Auto' )
} #Not all hyperparameters are feasible to tune directly
#see: https://docs.aws.amazon.com/sagemaker/latest/dg/object-detection-tuning.html
#For these we run model tuning jobs in parallel using a for loop
#We take this approach for tuning over different model architectures #and different feature engineering configurations
use_pretrained_options = [0, 1]
base_network_options = ['resnet-50', 'vgg-16'] for use_pretrained, base_network in itertools.product(use_pretrained_options, base_network_options): static_hyperparameter_configuration = { **static_hyperparameters, 'use_pretrained_model' : use_pretrained, 'base_network' : base_network } object_detection_estimator.set_hyperparameters( **static_hyperparameter_configuration ) tuner = HyperparameterTuner( estimator = object_detection_estimator, objective_metric_name = 'validation:mAP', strategy = 'Bayesian', hyperparameter_ranges = hyperparameter_ranges, max_jobs = 24, max_parallel_jobs = 2, early_stopping_type = 'Auto', ) tuner.fit( inputs = data_channels ) print(f'Started tuning job: {tuner.latest_tuning_job.name}') #wait a bit before starting next job so auto generated names don't conflict sleep(60)

This code uses version 1.72.0 of the Amazon SageMaker Python SDK, which is the default version installed in Amazon SageMaker notebook instances. Version 2.X introduces breaking changes. For more information, see Use Version 2.x of the SageMaker Python SDK.

We used powerful GPU hardware (p3.2xlarge instances), and it took us just 1 week and approximately $1,500 to explore 455 unique parameter configurations. Of these configurations, Amazon SageMaker found that a fine-tuned Faster R-CNN model with text cropping performed the best, with a mean average precision score of 0.93. This aligned with results from our prior work in this space, which found that two-stage detectors generally outperform single-stage detectors in processing menus.

The following is an example of how the object detection model processed a menu. In this image, the purple boxes are the predicted bounding boxes from the menu section detection model. Black boxes are the text detection bounding boxes provided by Amazon Textract.

Using Amazon SageMaker to build rule- and ML-based text classifiers

The final component in the solution was a layer of text classification. To enable our enhanced search functionality, we had to know if each detection within a menu section was the menu section title, name of a dish, price of a dish, or something else (such as a description of a dish or the name of the restaurant). To this end, we developed a hybrid rule- and ML-based text classification system.

The first step of the classification was to use a rule to determine if a detection was a price or not. This rule simply calculated the proportion of numeric characters in the detection. If the proportion was greater than 40%, the detection was classified as a price. Although simple, this classifier worked well in practice. We used Amazon SageMaker notebook instances as a convenient interactive environment to develop this and other rules.

After the prices were filtered out, the remaining detections were classified as dish or not dish. From our experience in processing menus, we intuitively knew that in many cases, the location of prices was sufficient to do this classification. For these menus, dishes and prices are listed side by side, so simply classifying detections located to the left of prices as dishes worked well.

The following example shows how the rules-based text classification system processed a menu. Green boxes are detections classified as dishes (by the price location rule). Red boxes are detections classified as not dishes (by the price location rule). Blue boxes are detections classified as prices. Final dish detections are on the right.

Some menus might include lengthy dish descriptions or may not list prices next to individual dishes. These menus violate the assumptions of the price location rules, so we turned to model-based text classification. We used Amazon SageMaker training jobs to experiment with many modeling approaches in parallel, including an XGBoost model trained on hashed word count vectors. In the end, we found that a fine-tuned BERT model from GluonNLP achieved the best performance with an AUROC score of 0.86.

The following image is an example of how the model-based text classification system processed a menu. Green boxes are detections classified as dishes (by the BERT model). Red boxes are detections classified as not dishes (by the BERT model). Blue boxes are detections classified as prices. The final dish detections are on the right.

Of the remaining detections (those not classified as prices or dishes), a final round of classification identified menu section titles. We created features that captured the font size of the detection, the location of the detection on the menu, and the length of the words within the detection. We used these features as inputs to a logistic regression model that predicted if a detection is a menu section title or not.

Key features of Amazon SageMaker

In the end, we found that doing OCR was as simple as making an API call to Amazon Textract. However, our use case required additional customization. We selected Amazon SageMaker as an ML platform to develop this customization because it offered several key features:

  • Amazon SageMaker Notebooks made it easy to spin up Jupyter notebook environments for prototyping and testing rules and models.
  • Ground Truth helped us build and deploy a custom image annotation tool with no front-end experience required.
  • Amazon SageMaker automatic tuning enabled us to run massive hyperparameter tuning jobs on powerful hardware, and included an intuitive interface for tracking the results of hundreds of experiments. You can implement tuning jobs with early stopping conditions, which makes experimentation cost-effective.

Amazon SageMaker offers additional integration benefits from including all the preceding features in a single platform:

  • Amazon SageMaker Notebooks come pre-installed with all the dependencies needed to build models that can be optimized with automatic tuning.
  • Ground Truth offers easy access to labelers from Mechanical Turk or AWS Marketplace.
  • Automatic tuning can directly ingest the manifest files created by Amazon SageMaker Ground Truth.

Putting it all together

Our menu digitization system can extract text from images of menus, group it by menu section, extract the title of the section, extract the dishes within each section, and pair each dish with its price. The following is a visualization of the end-to-end solution.

The workflow contains the following steps:

  1. The input is an image of a menu.
  2. Amazon Textract performs OCR on the input image.
  3. An ML-based computer vision model predicts bounding boxes for menu sections in the menu image.
  4. A rules-based classifier classifies Amazon Textract detections as price or not price.
  5. A rules-based classifier (5a) attempts to use the location of price detections to classify the not price detections as dish or not dish. If this rule doesn’t successfully classify most of the detections on the page, an ML-based classifier is used instead (5b).
  6. The ML-based classifier uses hand-crafted features to classify not dish detections as menu section title or not menu section title.
  7.  The menu text is structured by combining the menu section detections and the text classification results.

The following image visualizes a sample output of the system. Green boxes are detections classified as dishes. Blue boxes are detections classified as prices. Yellow boxes are detections classified as menu section titles. Purple boxes are predicted menu section bounding boxes.

The following code is the structured output:

[ { "title":{ "text":"Shrimp Dishes" }, "dishes":[ { "text":"Shrimp Masala", "price":{ "text":"140" } }, { "text":"Shrimp Biryani", "price":{ "text":"170" } }, { "text":"Shrimp Pulav", "price":{ "text":"160" } } ] }, ...
]

Conclusion

We built a system that uses ML to digitize menus without any human input required. This system will improve user experience by powering new features such as advanced dish search and review highlight verification. Our content team will also use it to accelerate creating menus for online ordering.

To explore these capabilities of Amazon Textract and Amazon SageMaker in more depth, see Automatically extract text and structured data from documents with Amazon Textract and Amazon SageMaker Automatic Model Tuning: Using Machine Learning for Machine Learning.

The Amazon ML Solutions Lab helped us accelerate our use of ML by pairing our team with ML experts. The ML Solutions Lab brings to every customer engagement learnings from more than 20 years of Amazon’s ML innovations in areas such as fulfillment and logistics, personalization and recommendations, computer vision and translation, fraud prevention, forecasting, and supply chain optimization. To learn more about the AWS ML Solutions Lab, contact your account manager or visit Amazon Machine Learning Solutions Lab.


About the Authors

Chiranjeev Ghai is a Machine Learning Engineer. In his current role, he has been aiding automation at zomato by leveraging a wide variety of ML optimisations ranging from Image Classification, Product Recommendation, and Text Detection. When not building models, he likes to spend his time playing video games at home.

Ryan Cheng is a Deep Learning Architect in the Amazon ML Solutions Lab. He has worked on a wide range of ML use cases from sports analytics to optical character recognition. In his spare time, Ryan enjoys cooking.

Andrew Ang is a Deep Learning Architect at the Amazon ML Solutions Lab, where he helps AWS customers identify and build AI/ML solutions to address their business problems.

Vinayak Arannil is a Data Scientist at the Amazon Machine Learning Solutions Lab. He has worked on various domains of data science like computer vision, natural language processing, recommendation systems, etc.

Source: https://aws.amazon.com/blogs/machine-learning/zomato-digitizes-menus-using-amazon-textract-and-amazon-sagemaker/

Continue Reading
AI1 hour ago

5 Work From Home Office Essentials

AI2 hours ago

zomato digitizes menus using Amazon Textract and Amazon SageMaker

AI2 hours ago

zomato digitizes menus using Amazon Textract and Amazon SageMaker

AI2 hours ago

zomato digitizes menus using Amazon Textract and Amazon SageMaker

AI2 hours ago

How to Improve Your Supply Chain With Deep Reinforcement Learning

AI2 hours ago

Video streaming and deep learning: Using Amazon Kinesis Video Streams with Deep Java Library

AI2 hours ago

Video streaming and deep learning: Using Amazon Kinesis Video Streams with Deep Java Library

AI2 hours ago

Video streaming and deep learning: Using Amazon Kinesis Video Streams with Deep Java Library

AI2 hours ago

Video streaming and deep learning: Using Amazon Kinesis Video Streams with Deep Java Library

AI5 hours ago

Conversation Designers: who are they and what do they do?

AI5 hours ago

Automating Bot Testing at Haptik

AI7 hours ago

Why Facebook’s New Machine Translation Model is a Great Step for AI

AI1 day ago

Bringing real-time machine learning-powered insights to rugby using Amazon SageMaker

AI1 day ago

Bringing real-time machine learning-powered insights to rugby using Amazon SageMaker

AI1 day ago

Bringing real-time machine learning-powered insights to rugby using Amazon SageMaker

AI1 day ago

Bringing real-time machine learning-powered insights to rugby using Amazon SageMaker

AI1 day ago

Bringing real-time machine learning-powered insights to rugby using Amazon SageMaker

AI1 day ago

Bringing real-time machine learning-powered insights to rugby using Amazon SageMaker

AI1 day ago

Bringing real-time machine learning-powered insights to rugby using Amazon SageMaker

AI1 day ago

Bringing real-time machine learning-powered insights to rugby using Amazon SageMaker

Trending