Connect with us

AI

The “why” behind the Multilingual AI Agent Assistants (MAIA) project: A conversation with Dr. Graham Neubig

Unbabel recently announced that the company’s AI research team is partnering with Carnegie Mellon University, INESC-ID, and Instituto de Telecomunicações. The goal? Reduce language-based borders by making multilingual chat work better. The MAIA: Multilingual AI Agent Assistants large scale research project will augment customer service agents with AI, making it more efficient for enterprises to …

The post The “why” behind the Multilingual AI Agent Assistants (MAIA) project: A conversation with Dr. Graham Neubig appeared first on Unbabel.

Published

on

Unbabel recently announced that the company’s AI research team is partnering with Carnegie Mellon University, INESC-ID, and Instituto de Telecomunicações. The goal? Reduce language-based borders by making multilingual chat work better.

The MAIA: Multilingual AI Agent Assistants large scale research project will augment customer service agents with AI, making it more efficient for enterprises to deliver chat in 30 languages and improve customer satisfaction through human empathy.

We at Unbabel are working closely with universities and researchers like Dr. Graham Neubig on MAIA. We know customer satisfaction requires human empathy, and speaking the customer’s language is a powerful way to demonstrate this.

Recently, we sat down with Dr. Neubig to learn more about his background, what led him to his field of research, why he’s excited about MAIA and the biggest challenges on the road ahead. Read more below.

Interview with Dr. Graham Neubig of Carnegie Mellon University

Unbabel: How many languages do you speak? 

Dr. Graham Neubig: I speak two languages. English is my native language. I learned Japanese by studying it in university and then living in Japan for 11 years. I also can read a bit of Chinese, Korean, and Spanish.

Unbabel: How has your experience led you to the research you do?

Dr. Graham Neubig: When I started university, I was most interested in music processing. I didn’t really get interested in natural language processing until I studied abroad in Japan my junior year and started learning the language. That really sparked my interest, and it has grown from there.

Unbabel: Your areas of research include machine learning approaches that are both linguistically motivated and tailored to applications (such as machine translation and natural language understanding). What does “linguistically motivated” mean to you?

Dr. Graham Neubig: Basically, it means taking into consideration what we know about various aspects of linguistics: phonology, syntax, semantics. For example, we know that all humans produce speech using the articulators (anatomical parts of their mouths and throats). So all speech in different languages is going to be similar in some ways. We also know that all languages have some variety of recursive structure. And we know that languages evolve from each other and naturally share many similarities. We can use all of this linguistic knowledge to build more effective machine learning models.

Unbabel: What drove your interest in the MAIA research project specifically? 

Dr. Graham Neubig: I’m very interested in communication in the conversational context. Conversation poses some unique challenges and opportunities. For example, conversations are highly context-dependent. It may not be clear how to translate a particular sentence unless you have a model that can handle the whole, broader context. 

Also, you need to hit the appropriate register. In other words, you don’t want to be talking to a business customer using slang or impolite expressions. Plus, conversations have active participants. We may be able to leverage this fact to inform the model about how it should be translating some terms or to provide guidance about when it’s doing well or doing poorly.

Unbabel: Why do you think it’s important for academics and the industry to partner up when it comes to this type of research? 

Dr. Graham Neubig: It’s great to be able to partner with Unbabel! I know and greatly respect many of the technical people they have on the team, and it’s great to see all the success they’ve had through innovation on the business side as well.

On a broader level, joint projects between academia and the industry are interesting, because the industry often has complex, real-world problems that beg to be solved on a faster timeline. In academia, we can often take a slightly more deliberate pace to break problems down to their essences, formulate them cleanly, and come up with long-term solutions. It’s nice to partner up, because it forces us to strike a balance.

Unbabel: Speaking of balance, how do you think about balancing the theoretical with the practical when it comes to your research?

Dr. Graham Neubig: My personal research style is often to start with a practical problem, get a good understanding of the problem, and then try to generalize or simplify the problem setting. Then we can run carefully designed experiments that may allow us to test general theories. In that way, I think I’m on the more practical side as far as academic researchers go. I really like when we can solve big problems that affect many practical situations in one fell swoop.

Unbabel: What are the hottest frontiers in natural language processing right now, in your opinion? 

Dr. Graham Neubig: We’re in a very exciting time for NLP right now, with large-scale neural models consistently improving accuracy and expanding the boundaries of what we can do. Long-tail languages are certainly an issue close to my heart, as well as models that are more specific to different topical domains or dialects. In addition, bias, model interpretability, and further expanding the application scenarios of NLP technology are all fascinating topics to me.

Unbabel: What are some of the most difficult languages to build NLP models for? Why?

Dr. Graham Neubig: Given the current technology, languages with fewer resources are hardest to build models for, especially when they are not very similar to other languages that have more resources. For languages that have few resources but are similar lexically or syntactically to another language with more resources, we now have reasonably good tools. We can use data from the higher-resourced language to improve accuracy on the lower-resourced languages. There are certainly some factors of languages that make them a bit harder to model (e.g. complicated morphology, when information is left implicit, etc.) However, it seems that, with our current models, presence or absence of lots of data really is the overwhelming factor.

Unbabel: What real-world applications of NLP do you find the most interesting or important and why?

Dr. Graham Neubig: With regards to important, there are widely used ones, such as machine translation, speech recognition, chatbots, and questioning answering. But I’m also interested in some more esoteric or emerging applications, such as natural language programming (where humans command computers in natural language vs. code), and language learning applications for low-resource and/or endangered languages.

The MAIA Project: Borderless customer conversations

Over the course of 36 months, the research team proposes to build a toolbox of machine learning technologies for online multilingual customer service, which includes context-aware machine translation, automatic answer generation, and conversational quality estimation. 

This will significantly expand the current portfolio of Unbabel AI technologies, creating a new model: an agent assistant that will facilitate the communication between human agents and international customers, making live chat customer service platforms more multilingual, scalable and capable of ensuring higher customer satisfaction. 

It will also help agents with new dialogue-oriented productivity tools, and expand Unbabel’s renowned quality estimation technology to assess conversational quality.  

We look forward to reporting back on our progress together along the way.

AI

Graph Convolutional Networks (GCN)

In this post, we’re gonna take a close look at one of the well-known graph neural networks named Graph Convolutional Network (GCN). First, we’ll get the intuition to see how it works, then we’ll go deeper into the maths behind it. Why Graphs? Many problems are graphs in true nature. In our world, we see many data are graphs, […]

The post Graph Convolutional Networks (GCN) appeared first on TOPBOTS.

Published

on

graph convolutional networks

In this post, we’re gonna take a close look at one of the well-known graph neural networks named Graph Convolutional Network (GCN). First, we’ll get the intuition to see how it works, then we’ll go deeper into the maths behind it.

Why Graphs?

Many problems are graphs in true nature. In our world, we see many data are graphs, such as molecules, social networks, and paper citations networks.

Tasks on Graphs

  • Node classification: Predict a type of a given node
  • Link prediction: Predict whether two nodes are linked
  • Community detection: Identify densely linked clusters of nodes
  • Network similarity: How similar are two (sub)networks

Machine Learning Lifecycle

In the graph, we have node features (the data of nodes) and the structure of the graph (how nodes are connected).

For the former, we can easily get the data from each node. But when it comes to the structure, it is not trivial to extract useful information from it. For example, if 2 nodes are close to one another, should we treat them differently to other pairs? How about high and low degree nodes? In fact, each specific task can consume a lot of time and effort just for Feature Engineering, i.e., to distill the structure into our features.

graph convolutional network
Feature engineering on graphs. (Picture from [1])

It would be much better to somehow get both the node features and the structure as the input, and let the machine to figure out what information is useful by itself.

That’s why we need Graph Representation Learning.

graph convolutional network
We want the graph can learn the “feature engineering” by itself. (Picture from [1])

If this in-depth educational content on convolutional neural networks is useful for you, you can subscribe to our AI research mailing list to be alerted when we release new material. 

Graph Convolutional Networks (GCNs)

Paper: Semi-supervised Classification with Graph Convolutional Networks (2017) [3]

GCN is a type of convolutional neural network that can work directly on graphs and take advantage of their structural information.

it solves the problem of classifying nodes (such as documents) in a graph (such as a citation network), where labels are only available for a small subset of nodes (semi-supervised learning).

graph convolutional network
Example of Semi-supervised learning on Graphs. Some nodes dont have labels (unknown nodes).

Main Ideas

As the name “Convolutional” suggests, the idea was from Images and then brought to Graphs. However, when Images have a fixed structure, Graphs are much more complex.

graph convolutional network
Convolution idea from images to graphs. (Picture from [1])

The general idea of GCN: For each node, we get the feature information from all its neighbors and of course, the feature of itself. Assume we use the average() function. We will do the same for all the nodes. Finally, we feed these average values into a neural network.

In the following figure, we have a simple example with a citation network. Each node represents a research paper, while edges are the citations. We have a pre-process step here. Instead of using the raw papers as features, we convert the papers into vectors (by using NLP embedding, e.g., tf–idf).

Let’s consider the green node. First off, we get all the feature values of its neighbors, including itself, then take the average. The result will be passed through a neural network to return a resulting vector.

graph convolutional network
The main idea of GCN. Consider the green node. First, we take the average of all its neighbors, including itself. After that, the average value is passed through a neural network. Note that, in GCN, we simply use a fully connected layer. In this example, we get 2-dimension vectors as the output (2 nodes at the fully connected layer).

In practice, we can use more sophisticated aggregate functions rather than the average function. We can also stack more layers on top of each other to get a deeper GCN. The output of a layer will be treated as the input for the next layer.

graph convolutional network
Example of 2-layer GCN: The output of the first layer is the input of the second layer. Again, note that the neural network in GCN is simply a fully connected layer (Picture from [2])

Let’s take a closer look at the maths to see how it really works.

Intuition and the Maths behind

First, we need some notations

Let’s consider a graph G as below.

graph convolutional network
From the graph G, we have an adjacency matrix A and a Degree matrix D. We also have feature matrix X.

How can we get all the feature values from neighbors for each node? The solution lies in the multiplication of A and X.

Take a look at the first row of the adjacency matrix, we see that node A has a connection to E. The first row of the resulting matrix is the feature vector of E, which A connects to (Figure below). Similarly, the second row of the resulting matrix is the sum of feature vectors of D and E. By doing this, we can get the sum of all neighbors’ vectors.

graph convolutional network
Calculate the first row of the “sum vector matrix” AX
  • There are still some things that need to improve here.
  1. We miss the feature of the node itself. For example, the first row of the result matrix should contain features of node A too.
  2. Instead of sum() function, we need to take the average, or even better, the weighted average of neighbors’ feature vectors. Why don’t we use the sum() function? The reason is that when using the sum() function, high-degree nodes are likely to have huge v vectors, while low-degree nodes tend to get small aggregate vectors, which may later cause exploding or vanishing gradients (e.g., when using sigmoid). Besides, Neural networks seem to be sensitive to the scale of input data. Thus, we need to normalize these vectors to get rid of the potential issues.

In Problem (1), we can fix by adding an Identity matrix I to A to get a new adjacency matrix Ã.

Pick lambda = 1 (the feature of the node itself is just important as its neighbors), we have Ã = A + I. Note that we can treat lambda as a trainable parameter, but for now, just assign the lambda to 1, and even in the paper, lambda is just simply assigned to 1.

By adding a self-loop to each node, we have the new adjacency matrix

Problem (2)For matrix scaling, we usually multiply the matrix by a diagonal matrix. In this case, we want to take the average of the sum feature, or mathematically, to scale the sum vector matrix ÃX according to the node degrees. The gut feeling tells us that our diagonal matrix used to scale here is something related to the Degree matrix D̃ (Why , not D? Because we’re considering Degree matrix  of new adjacency matrix Ã, not A anymore).

The problem now becomes how we want to scale/normalize the sum vectors? In other words:

How we pass the information from neighbors to a specific node?

We would start with our old friend average. In this case, D̃ inverse (i.e., D̃^{-1}) comes into play. Basically, each element in D̃ inverse is the reciprocal of its corresponding term of the diagonal matrix D.

For example, node A has a degree of 2, so we multiple the sum vectors of node A by 1/2, while node E has a degree of 5, we should multiple the sum vector of E by 1/5, and so on.

Thus, by taking the multiplication of D̃ inverse and X, we can take the average of all neighbors’ feature vectors (including itself).

So far so good. But you may ask How about the weighted average()?. Intuitively, it should be better if we treat high and low degree nodes differently.

We’re just scaling by rows, but ignoring their corresponding columns (dash boxes)
Add a new scaler for columns.

The new scaler gives us the “weighted” average. What are we doing here is to put more weights on the nodes that have low-degree and reduce the impact of high-degree nodes. The idea of this weighted average is that we assume low-degree nodes would have bigger impacts on their neighbors, whereas high-degree nodes generate lower impacts as they scatter their influence at too many neighbors.

graph convolutional network
When aggregating feature at node B, we assign the biggest weight for node B itself (degree of 3), and the lowest weight for node E (degree of 5)
Because we normalize twice, we change “-1” to “-1/2”

For example, we have a multi-classification problem with 10 classes, F will be set to 10. After having the 10-dimension vectors at layer 2, we pass these vectors through a softmax function for the prediction.

The Loss function is simply calculated by the cross-entropy error over all labeled examples, where Y_{l} is the set of node indices that have labels.

The number of layers

The meaning of #layers

The number of layers is the farthest distance that node features can travel. For example, with 1 layer GCN, each node can only get the information from its neighbors. The gathering information process takes place independentlyat the same time for all the nodes.

When stacking another layer on top of the first one, we repeat the gathering info process, but this time, the neighbors already have information about their own neighbors (from the previous step). It makes the number of layers as the maximum number of hops that each node can travel. So, depends on how far we think a node should get information from the networks, we can config a proper number for #layers. But again, in the graph, normally we don’t want to go too far. With 6–7 hops, we almost get the entire graph which makes the aggregation less meaningful.

graph convolutional network
Example: Gathering info process with 2 layers of target node i

How many layers should we stack the GCN?

In the paper, the authors also conducted some experiments with shallow and deep GCNs. From the figure below, we see that the best results are obtained with a 2- or 3-layer model. Besides, with a deep GCN (more than 7 layers), it tends to get bad performances (dashed blue line). One solution is to use the residual connections between hidden layers (purple line).

graph convolutional network
Performance over #layers. Picture from the paper [3]

Take home notes

  • GCNs are used for semi-supervised learning on the graph.
  • GCNs use both node features and the structure for the training
  • The main idea of the GCN is to take the weighted average of all neighbors’ node features (including itself): Lower-degree nodes get larger weights. Then, we pass the resulting feature vectors through a neural network for training.
  • We can stack more layers to make GCNs deeper. Consider residual connections for deep GCNs. Normally, we go for 2 or 3-layer GCN.
  • Maths Note: When seeing a diagonal matrix, think of matrix scaling.
  • A demo for GCN with StellarGraph library here [5]. The library also provides many other algorithms for GNNs.

Note from the authors of the paper: The framework is currently limited to undirected graphs (weighted or unweighted). However, it is possible to handle both directed edges and edge features by representing the original directed graph as an undirected bipartite graph with additional nodes that represent edges in the original graph.

What’s next?

With GCNs, it seems we can make use of both the node features and the structure of the graph. However, what if the edges have different types? Should we treat each relationship differently? How to aggregate neighbors in this case? What are the advanced approaches recently?

In the next post of the graph topic, we will look into some more sophisticated methods.

graph convolutional network
How to deal with different relationships on the edges (brother, friend,….)?

REFERENCES

[1] Excellent slides on Graph Representation Learning by Jure Leskovec (Stanford):  https://drive.google.com/file/d/1By3udbOt10moIcSEgUQ0TR9twQX9Aq0G/view?usp=sharing

[2] Video Graph Convolutional Networks (GCNs) made simple: https://www.youtube.com/watch?v=2KRAOZIULzw

[3] Paper Semi-supervised Classification with Graph Convolutional Networks (2017): https://arxiv.org/pdf/1609.02907.pdf

[4] GCN source code: https://github.com/tkipf/gcn

[5] Demo with StellarGraph library: https://stellargraph.readthedocs.io/en/stable/demos/node-classification/gcn-node-classification.html

This article was originally published on Medium and re-published to TOPBOTS with permission from the author.

Enjoy this article? Sign up for more computer vision updates.

We’ll let you know when we release more technical education.

Continue Reading

AI

Microsoft BOT Framework — Loops

Published

on

Loops is one of the basic programming structure in any programming language. In this article, I would demonstrate Loops within Microsoft BOT framework.

To follow this article clearly, please have a quick read on the basics of the Microsoft BOT framework. I wrote a couple of articles sometime back and the links are below:

Let’s Get Started.

I would be using the example of a TaxiBot described in one of my previous article. The BOT asks some general questions and books a Taxi for the user. In this article, I would be providing an option to the user to choose there preferred cars for the ride. The flow will look like below:

Create a new Dialog Class for Loops

We would need 2 Dialog classes to be able to achieve this task:

  1. SuperTaxiBotDialog.cs: This would be the main dialog class. The waterfall will contains all the steps as defined in the previous article.
  2. ChooseCarDialog.cs: A new dialog class will be created which would allow the user to pick preferred cars. The loop will be defined in this class.

The water fall steps for both the classes could be visualized as:

The complete code base is present on the Github page.

Important Technical Aspects

  • Link between the Dialogs: In the constructor initialization of SuperTaxiBotDialog, add a dialog for ChooseCarDialog by adding the line:
AddDialog(new ChooseCarDialog());

1. 8 Proven Ways to Use Chatbots for Marketing (with Real Examples)

2. How to Use Texthero to Prepare a Text-based Dataset for Your NLP Project

3. 5 Top Tips For Human-Centred Chatbot Design

4. Chatbot Conference Online

  • Call ChooseCarDialog from SuperTaxiBotDialog: SuperTaxiBotDialog calls ChooseCarDialog from the step SetPreferredCars, hence the return statement of the step should be like:
await stepContext.BeginDialogAsync(nameof(ChooseCarDialog), null, cancellationToken);
  • Return the flow back from ChooseCarDialog to SuperTaxiBotDialog: Once the user has selected 2 cars, the flow has to be sent back to SuperTaxiBotDialog from the step LoopCarAsync. This should be achieved by ending the ChooseCarDialog in the step LoopCarAsync.
return await stepContext.EndDialogAsync(carsSelected, cancellationToken);

The complete code base is present on the Github page.

Once the project is executed using BOT Framework Emulator, the output would look like:

Hopefully, this article will help the readers in implementing a loop with Microsoft BOT framework. For questions: Hit me.

Regards

Tarun

Source: https://chatbotslife.com/microsoft-bot-framework-loops-fe415f0e7ca1?source=rss—-a49517e4c30b—4

Continue Reading

AI

The Bleeding Edge of Voice

This fall, a little known event is starting to make waves. As COVID dominates the headlines, an event called “Voice Launch” is pulling…

Published

on

Tapaan Chauhan

This fall, a little known event is starting to make waves. As COVID dominates the headlines, an event called “Voice Launch” is pulling together an impressive roster of start-ups and voice tech companies intending to uncover the next big ideas and start-ups in voice.

While voice tech has been around for a while, as the accuracy of speech recognition improves, it moves into its prime. “As speech recognition moves from 85% to 95% accuracy, who will use a keyboard anymore?” says Voice Launch organizer Eric Sauve. “And that new, more natural way to interact with our devices will usher in a series of technological advances,” he added.

Voice technology is something that has been dreamt of and worked on for decades all over the world. Why? Well, the answer is very straightforward. Voice recognition allows consumers to multitask by merely speaking to their Google Home, Amazon Alexa, Siri, etc. Digital voice recording works by recording a voice sample of a person’s speech and quickly converting it into written texts using machine language and sophisticated algorithms. Voice input is just the more efficient form of computing, says Mary Meeker in her ‘Annual Internet Trends Report.’ As a matter of fact, according to ComScore, 50% of all searches will be done by voice by 2020, and 30% of searches will be done without even a screen, according to Gartner. As voice becomes a part of things we use every day like our cars, phones, etc. it will become the new “norm.”

The event includes a number of inspiration sessions meant to help start-ups and founders pick the best strategies. Companies presenting here include industry leaders like Google and Amazon and less known hyper-growth voice tech companies like Deepgram and Balto and VCs like OMERS Ventures and Techstars.

But the focus of the event is the voice tech start-ups themselves, and this year’s event has some interesting participants. Start-ups will pitch their ideas, and the audience will vote to select the winners. The event is a cross between a standard pitchfest and Britain’s Got Talent.

Source: https://chatbotslife.com/the-bleeding-edge-of-voice-67538bd859a9?source=rss—-a49517e4c30b—4

Continue Reading
AI2 hours ago

Graph Convolutional Networks (GCN)

AI4 hours ago

Microsoft BOT Framework — Loops

AI4 hours ago

The Bleeding Edge of Voice

AI19 hours ago

Using Amazon SageMaker inference pipelines with multi-model endpoints

AI19 hours ago

Using Amazon SageMaker inference pipelines with multi-model endpoints

AI19 hours ago

Using Amazon SageMaker inference pipelines with multi-model endpoints

AI19 hours ago

Using Amazon SageMaker inference pipelines with multi-model endpoints

AI19 hours ago

Using Amazon SageMaker inference pipelines with multi-model endpoints

AI19 hours ago

Using Amazon SageMaker inference pipelines with multi-model endpoints

AI19 hours ago

Using Amazon SageMaker inference pipelines with multi-model endpoints

AI19 hours ago

Using Amazon SageMaker inference pipelines with multi-model endpoints

AI19 hours ago

Using Amazon SageMaker inference pipelines with multi-model endpoints

AI19 hours ago

Using Amazon SageMaker inference pipelines with multi-model endpoints

AI19 hours ago

Using Amazon SageMaker inference pipelines with multi-model endpoints

AI19 hours ago

Using Amazon SageMaker inference pipelines with multi-model endpoints

AI19 hours ago

Using Amazon SageMaker inference pipelines with multi-model endpoints

AI19 hours ago

Using Amazon SageMaker inference pipelines with multi-model endpoints

AI19 hours ago

Using Amazon SageMaker inference pipelines with multi-model endpoints

AI19 hours ago

Using Amazon SageMaker inference pipelines with multi-model endpoints

AI19 hours ago

Using Amazon SageMaker inference pipelines with multi-model endpoints

Trending