#nlu

Yes, my first job was related to exploring self-organizing maps (so-called Kohonen nets). I don’t remember all project details, but we worked on syntactic dependency structures and tried to represent it in Kohonen maps for the German language. After we tried different IE approaches, text classification and run other experiments. That was a great time for learning. I had done an internship during my studies as well. It was in Paris, at a software company called Arisem, so I could also practice my French. It was the B2B company which focused on semantic search, dedicated one, including crawling. Then I came back to finish my master thesis.

What was master thesis topic about?

It was about the numerical representation of text corpora including how can we represent corpora for classification. I tried LSA that time also, but the topic was like a meta-analysis of different approaches.

Then Ph.D. happened to you.

Yes, at some point after I decided to stay in academia, to do a Ph.D. I went to Jena university, a big move from Cologne. But it was not only a Ph.D. position but a research assistant position in a European project BootStrep. The focus was on biomedical text processing: text mining in biology, semantic search over the publication of medicine/bio published research. There is a huge database PubMed which has millions of citations and which continue to grow quickly. And, obviously, a problem for a biologist is to find relevant information in such an enormous amount of data. So, preprocessing of data, named entity recognition (NER), normalization of extracted entities and relation extraction, are of particular interest here. My personal focus was on relation extraction, e.g. how a researcher describes gene expression processes.

Did you have medicine ontology for named entities?

We had a couple of Ph.D. students, which helped to develop the ontology in our group, of course using terminology from established sources. It reminds also what else was great about the project group: everybody had a specific skill-set and the tasks were assigned well and according to a person focus: somebody worked on NER and fast annotation using active learning, someone - preprocessing, another person cared about the ontology, search engines. I focused on detection of events and relations. It was a great experience to have such a skilled team.

Do you remember a day when you realized that you need to leave the project?

I continued working on the project during my Ph.D. I started later and the main result I would say was my participation in BioNLP 2009 shared task, where I got a second place once evaluated. After that, I elaborated on my topic. On 2012 I’ve completed my Ph.D. and started to look for a new challenge. I could have stayed in the Biomedical domain, but I was open to other topics also as I studied a lot while reading about different topics, including dependency parsing, collecting data in general. Then I found an open position at Nuance, there were not many at that moment in Germany. So, I became one of the first joining the NLU (Natural Language Understanding) team and moved to Aachen, which I also like as it’s close to Cologne.

How many people in Nuance NLU team now?

There are about 60 people in Automotive cloud NLU, which includes Aachen, Montreal, and Burlington offices and people working remotely. Company-wide there are more NLU people (100+).

NLU is a challenge by the name. So, tell us, what do you do and how do you overcome the challenges?

First, our main application area is an automotive domain. Our team works at the moment on a classification of user intents and named entity recognition. So, you have one-two step dialog, one-shot query, which requires a classification of the intent. I’d say that it’s now for the navigation system, office system in the car.

Well, actually from my experience I remember around a year ago participating in a hackathon organized by Nuance NLU system. And if I recall correctly, for NLU system you need to provide not only intents but also labels, concepts to train it, am I right?

Yes, you also need to provide concepts which need to be detected.

Would be nice if you can share an example of a use-case.

Ok, the simple example is a question about the weather: “What will be the weather tomorrow in Trento?” So, we need to recognize the intent: weather, the date: tomorrow, the location: Trento. Another example, you can: would it be sunny tomorrow in Trento? So, we do have multiple steps, relying on statistical models and many features, like named entities, and lexical information (keywords sun, weather, etc). Both are possible: you can do intent classification first and then named entities or the other way around.

As I remember from the mentioned hackathon, you have two interfaces: speech and text.

You are right, but it’s another project, it’s Nuance MIX you mean, our project. In our solution, we provide an ability to type, use speech interface and handwriting.

You haven’t told us a lot of internal details yet ;) Ok, what languages do you support?

We support over 20 languages for Automotive cloud NLU, additionally to major European languages we have Czech, Swedish, Turkish, for Asia - Japanese, Cantonese, Mandarin, and others.

It leads me to the question: do you reuse models available or develop all yourself?

We develop all internally. For example, we have developers graduated from the Charles University in Prague, who work on Czech support.

That’s an interesting story about computational linguistics in Czech, though I wouldn’t call it as widely spoken as others in Europe, Charles University has two or three groups which develop universal dependencies for the language, though some more representative languages have none.

Alright, what do you work on currently?

It’s mostly improving accuracy for automotive-related projects (for features like navigation, weather search, and more), which includes adding of data. Also embedding extensions, and for that case, the main challenge is the proper evaluation, which helps to avoid a degradation in quality. We worked on a hybrid solution: embedded NLU and cloud NLU. As we have some overlap, we need to split the responsibilities in a clear way. We need to work on confidence for prediction. We are facing AI as well, I mean complex request, e.g. a user could ask: find me a good restaurant and a parking slot around. So, a combination of intents brings an interesting challenge.

So, let’s come back quickly to language sources: do you plan to release the language resources to the language developers community.

I have no insight regarding this from the business.

It is a company which was bought by Microsoft, Maluuba, which developed an evaluation dataset, NewsQA. So, releasing an evaluation dataset can be a good step from Nuance. Thank you for the talk and I wish you good luck with a challenge of multiple intents.

Thanks, I was happy to share the knowledge and what we do.

Image 1 is published with an agreement of K. Kruchinina