Hi Katja, let me ask about your professional experience: how did you find your interest in computational linguistics and developed through it?
Hi, I started my university studies in Rostov-on-Don, Russia. My subject was German language and literature, with emphasis on literature and translation studies. I studied for 2 years and then continued my studies at the University of Cologne in Germany. I started there from the first year, as in Germany similar program starts two years later than in Russia. Additionally, I took another topic of interest, French, as I was keen to learn it. But after a couple of years, I realized I had high interest in linguistics, especially after taking courses in modern linguistics and formal syntax in university. But honestly, I was not aware of computational linguistics at that point. One day I found that there is a study subject ‘Linguistic data processing’ at the University of Cologne and I joined the class after a talk with a professor. After a couple of years I started to work at the department, and of course, it was a good time to learn programming, which I really enjoyed. At that point, I realized much more about computer science. We studied Java as a first language, though many in the field start now with Python. I remember we programmed a search engine over a summer.
It reminds me a talk to Natalia Karlova-Burbonus. Natalia has a very similar story: going from interest to the German language to Computational Linguistics in Germany.
My next question whether you remember your first project or last at that time.
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
Аляксей Северын. "Нейронными сетями не занимается только ленивый", ч. 2/2
Гэта другая частка гутаркі з Аляксеем Северыным. Першую можна знайсці тут.
Што вы выкарыстоўваецце зараз для задачы па summarization?
Мы используем machine learning (ML), нейронные сети тоже. Мне кажется, что об этом на wired есть статья. Я не могу говорить детали, помимо того, что уже публично известно.
Давай падумаем пра будучыню. Што чакае human-computer interface? Якія змяненні будуць?
Все больше и больше end-to-end систем, neural-based. Они будут вообще полностью заменять огромные массивные компоненты, например, machine translation. Изначально это была очень сложная система с большим compressing компонентом, с pipelines. И сейчас Google анонсировал, что уже несколько языковых пар, например, Chinese-English – это просто end-to-end система machine translation. Тоже самое будет происходить в других областях, speech recognition и question answering. На сегодняшний день в академии это есть, но момент, когда это все окончательно сформируется в production, пока еще неясeн.
Год, два, тры?
Я думаю, что в ближайший год, но это будет инкрементно. Мое мнение, что через два года большие куски сложных систем, где используется machine learning, станут просто одной большой нейронной сетью. Все неизбежно движется в этом направлении.
Адна з мэтаў суполкі – развіваць цікаўнасць да вобласці. Таму, што можаш пажадаць маладым даследчыкам?
Смотря какие цели. Они могут быть разные: построить карьеру, найти хорошую работу, продвинуться в своем исследовании...
Напрыклад, давай уявім сабе маладога даследчыка, які толькі пачаў развівацца ў накірунку задачы question answering. У яго сomputer science background пасля беларускага ўніверсітэта. Што яму рабіць?
Путь не такой быстрый. Человеку, который находится на одном уровне, сложно посмотреть назад и по полочкам разложить. Мне кажется, что я даже ни одной книги не прочитал полностью. Когда занимаешься исследованием, у тебя просто нет такой возможности, как сесть и прочитать полностью книгу.
Я читал только научные статьи. Мой совет: если прочитал статью и понял, что очень нравится идея, то сразу же нужно сесть и воспроизвести ее в коде. Сейчас задача упрощается, потому что стало очень популярно релизить код. Когда я учился, около 5 лет назад, это было не настолько распространено - код публиковался в порядке исключения. На сегодняшний день релиз кода - это признак хорошего тона.
Кстати, у меня есть очень хороший совет для начинающих - участвовать в соревнованиях на Kaggle. В свое время я достаточно сильно прокачался, даже в какой-то момент на 26 место "заполз", что достаточно круто. Меня это многому научило. Т.е. надо уметь работать руками, потому что самые лучшие идеи приходят в процессе практики.
У исследователя очень хорошо должна быть развита интуиция, тогда действительно могут прийти какие-то дельные идеи. Для того, чтобы понять недостатки или ограничения какой-то модели, в первую очередь, нужно с ней поработать практически. На мой взгляд, очень много внимания уделено созданию различных архитектур и моделей, но что касается nlp, то здесь очень большое количество задач сводится к sequence prediction, например, как в machine translation. То есть способы, которые позволяют генерировать structured output, что выражается, например, в использовании правильной loss-function. Сегодня большинство использует cross-entropy loss, т.е. оптимизируют, там где сеть выдает по одному слову за раз. Но есть достаточно большое количество статей, где используется reinforcement learning, CRF Objective function поверх всего. Это все помогает, все полезно, потому очень важно понимать, как всё работает.
Frameworks?
Я работаю с TensorFlow сейчас, когда учился - Theano. На сегодняшний день, я считаю, что Theano пока самый удачный framework, причем простой в использовании. И сообщество породило много надстроек, напр. Lasagne, Keras. Когда я начинал в своё время - я писал все с нуля.
Сейчас нейронными сетями не занимается только ленивый.
Аляксей Северын. "Нейронными сетями не занимается только ленивый", ч. 1/2
Сёння мы размаўляем з Аляксеем Северыным, хлопцам з Беларусі, які займаецца камп’ютарнай лінгвістыкай.
Прывітанне, Леша. Калі ласка, раскажы пра сябе, пра асноўныя этапы развіцця ў галіне: з чаго пачыналася, як з’явілася цікаўнасць увогуле?
В общем, мой путь был, наверное, не похож на путь многих. Меня всегда интересовала физика. Я учился в лицее БГУ на физике и потом поступил на радиофизику, потом я пошел работать в компанию в Минске, она называлась ScienceSoft. У ScienceSoft была дочерняя компания, которая называлась Nilitis. Она занималась разработкой систем автоматической торговли.
Магчыма працаваў з Цурікавым?
Да, он был моим руководителем, и меня хорошо знает. У меня там быстро сложилось карьера, и мне все нравилось, но решил, что мне нужно прокачаться в machine learning (ML). На тот момент в Беларуси программ таких было немного, может совсем чуть-чуть в БГУ на ФПМИ, определенные аспекты там затрагивали. Сложилась так, что, еще до того как я пошел работать в Nilitis, я подал заявки в несколько университетов на различные программы, и одна из них была связана с ML. Это то, что мне было интересно.
Я подумал, почему бы мне не прокачаться в ML, год или два поучиться и вернуться уже с совсем другим багажом знаний и умений. Потому что как раз то, над чем я начал работу в Nilitis, было Support Vector Machine (SVM). Это было еще все очень новое и интересное.
Поэтому я поехал по программе Erasmus Mundus. Первый год был в Италии, в Тренто, а второй год в Бонне, Германия. Программа была computer science, data science, но была достаточно большая возможность выбирать курсы по ML. Мне это понравилось, я закончил первый год в Тренто, и не закончив master degree, сразу переключился на PhD. Мой advisor, Alessandro Moschitti. Он меня заприметил на своем курсе и предложил после магистратуры присоединиться к его группе на время PhD. В тот момент я встретил свою будущую жену, как раз в Тренто, и я не захотел уезжать в Бонн, таким образом у меня возникла возможность сразу переключиться на PhD, чего я никогда не планировал. У меня не было интереса к академической карьере, но, тем не менее, я начал делать PhD по ML и мне это стало нравиться.
Якая была тэма?
Я занимался разработкой моделей, в которых необходимость feature engineering сводилась к минимуму. Мы использовали kernel functions, потому что это было основное направление исследования моего профессора. Он в этом эксперт. Я работал над проектом IBM Watson.
Што ты маеш на ўвазе “працаваў”?
У профессора было сотрудничество с исследователями из IBM Watson. Мы пытались использовать технологии Tree Kernel Functions, SVM для проекта, чтобы добавить один из сигналов в их систему ранжирования. Поскольку, в принципе, специфика тех вопросов, с которыми они работают, достаточно сильно зависит от синтаксиса, то этот сигнал оказался очень полезным. И он освобождал от необходимости вручную извлекать эти features, придумывать, какие из них должны работать, потому что, фактически, feature space бесконечен. Это то, чем я занимался в первые несколько лет, и потом уже переключился на нейронные сети, deep learning.
Это был последний год, но на тот момент я уже созрел как исследователь, то есть мне уже руководство не было необходимо, мог заниматься этим направлением самостоятельно.
Атрымоўваецца, што ты пачаў прымяняць DL да тых задач, якія былі дагэтуль?
Две основные задачи, которые меня интересовали на тот момент – это sentiment analysis и question answering (QA). И, фактически, я построил две системы, которые в то время стали state-of-the-art. Я победил в конкурсе по sentiment analysis на Semeval 2015. Моя модель была на первом месте и по question answering у нас получились топовые результаты по стандартному benchmark.
Гэта быў community question answering ці open domain?
Это была задача factoid question answering.
Пасля чатырох год у Італіі, у 2015 годзе ты пачаў працаваць у кампаніі Google, на якой пазіцыі? Якія задачы былі пад тваей адказнасцю?
На позиции research scientist. Я работаю над системой следующего типа. Например, если ты в гугл задаешь вопрос “Ok, Google, why is the sky blue?”. Ты получаешь ответ. Я работаю над системой question answering и моя задача – делать summarization ответов, которые воспроизводятся голосом. Потому что, когда ты видишь ответ на экране, ты его можешь прочитать быстро. Но, когда ты вынужден его слушать, чем длиннее ответ, тем сложнее его слушать.
Атрымоўваецца, што тут есць дзьве часткі задачы, за якія ты адказны – вылучэнне пытання і summary.
Нет, я отвечаю за часть по summary. Система большая и сложная, там куча всяких разных сигналов.
Якім чынам адрозніваецца text summarization ад voice-based summarization?
Они ничем не отличаются. Ты саммаризируешь текст, который потом подается на вход text-to-speech системе.
Ці есць афіцыйныя benchmarks ці нейкія baselines, community tasks на Semeval?
По этой задаче – нет. По question answering summarization, насколько я знаю, нет в академии. Есть много benchmarks по summarization в целом и есть много различных типов summarization: document level summarization, sentence level summarization. Но в question answering summarization сложность состоит в том, что тебе нужно сделать summarization таким образом, чтобы ты сохранял ответ, то есть не потерял суть. Например, ты задаешь вопрос и хочешь сжать ответ, но таким образом, чтобы смысл ответа остался.
У другой часцы размовы мы закранем бачанне бліжэйшай будучыні і абсудзім некаторыя парады для маладых даследчыкаў.
Today we talk to Barbara Plank, well-known computational linguistics researcher. Barbara, let me ask you to introduce yourself: some words about your experience and background, please.
I came from Computer Science, I started in Bolzano, and then went to Amsterdam for the LCT master program; Amsterdam was part of LCT back then.
During my bachelor I took the optional 'Introduction to Computational Linguistics' class with Raffaella Bernardi. The course showed me that "Hey! There is something else than databases, software engineering and the other related stuff". It was very exciting for me. So, that's why I became interested in CL/NLP and took the LCT master.
How was your interest changed during the career?
In Bolzano it was mainly CS, but in Amsterdam it was Statistical NLP and Machine Learning. And it was funny, because I remember at some point I was asked whether I want to go towards ML and I was thinking: "Not sure I want to go that direction..." And it turns out to be what I'm doing all the time nowadays.
Do you remember your first paper?
My first paper was in my undergrad studies. That was about a summer internship I've been doing in Bolzano at our faculty, on extending the current catalog of the library, which was monolingual, to several languages, i.e., cross-lingual information retrieval. Two years later I published my master thesis on 'Sub-domain driven parsing'.
The idea there was to assume that a treebank is a mix of domains, heterogeneous, and one wants to weight some domains more. This is an interest of mine that persists till today, I continued to work along this line during my PhD in Groningen, which I finished in 2011. After that I moved to Trento for a year. That time I started to work on relation extraction, and I could explore a new area, working on a new task and method (tree kernels). After that I went to Copenhagen, to work in a great group.
A view of Bolzano
You've attending ACL-2016 in Berlin. Could you please share your impression? Probably, focusing on what was important for you in terms of your interest and in general NLP.
ACL this year was huge (1600 participants), the whole field is growing. Every year our conferences get more papers, more people. It will be an issue to host such events in future, but it is also a very positive sign, with growing interest in our field. Also businesses and companies are getting an idea: "Oh, wow! There is this thing called NLP, we want NLP!", though, it is not just one thing, but interest is growing, both from academia and industry.
Could you, please, share the general idea of the class you were teaching at ESSLLI-2016 in Bolzano?
Yes, the name of the course is 'Fortuitous data to improve language technology'. Fortuitous data means that instead of training models on a certain task, certain language, certain domain, we should really try to go beyond the domain and language, and look at signals, where they are not always obvious. A colleague of mine said, that, in other words, you have data and you can do end-to-end learning, that works amazingly well, but you can't do that for everything you are interested in: typically, you don't have supervised data. For some tasks you do have a lot of supervised data and you get some 'learning power' from it. So, the motivation was, like I show in the picture.
So, we have limited labeled data which is really small and we want to extend it, so, hopefully, going to better models and also combining learning signals from various sources. You can read more about fortuitous data in my recent position paper.
What was the paper that inspired you most in your research life?
Yoav Goldberg's tutorial about Neural Networks was very interesting for me recently. I can also refer to Joan Bresnan talk at ACL this year for the reward of 'Life-time achievement'. She had a nice analogy of going from a garden, 'armchair' linguistics, out into the bush. And the bush for her was a way like these nice grammatical structures, so in a real world. But that also linked nicely to her life, because she worked first mostly on English and later on many African languages: so, really, she moved into the bush, doing field-studies. I liked the question by Joakim Nivre: "Is our young generation actually living in the bush?" So, rephrasing it, if we now teach young people only deep-nets stuff, then they miss out a lot of basic things in Computational Linguistics. I think this is important to keep in mind, to balance new and old, the Jurafsky and Manning's book is a great reference here.
This goes back to my general philosophy: use as much as you know, and reuse instead of starting from scratch. Luckily also in research, fields start to talk to each other, interact, a lot happens these days with NLP and Computer Vision, and I hope the same will happen with Speech Processing and Cognitive Science.
Interview with Jon Dehdari, computational linguist, part 2/2
Here you find the second part of our talk to Jon Dehdari. First part can be accessed via the link.
Another interesting thing about the N400: it doesn't depend on the word order. If there is a word that might be semantically congruent but might not fit in with the syntactic structure, there is no N400 activity going on. There are other activities going on in the brain.
One called the P600 - a positivity in the brain, around the top back part of the brain, around 0.6s after we hear a word that is syntactically not well-formed.
There are other language related ERP as well. For example, if a part of speech or intonation are not what we are expecting. But N400 primarily is concerned with semantic coherence. I thought it would be useful as a language model to model that. Especially when we have other language models that can help account for local grammatical coherency or long distance grammatical coherency with syntactic language model, but there hadn’t been as much work done on longer distance semantic coherency. That was something that I worked on.
Is the developed LM available online?
It’s currently not available, but I want to put it online. I think this is something I should probably do :)
This achievement should be available for other researchers. I know that you teach some courses at Saarland University. What kind of courses and what is the content?
I teach master’s level courses every semester at Saarland University in the CoLi (Computational Linguistics) department and I teach classes in the translation department as well. There are going to be unifying in the next few months. For the CoLi class it’s usually related to language modelling or machine translation. For the translation department, it’s either related to machine translation or Python and shell programming.
What is your favourite toolset in terms of programming and NLP? What is your daily toolset, computational linguistics related tools?
Normally, I try to use the right tool for the job :) It depends on what you are going to be doing. If you are new to programming, Python is a safe choice, it’s good to learn it because it’s quite common. If you are going to be programming something that is computationally very expensive, then C is a good language for mature programmers. Swift and Go also look nice.
I wanted to know about special tools. Moses, you’ve mentioned already, right? What are others would you recommend to have a look?
Moses has been around for a while. It’s very mature and very well developed and incorporates a lot of different ways to do machine translation, both phrase based and hierarchical or syntax oriented. The field in general is undergoing a lot of changes right now, so there is a lot of work being done on neural machine translation, which tends to be very minimalistic in terms of software. So the trade-off between simpler algorithms and simpler software but being more computationally expensive. And I think that researchers today would be well off to learn some of the tensor-processing libraries, like TensorFlow or Torch, Theano, or other software that works with neural networks. If you are a Python programmer, Keras makes building neural networks very simple, and it has either Theano or TensorFlow as a backend, which allows you to build a distributed environment. DL4J is another example of building distributed environment for NNs. Most of the others run on a single box.
State-of-art in neural networks, in your opinion? What the state of the field?
This is rapidly changing field. Every month there is something new created. Currently, people are working mainly with bidirectional LSTMs and GRUs. Also, recently works on character-based neural machine translation appeared. This is actually an area that I and a colleague of mine Georg Heigold were pursuing a few month ago for a collaborative workshop. And while we were in the process of developing it, we discovered that some other researchers from Carnegie Mellon as well as UPC, Barcelona were thinking along the same lines. So, I think it’s a good idea.
Even though it sounds like translating characters to characters, it’s not the case, but rather the idea is to look below the level of words, so to use sub-word units and try to translate these sub-word units, which makes sense for anyone who is familiar with a morphologically rich language, like Turkish, which has a lot long words. The problem with translating morphemes is that it’s very time-consuming to develop a morphological analyser. I’ve done that before and it’s a very difficult process. Sometimes, what a human thinks is a good morphological unit might be not good to translate. Sometimes, it’s good to translate fairly small morphological units, sometimes it makes more sense to translate larger units or an entire word/phrase. So, it requires a flexibility: what is a good level to translate a linguistic unit, and neural nets are quite flexible in the way what they can ‘learn’.
So, some people from Harvard and New York University have worked on combining convolutional neural nets and recurrent neural nets, and have done a good job at finding good sub-word units and working on those levels to translate. It’s based on a sequence-to-sequence type of language model. A lot more people are working on it now. Just a few years ago, Neural MT was not as competitive as phrase-based MT. I think it’s getting better and better with model, algorithmic, software, and hardware improvements. Achievements for some language pairs are very promising.
We were talking to Dzmitry Bahdanau about other libraries for Neural MT called Blocks and Fuel. So I guess we can add those to the list of tools.
Do you think in the following few months we can expect more research papers on CharNN MT?
There is lot of interest and excitement in breaking down words into smaller units. But, it should be mentioned, that there is also work on higher-levels of linguistic units, like phrases, sentences, and documents.
Interview with Jon Dehdari, computational linguist, part 1/2
Today we talk to Jon Dehdari, a well-known computational linguist from Saarland University. Jon, let me ask you to introduce yourself: what did you study before, how did you reach the point you are at now?
Hi, sure, I’m Jon Dehdari. I’m doing a post-doc here at University of Saarland and DFKI, Germany. And before that I was working on a PhD at Ohio State University in the US. I was working on different kinds of NLP related topics. I started out working on parsing and formal analysis of syntax as well. And then I drifted into statistical NLP, machine translation, and then to neuroscience-informed NLP, I guess.
I worked with professor William Schuler on an EEG-inspired language model. EEG is a way of studying how the brain is working at a large, macroscopic level.
The timing is very fast, so you are able to see in real time how the brain is processing language and other things that are happening.
Well, it’s wide range of topics – parsing, syntax, brain processing from an EEG point of view. But do you have a main topic of interest now?
Machine translation and language modeling are main two areas I’m currently interested in. Language modeling (LM) is a general field within NLP and related areas. It’s basically just modeling language, and typically people think of LM as incremental language modeling. So, within the context of a sentence, given previous few words, it tries to predict what word will come next. That is not necessarily the only usage of language modelling, but very common scenario.
So, for both machine translation and speech recognition you have input coming in incrementally and in real time, and you want to know what word somebody spoke in to microphone or how a given word/phrase should be translated. So typical decoders for machine translation or speech recognition work incrementally - that’s the most efficient way of doing with those inputs.
What languages do you speak? It’s very important for computational linguist to have different background of some languages.
That’s true. So, English is my first language, I am also fluent in Spanish, and I have studied extensively Persian and Arabic, also German. I’ve studied to lesser extend five or ten other languages. Enough to get into trouble, but not enough to get out of trouble, I would say :)
I do have a linguistics background and some computer science background as well. I have an appreciation and fascination with human languages. I am always trying to learn new languages. Obviously, to acquire complete fluency in any language you need years and years of practice.
Let’s talk about your PhD work. What was the main problem, the proposed solution, main challenges, in your opinion?
I was initially interested in an unsupervised parser. Parsing is the task of annotating sentence with a syntactic structure, often in the form of a constituency tree. I was trying to find phrases and phrases of those phrases recursively, or trying to figure out what dependency relationships exists between words in a sentence. Usually that is done by learning from a treebank, an existing dataset of annotated syntactic trees. But I was interested in whether it is possible to learn those structures and relationships without the use of labeled data. There is plenty of labeled data for English and a few other languages, but for vast majority of the world’s languages there are simply no labeled data or very little of it. I am quite interested in part from the theoretical perspective with NLP. But also, as a linguist, I am interested in developing algorithms for all of the world’s languages or for most of them. So, unsupervised learning is an obvious choice of machine learning paradigms working with all the world languages.
As I worked more and more with unsupervised parsing and unsupervised grammar induction, it came apparent that it wasn’t that useful for machine translation decoder to have that information or it was only marginally useful to have that. There are a lot of different ways the machine translation goes wrong, and any given approach will have a small impact and I became less interested in unsupervised parsing and more interested in language modeling, which is a related area but a little bit different emphasis. N-gram language modeling is and has been a common technique for modeling language, where you just simply take the previous few words and base the probability of a next word that is going to follow on the previous few words. But it requires the previous few words to exist as it is in your training set (the data you used to train your language model). That does not work very well for free word order, which is quite common for morphologically-rich languages. An n-gram based language model works well for English and for other languages with fixed word order. That’s why I was interested in working with more morphologically-rich languages.
At that time neural language models were very new and were extremely slow. As more and more modelling data became available it was apparent for me to develop a language model first for these languages that can accommodate large amounts of data. I was interested in working within the constraints of a language model that was both fast and didn’t require any manual annotations and could make use of a longer history than just the previous three or four words. I developed a language model that unified a lot of existing techniques and also extended it so that it was both fast and accommodated freer word order as well. It was also inspired by the way that the brain processes semantic information at an incremental level. There is an event-related potential (ERP), which is an activity in the brain, that is called N400.
What we see is that at the particular part of the brain, the left side usually, in the middle, whenever we hear a part of the sentence and then we hear another word that is semantically incongruent with the words preceding it, then we're surprised a little bit to hear that word. What happens is, there is a change in voltage that becomes more negative than is typical around 0.4s after we hear that word coming in.
The classic example is - "A sparrow is not a ... boat" or "a sparrow is not a ... car". When I say “a sparrow is not a car”, that is logically true, but typically when we hear the word “sparrow”, we are going to think about some other animal-type word to follow or something that would be semantically associated with it. And when we hear the word “car”, it’s surprising to a certain extent and there is N400 would spike in that context. People would look at the N400 from a variety different angles and try to pen down when it happens and how it happens, in what context does it occur and it what context it doesn’t occur. What people found was that the more words that we hear in a given sentence, the less surprised we are going to be. Especially if all the preceding history is congruent, semantically. But at the beginning of the sentence we don’t know what the first word of that sentence will be. So the history can help to eliminate some of the surprisal that we can hear in the input, I developed a language model that reflects those patterns in the N400.
As more and more history comes in then the less surprisal it’s going to be, especially as words are congruent with the preceding words
The second part will be published soon and will tell you more things about N400 and neural networks applied to machine translation
Last Saturday, Feb 27, we had an incredible meetup, the third one of our group.
First of all, thank you all for joining us and being very active!
For those who was able to make it, in the post you can refresh some moments, ideas, thoughts; for others - you can find resourceful presentations and some pictures ;)
Alexey Cherkes was the first speaker. He introduced new text-to-speech synthesis for Belarusian and Russian. It’s developed as a part of the project http://mbook.by/, a multimedia library for pupils. For the moment, the app is already available for Android. TTS is a nice feature of the app: finally, a pupil can listen school-books, not just read.
Alexey did a short overview of approaches to TTS, and told about modern tools and methods. Now we know that TTS feature is based on HTS; audience also enjoyed some online-demos of different synthesis systems. We were talking about the issue of a lack of linguistic resources for Belarusian and how to join our efforts to develop it. Also, we were told that prosody implementation is the next step to improve the quality of tts.
Demonstration of Belarusian tts for Android at #NLProc meetup pic.twitter.com/I4L4y4uf2Y
— nlproc.by (@nlprocby)
February 27, 2016
You can find the presentation here, but we can say that the value of it was actual live-demos.
The second speaker was Denis Postanogov, also very experienced NLProc-engineer. Denis with a team have been developing a full-featured commercial product for big business-players. But nobody was bored as it happens when people talk about business solutions, and actually people asked a lot of questions, Denis replied to all of them. We even had to extend the time of the meetup for an hour to give a chance to explain everything people asked about.
Sentiment analysis which is based on predictive QA in IHS Goldfire, discussing at #nlproc meetup in Minsk, @nlprocby pic.twitter.com/C0qdX0eG1t
— Yauhen (@YauhenMinsk)
27 February 2016
Here are just a couple of topics discussed: semantic relations, QA, machine translation, sentiment analysis, pdf parsing.
It’s hard to explain in this short post everything Denis was talking about, thus we recommend to look at the presentation and imagine the complexity of the product (attention: nice pics).
Traditionally, we were happy to say “Thank you!" to speakers by giving our nice T-shirts.
You can find more pictures here. Join us next time.
Blocks гэта бібліятэка кампанентаў, якія выкарыстоўваюцца як часткі сістэмы DL: есць FF, Convolutional, Recurrent, Attention, таксама кампаненты для регулярызацыі. Есць нават канструктар, які дапамагае збіраць такія сістэмы для MT і SR (speech recognition), але ён атрымаўся вельмі складаным, таму я зараз працую над наступнай версіяй.
У нас атрымалася вельмі добрая імплементацыя, як арганізаваць цыкл навучання нейроннай сеткі. Можна дадаваць кампаненты, рабіць валідацыю на асобным dataset, есць магчымасць рэгулярна захоўваць параметры, улічваючы, што працэс застаецца эффектыўным.
Fuel - бібліятэка, каб дадзеныя хутка ператвараць у паток batches. Усе нашыя алгарытмы навучання заснаваны на паслядоўнасцях batches, таму нам неабходна мець software, якое дазволіць хутка, у тэрмінах коду, і эфектыўна пабудаваць неабходную паслядоўнасць. Гэта таксама агульны фармат для захоўвання вялікага аб’ёму datasets. Зараз у нас агульны фармат на базе hdf5. У нас атрымоўваецца захоўваць дадзеныя для машыннага зроку i машыннана перкладу, распазнавання маўлення ў адзін фармат. Гэта вельмі зручна.
Вялікай міжнароднай папулярнасці праекты не набылі. Здаецца, што адна з прычынаў, што нам трэба мець вельмі гнуткую прыладу працы, якая дазваляе змяніць абсалютна ўсе, таму што мы працуем на самым нізкім узроўні нейронных сетак, а большасці людзей патрэбны ўжо гатовыя сеткі.
Што чакаць у бліжэйшы час ад навуковага асяроддзя па тэме нейронных сетак? Якія, на тваю думку, будуць breakthroughs?
Галоўная крыніца breakthroughs на дадзены момант - індустрыяльныя лабараторыі, якія за апошні час нанялі вельмі шмат PostDocs, прафесараў, прафесіяналаў ML з багатым вопытам. Зараз у іх фокусе reinforcement learning (RL). Гэта якраз і выглядае наступным крокам. Усе гэта з нагоды, што хочацца менш залежыць ад наяўнасці шматлікіх прыкладаў, якія зараз спецыяльна рыхтуюцца жывымі людзьмі. Па-праўдзе, мае цікаўнасці патроху змяшчаюцца ў гэтым накірунку таксама. Таксама unsupervised learning (UL) для неразмечаных дадзеных, ідэя не новая, але ў апошні час гэты накірунак атрымлівае шмат увагі. Праблема ў тым, што не існуе метадаў UL, якія могуць вылучаць універсальныя, карысныя ўяленні дадзеных, якія можна хутка і лёгка выкарыстоўваць ва ўсіх сістэмах. Так што RL i UL, на маю думку, два самыя папулярныя buzzwords.
Я таксама зараз спрабую прымяніць некаторыя падыходы з RL у кантэксце задач, над якімі мы працуем. На маю думку, акрамя папулярных у гэты момант рашэнняў для гульняў (напр., Go ад Google DeepMind), есць яшчэ шмат невырашаных задач для supervised ML (speech recognition, machine translation), і на іх можна пратэстатваць шмат цікавых ідэй.
Магу адзначыць, што зараз набірае папулярнасць цікаўнасць да тэмы вылучэння алгартымаў пры дапамозе DNN (карацей, алгарытмы па прыкладах, ці Neural Turing Machine(NTM)). Пакуль я маю скептычныя адносіны да гэтай новай ідэі, плыні, таму, што, на маю думку, людзі так не робяць. Нядаўна чуў на дакладзе на NIPS, што ў Google навучылі сетку знаходзіць найкарацейшыя шляхі (на ўваходзе апісанне графа, на выхадзе - апісанне найкарацейшага шляху), але, зноў, мы так не робім - мы прыдумваем алгарытм: мы не маем шмат прыкладаў, каб распрацаваць дакладны метад. Магчыма, я пакуль не зразумеў важнасць падыходу, і пакуль мне цікавы больш задачы, для якіх людзі не могуць распрацаваць алгарытмы. Напрыклад, мне вельмі цікава, ці можна ўжываць end-to-end падыход у задачах хіміі і біялогіі, как будаваць карысныя малекулы і злучэнні.
Можа быць, NTM плануецца ў будучыні выкарыстоўваць для робатаў у задачах планавання, але зноў, мне здаецца, што людзі не выкарыстоўваюць алгарытмы ў планаванні, а больш простыя эўрыстычныя метады.
Што параіш з літаратуры для тых, у каго цікаўнасць да NN з’явілася нядаўна?
Добрае пытанне, на якое не маю адказу :) Я чытаў вельмі шмат навуковых артыкулаў, зараз есць кніга ад Yoshua, хаця не паспеў пакуль яе пачытаць. Ведаю, што яна вельмі добрая ў якасці зборніку спасылак на асноўныя навуковыя артыкулы. Таксама, ад Microsoft есць нейкая кніга, я пакуль не паспеў яе глянуць. Думаю, што хутка з’явіцца больш карысных матэрыялаў, а пакуль раю чытаць навуковыя паперы і недапісаную кнігу Y. Bengio, праходзіць анлайн-курсы.
Інтэрв’ю з Дзмітрыем Багданавым, даследчыкам у галіне deep learning. Частка 2/3
Працягваем размову з Дзімай, ніжэй другая частка гутаркі.
Дзіма, у працяг тваёй навуковай паперы, раскажы пра attention mechanism.
Раскажу пра attention mechanism менавіта ў кантэксце Deep Learning, бо ў іншых галінах навукі гэты тэрмін можа мець іншае значэнне. Дарэчы, не магу сказаць, што я быў адзіным, хто гэты метад вынайшаў, праз 3 месяцы вельмі падобныя рэчы былі прэзентаваны іншымі людзьмі, і яшчэ за 2 месяцы дагэтуль падобны падыход быў прапанаваны для апрацоўкі малюнкаў (image processing). Паспрабую распавесці, як я прыйшоў да АМ.
Традыцыйная feed-forward neural network мае дастаткова ўсталяваную архітэктуру, якую складана змяніць, такім чынам уваход і выхад маюць фіксаваныя памеры. Іншыя тыпы, калі памер уваходных дадзеных нефіксаваны, выходных - фіксаваны, напрыклад для задач sentiment classification, для задачы time-series prediction маем паслядоўнасці на ўваходзе і выхадзе аднолькавага памеру, і апошні тып - уваход і выхад нефіксаванага памеру. Вось менавіта апошні тып для мяне і асабліва цікавы, таму што вельмі складаны: як некаторыя часткі выхаду залежаць ад некаторых уваходных.
Насамрэч, як людзі робяць пераклад? Гледзячы на сказ, я напішу адное-два словы, паглядзеўшы на адно месца ў сказе, потым яшчэ некалькі словаў, глянуўшы ў іншае. Таму хочацца мець інтэрактыўнасць, умець выбіраць важную інфармацыю. Улічваючы такія абставіны, мы дадаем RNN, якая на кожным кроку, перад тым як выбраць наступнае слова, робіць пошук неабходнай інфармацыі. Вельмі складана выбраць тое, што неабходна, таму NN робіць soft выбар, дадаючы кожнай частцы ўваходу сваю вагу, далей мы робім нармалізацыю, потым атрымоўваем ўзважаную суму ўваходных дадзеных (weighted sum of inputs). Вось так мы атрымоўваем інфармацыю, якая карысна на наступным кроку. У выніку мы маем не тое, што раней, калі мы канцэнтравалі інфармацыю ў адно прадстаўленне, зараз па-іншаму: канфігурацыя свабодна мяняецца ад аднога сказу да іншага, без bottlenecks. Неабходна заўважыць, што гэтая дадатковая сетка прымяняецца шмат разоў, каб падлічыць усе камбінацыі ўваходых і выхадных частак. Калі гэта зрабіць вельмі акуратна, то не будзе асаблівага ўплыву на час трэніроўкі. Тым больш, можна сказаць, што мы маем shallow network (мы падлічваем першы слой для абодвух сказаў перад тым як прымяніць АМ), застаецца скласці два вектары, прымяніць activation function, і падлічыць dot product, які па скрытым слоі дае адну лічбу; у выніку мы маем О(mn) complexity з невялікай канстантай, дзе n i m - даўжыні кожнага са сказаў.
У выніку гэта практычна. Здаецца, Google Brain актыўна выкарыстоўвае АМ зараз: шмат артыкулаў ад іх з’явілася ў апошні час, а яны вельмі любяць практычная падыходы.
Якія праблемы бачыш у АМ? Што можна зрабіць для паляпшэння якасці machine translation?
Не магу сказаць, што я займаўся machine translation, нягледзячы на тое, што праца мая: MT для мяне адна з задач агульнага падыходу. Але ў мяне былі канкрэтныя меркаванні і ідэі. Таксама трэба адзначыць, што я не паспяваю сачыць за ўсемі публікацыямі. Таму да гэтага часу я ня ведаю, ці зрабіў хто-небудзь one-to-many attention. Напрыклад, у нямецкай мове ёсць прыстаўкі, якія не заўседы з’яўляюцца часткай слова да якого адносяцца: “Ich mache am Morgen um 10 Uhr das Radio an.”, an - гэта адлучаная прыстаўка дзеяслова anmachen. Неабходна звярнуць увагу і на mache, і на an, каб правільна перакласці. На маю думку, у такіх выпадках трэба рабіць незалежна і паралельна два пошукі, і знаходзіць дзьве пазіцыі. Як я ўжо адзначыў, не магу сказаць, ці хто публікаваў штосьці падобнае. Зараз я больш увагі звяртаю на такія ж сістэмы, але для распазнавання маўлення, хаця магчыма хутка зноў больш часу буду траціць на машынны пераклад.
Інтэрв’ю з Дзмітрыем Багданавым, даследчыкам у галіне deep learning. Частка 1/3
Сябры, мы запісалі вельмі цікавую размову з сябрам нашай суполкі Дзмітрыем Багданавым. Ніжэй - першая частка.
Прывітанне, Дзіма. Раскажы пра сябе, калі ласка: дзе вучыўся, якія асноўныя цікаўнасці?
Я вучыўся на ФПМІ БДУ, заўседы цікавіўся праграмаваннем і матэматыкай. Падчас вучобы на палову працоўнага дня ўладкаваўся ў кампанію “Яндекс”, дзе займаўся алгарытмам ранжыравання. Больш дэталева задачу можна апісаць наступным чынам: у цябе есць вынікі пошуку, але ж яны не сартаваны, трэба вырашыць, у якім парадку яны павінны з’явіцца на старонцы карыстальніка. Насамрэч, ранжыраванне ў “Яндексе” - гэта адзін вялікі алгарытм machine learning, вельмі складаны, што патрабуе шмат людзей для таго, каб яго паляпшаць. Вось я гэтым і займаўся.
Падчас працы я заўважыў, што ў мяне больш жадання прымяняць самыя новыя, складаныя метады, напрыклад з NLProc, графічных мадэляў. Так я зразумеў, што мне ўсё ж такі трэба займацца даследчай дзейнасцю.
Таксама ў гэты час наведваў Школу Аналіза Дадзеных (ШАД). Падчас вучобы я вырашыў паступіць у Брэменскі ўніверсітэт, тым больш мяне хацеў бачыць мой прафесар Herbert Jäger.
Раскажы дэталёва пра свае поспехі ў спаборніцтвах па праграмаванні?
Найвышэйшае дасягненне адбылося ў 2012 годзе: ў камандзе разам з Юрай Пісарчыкам і Сяргеем Собалем мы атрымалі срэбраны медаль сусветнага чэмпіянату ACM/ICPC.
Ты зараз у Манрэалі. Як ты туды трапіў і чым займаешся зараз?
Да, гэта праўда. Падчас вучобы на Нямеччыне ў мяне было жаданне таксама папрацаваць у групе не толькі майго прафесара, але і ў іншых. З гэтай нагоды прафесар Jäger параіў звярнуцца да Yoshua Bengio, які якраз працуе ў Канадзе і ўзначальвае лабараторыю па Deep Learning. Yoshua дастаткова адкрыты ў плане запрашэння студэнтаў на кароткія тэрміны для сумеснай працы. Мы сустрэліся, пагутарылі, я расказаў пра свой вопыт у machine learning, і быў у выніку запрошаны на стажыроўку. У гэты перыяд мы напісалі дастаткова вядомую зараз навуковую працу "Neural Machine Translation by Jointly Learning to Align and Translate”.
Пасля гэтага я вярнуўся ў Брэмен, дзе думаў працягваць працу над падобнымі сістэмамі, але толькі для распазнавання маўлення. Ідэя наступная: нам не хочацца распрацоўваць вельмі розныя сістэмы для розных задач, якія вельмі падобныя адна на адну, нам хацелася бы выкарыстоўваць агульныя прынцыпы навучання, падобныя архітэктуры, але з умовай, што навучанне робіцца аўтаматычна. Практычна такі ж самы падыход, які быў выкарыстаны для machine translation, мы, працуючы з калегамі з Канады і Польшчы, паспяхова выкарысталі для speech recognition. Пасля магістратуры ў Германіі я паехаў працаваць над PhD да Yoshua ў Канаду. Вось зараз там і знаходжуся, працягваючы працу над тым жа падыходам: end-to-end systems, end-to-end training. Зараз стараюся знаходзіць новыя метады.
Раскажы падрабязней пра стажыроўку, калі ласка. Якія задачы ставіліся спачатку? Які кантэкст склаўся на той момант, і як ты развіваўся ў гэтым кантэксце?
Калі я прыехаў, мне расказалі пра вельмі амбіцыйны праект Neural Machine Translation, ідэя якога была, што ўвесь машынны пераклад будзе рабіцца адной нейроннай сеткай, і гэта, насамрэч, радыкальная змена пардыгмы, якая была раней. Мая першая рэакцыя была: “Якое гэта глупства! Гэта ні ў якім разе не будзе працаваць таму, што немагчыма любы сказ перавесці ў вектар лічбаў фіксаванай даўжыні.”
Але ж праект вельмі цікавы, я пачаў працаваць над ім. Першае, што я зрабіў - перапісаў рэалізацыю на працягу месяца, каб трошкі пачысціць код і можна было працаваць з сістэмай. Далей мы пісалі невялічкія працы на семінары, на варкшопы. Гэта было таксама карысна, таму што дапамагала зразумець праблемы ў нашай сістэме. У гэты час мы ўсё больш прыходзілі да высновы, што галоўная праблема - зжаць сказ рэкурэнтнай сеткай (RNN).
У мяне атрымалася прыдумаць падыход, калі нейронная сетка актыўна шукае тое, што ёй трэба ў сказе падчас перакладу. Але гэта таксама подобна на папярэднія сістэмы машыннагу перакладу: у іх заўседы есць такі элемент, які суадносіць розныя часткі сказаў. У нашым выпадку было б нашмат прасцей, калі б хтосьці дадаткова сказаў, якая частка сказу на адной мове суадносіцца з якой часткой сказу на іншай мове. Па прычыне таго, што мы не маем такой інфармацыі, класічны end-to-end падыход - дадаць яшчэ адну нейронную сетку, якая будзе гэта падлічваць (рабіць прадказанні). Гэта адразу дало нечакана добрую якасць! Мне сказалі, што мы павінны напісаць працу, таму што вынікі будуць вельмі цікавы навуковаму асяроддзю, ды і былі канкурэнты, якія працавалі і працуюць у гэтым накірунку. Мы паслалі працу на канферэнцыю ICLR 2015, і я яе прэзентаваў у Каліфорніі, San Diego.
У другой часцы размовы мы апублікуем падрабязнасці пра attention mechanism, а ў трэцяй - пра праекты Blocks і Fuel.
@YauhenMinsk для @nlprocby, фотаздымак апублікаваны з дазволу Дзімы.
У гэты раз мы запрасілі сяброў суполкі, якія кожны дзень займаюцца рэальнымі задачамі камп'ютэрнай лінгвістыкі.
Першы выступ будзе ад Аляксея Чэркеса,
"Синтез белорусской речи в проекте
Мультимедийная Библиотека Школьника."
Лёша са сваёй камандай зусім нядаўна выпусцілі сінтэзатар маўлення. Падчас прэзентацыі паабяцалі зрабіць дэма і распавесці пра тэхнічныя дэталі.
Наступны выступ будзе ад Дзяніса Пастаногава,
"NLP в системе IHS Goldfire".
Дзяніс раскажа, як рашэнні розных задач NLProc працуюць у адной сістэме.
Рыхтуем пытанні. Мы, напрыклад, плануем спытаць у Дзяніса пра праблемы такенізацыі ў кітайскай мове, а ў Лешы - які падыход да сінтэзу выкарыстоўваўся (unit selection, neural nets).
A talk to speech scientist Dr. Sébastien Le Maguer, part 2 of 2
Dear friends, we’ve recently published the first part of our nice talk to Sébastien. In this post we share with you the second part, and actually, the last of two.
This part comprises the issues on a new voice creation, a toolset for a speech scientist, neural networks, evaluation approaches.
A talk to speech scientist Dr. Sébastien Le Maguer, part 1 of 2
Hi all, we had a nice and friendly talk to Dr. Sébastien Le Maguer, one of the leading specialist in speech technologies.
Today we’d like to share the first part of the talk with our readers & listeners.
With respect to an outline, Sebastien introduced himself at start, we continued by talking about the projects he was involved in, also the main focus of past and current research. Then we discussed the current state of MaryTTS, well-known text-to-speech synthesis system, and issues of a creating a new language.
Enjoy the talk.
Some links of the mentioned in the talk: Ingmar Steiner, DFKI
The second (and last) part will be published very soon.
Topics we discussed in the second part include issues on new voice creation, a toolset for a speech scientist, neural networks, evaluation approaches.
Author: @YauhenMinsk for @nlprocby, the photo is allowed to be published by Sébastien.
Natali Karlova-Bourbonus: from philologist to Computational Linguist
Hi, Natali. Now you are in Minsk for two-days workshop about Corpus Linguistics, and also to talk about EXMARaLDA, NLTK, Python.
Hi, Yauhen. You are right. I made a lecture about basic principles and steps of speech corpus compilation on August 27, and also had a class on the tool EXMARaLDA (Extensible Markup Language for Discourse Annotation). The tool is used for annotation, transcribing and speech corpora compilation. There was also a workshop on text processing in Python, esp. using NLTK library.
Natali photo
I really appreciate all administrative work on the workshop and thankful to the National Academy of Sciences of Belarus, esp. Yury Hetsevich and all the professional team of the Speech Synthesis and Recognition Laboratory of the United Institute of Informatics Problems. It’s a big opportunity to share knowledge and know more by communicating with professionals. My presentation is available at the page.
Natali, could you tell us how did you start your way in CompLinguistics?
Sure, I started to study at the Philology faculty at Belarusian State University in Minsk, the speciality was Deutsch and German literature. But I had dreamed to study in Germany even before I entered the university. Later, when I was a third-year student, I applied for the DAAD fund stipend and got a confirmation after a half of a year and invitation to University of Constance. It’s a tourist city near the Germany-Swiss border. I usually kid that I studied in the city where others have a rest. After the half a year of exchange I decided to continue my studies and applied for BA in German Literature. BA term is 3 years, but taking into account my study in Minsk, I entered the 3rd semester (out of 6). To pay university fees I worked at several places. Now I really thankful to my parents who supported me. I started to have an interest in Computational Linguistics during my BA, including base text statistics, making a concordance lists, finding collocations, speech synthesis, machine translation was a usual topic to discuss with foreigners.
Konstanz University campus picture
Thus after the BA there were no doubts about applying for CompLing. Btw, the terms are: 3 years are for BA/BS (Bachelor of Art/ Bachelor of Science), 2 - MA/MS, 3-4 years are for PhD. I chose Technical University of Darmstadt for my Master. Also, despite the popularity of Python, we studied Java at the first three semesters of Master program. I think the reason is that we, computational linguistics, had the same core program as other students including computer scientists, mathematicians, so requirements were high.
University of Darmstadt campus
Now you do your PhD at the department of applied and computational linguistics at University of Giessen. What is your main topic of interest?
My PhD topic is Automatic detection of contradictions in English texts.The goal is ambitious obviously. The first issue is to define the concept of contradiction. But there are some works on the topic, e.g. by Stanford University. They made a corpora of real-life contradictions, it’s the first one based on real not artificial pairs. My own corpora is based on news articles. Now I work on typology of contradictions. I must admit that my students helped me a lot with annotation of the corpora. So, I really appreciate their efforts. Speaking of typology, there are at least two base types: explicit and implicit. The first stands for obvious contradictions like “The cup is empty. The cup is not empty.” Implicit is much harder to find: there are a lot of knowledge about the world and logic inference are required. But there is not empty space here also, a good start is DIRT technique.
Good. Good luck with your research. What are hot topics now in Computational Linguistics in your opinion?
Thank you. I guess, Deep Learning now is proactive. Topic Modeling was really popular recently.
Could you suggest books or web-resources to start in CompLing?
I think for beginners it’s a good book by Ruslan Mitkov. Also can suggest classic books in the field:
“Statistical NLP” by Manning and Schutze;
Daniel Jurafsky and James H. Martin “Speech and Language Processing: An Introduction”.
Thank you, Natali. Good luck!
Thank you, good luck to NLProc.by!
The short Natali’s biography can be found by link
Author: @YauhenMinsk for @nlprocby, photos are taken privately from Natali and web-resources [1 and 2]
У суботу 3/10/2015 адбылася другая сустрэча нашай суполкі.
На гэты раз першы выступ быў ад Аляксея Чэвусава пра Word-based regular expressions. З яго мы даведаліся, што рашэнне паспяхова выкарыстоўваецца ў складанай лінгвістычнай сістэме, пра яе аснову - матэматычную мадэль, а таксама некалькі цікавых фактаў. Поўную прэзентацыю можна знайсці тут.
Другая частка сустрэчы была поўнасцю пра задачу сінтэзу мовы. Аляксандр Севэрін зрабіў якасны агляд сучасных інструментаў галасавога інтэрфейсу камп’ютэра: дэманстрацыя, параўнанне па розных параметрах, шмат прыкладаў, недахопы і моцныя бакі.
Espeak, speakingmouse, new phone. Talking about modern speech synthesis with Alexander Severin #nlproc #nlprocby pic.twitter.com/bdx7Hayu44
— nlproc.by (@nlprocby)
October 3, 2015
Працягнуў тэму Толя Бабеня раcповедам пра open-source сінтэзатар RHVoice. Карысныя слайды Толі можна глянуць тут.
Anatoli Babenja is talking RHvoice at #nlprocby meetup. Speech synthesis is incredibly important task of #nlproc pic.twitter.com/63ixmdWSON
— nlproc.by (@nlprocby)
October 3, 2015
Хацелася б адзначыць, што тэма выклікала актыўную дыскусію, тым больш былі прадстаўнікі з боку камерцыйных, навуковых устаноў, а таксама актыўныя карыстальнікі сінтэзатараў і праграмісты open-source рашэння (RHVoice).
Дзякуем усім за сустрэчу!
Нагадваю, што калі есць жаданне выступіць на цікавую тэму з камп’ютэрнай лінгвістыкі, звяртайцеся (кантакты суполкі на галоўнай старонцы)
Другая сустрэча суполкі Natural Language Processing Belarus
1) Алексей Чеусов, опытный разработчик Text Analysis tools,
"Аппарат расширенных регулярных выражений WRE"
Чакаецца цікавая гутарка пра word-based regular expression ад чалавека, які ведае ўсе пра wre і шмат-шмат чаго ў сумежных тэмах
2) Анатолий Бабеня, "Open-source синтезатор речи RHVoice"
Гутарка ад Python-гуру пра выкарыстанне сучаснага сінтэзатару маўлення RHVoice: дэма, магчымасці выкарыстання. Пагутарым пра праблемы сінтэзу мовы і сучасныя прылады.
Пажадана ўзяць ноўтбукі, каб можна было запусціць і пагуляцца з тым, пра што раскажуць Леша і Толя.
Mecца: Nordic Room ў Imaguru: Бизнес-Клуб/Startup Hub
Сябры, лабараторыя распазнавання і сінтэзу маўлення падзялілася з намі фота-аглядам і матэрыяламі з майстар-класа Наталлі Карлавай-Бурбонус па корпуснай лінгвістыцы.
Прэзентацыі глядзіце па спасылках ніжэй:
1. NLTK и Python для работы с текстами
2. Компиляция и транскрибирование корпуса устной речи с EXMARaLDA
3. Корпусная лингвистика: компиляция корпуса устной речи
