В LinkedIn в одной из групп разгорелась следующая дискуссия (привожу для точности на английском языке):
John SowaIn December, Google bragged about their neural MT system (GNMT), which made a vast improvement for English to and from German, French, Spanish, and Chinese. Last week, I tested GNMT on those languages. I agree that their syntax is better, but their semantics is just as bad. See slides 25 to 29 of
http://www.jfsowa.com/talks/nlu.pdf . I just ran the examples on slide 29 last week.
For examples of *cognitive* learning, see
http://www.jfsowa.com/talks/cogmem.pdf . These slides show 4 applications implemented for *paying* clients.
RasoolDear Dr. Sowa,
I'm interested in the modern works on artificial intelligence: machine learning, etc.
As a world-renowned expert in the field of knowledge engineering, I'd appreciate your opinion on the future of the field. Particularly, do you foresee any future breakthroughs in the field of semantics in natural language processing, for example in machine translation, associated with the use of ontologies and with an increased performance of computers?
Thank you.
RasoolDid you hear about ABBYY Compreno - Natural Language Processing technology developed in russian company ABBYY?
https://www.abbyy.com/compreno/John SowaDear Rasool, John K, and anyone else. The next major breakthrough for AI and NLP requires an integration of the traditional cognitive methods of AI with both perceptual *and* cognitive learning methods. I uploaded a revised version of
http://www.jfsowa.com/talks/cogmem.pdf that has a few more slides and makes a sharper distinction between perceptual learning and cognitive learning. For related issues of neuroscience, see
http://www.jfsowa.com/talks/vrmind.pdf .
Re Google's report on GNMT: I saw it, read it, and cited it in my nlu.pdf slides. Conclusion: it does better sentence-level syntax, but it can't do paragraph-level anaphora and semantics. Try running my examples on other languages.
Re ABBYY: I downloaded their 4-page report. The only new point I saw was their hierarchy of 130,000 concepts. That is not impressive. The Kyndi technology starts with a base ontology and can derive new concepts dynamically from any number of documents -- no annotations needed.
John SowaJohn K, I didn't have room in the previous note to say more about conceptual graphs. Basic principle: the names in circles represent predicates in predicate calculus; the names in boxes represent types of entities. Since the name 'Support' is in a box, it represents an entity type. If that name were inside an oval, it would represent a relation type (predicate). If you want to quantify over verbs or refer to them with a pronoun, put them in a box rather than a circle.
For an article that describes conceptual graphs and related systems, see
http://www.jfsowa.com/pubs/eg2cg.pdf . For more detail, see the references at the end of the article -- including the ISO standard for Common Logic.
Koos VanderwiltSome comments regarding this question from Jason Baldridge and some others on Twitter (of course limited format). Prof Manning's paper on the relationship between Deep Learning and (computational) linguistics is the subject of this Twitter exchange. I recommend reading:
http://www.mitpressjournals.org/doi/pdf ... LI_a_00239Koos VanderwiltDr Sowa's philosophy seems to foreshadow some Semantic Web devlopments? Another tangent: Walid Saba has consistently argued logic and the study of language belong together in various and sundry LinkedIn discussions.
John SowaKoos, I agree with Chris Manning and Walid Saba. I also agree with Martin Kay, who said that NL understanding is an AI complete problem: it won't be accomplished until *every* major problem in AI has been solved. No single method, by itself, is sufficient for NL understanding. But if you include *every* AI method that has been developed in the past half century, you can come close.
The one additional method that the cogmem.pdf slides introduce is the ability to encode arbitrary graphs in a form that allows graph search and retrieval in log(N) time, where N is the number of graphs in your knowledge base. When you add that to your AI toolkit, you enable *all* of the traditional AI tools to scale to the size of the WWW. That is a major, major breakthrough.
Note to students: Don't neglect logic. See
http://www.jfsowa.com/pubs/fflogic.pdf .
RasoolI'm not a high skilled expert in artificial intelligence.
Could we hope that progress is possible in the field of semantics in natural language processing associated with the use of Deep Learning? The success of AlphaGo demonstrated the possibility of using neural networks in modeling multi-step processes. Also, the success of Google-translator in the syntactic translation takes place.
Perhaps, it will be possible to use new and future results of research on the work of the human brain related to the ways of storing and processing concepts and entities in the cortex?
В связи с этим я ищу, где можно почитать по теме понимания текста на естественном языке. Для себя понял, что для понимания текста человеку или машине необходимы априорные знания о мире (как правило, бессмысленно рассказывать 5-летнему ребенку о строении синхрофазотрона).
Отсюда возникает вопрос о возможности использования классических баз знаний для понимания текста, принадлежащего одной из областей знания. Задача звучит почти фантастически, поэтому встает вопрос о границах возможного в понимании смысла текста, необходимых для коммерческого применения (семантический поиск, классификация документов, машинный перевод и т.д.).
.
