I am creating a chatbot using seq2seq. Normally we remove all punctuation and stop words while processing of Text Data and feed the same to Model.
So my questions will this not impact the readability of Output?
For example - a user input some question in Chatbot window and press enter to get an answer. Now if the user gets the answer without the punctuation and stop words, will this impact the readability?
It really depends on what type of Chatbot you want to create. Generally we have two types of ChatBots:
Retrieval-based: You train your model with lots of pairs. In inference phase, your model find most similar item to training example and return to that to the user. In this case, we find most similar item between user question and our questions. Then return to most similar question's response to the user. So if we do preprocessing, on question this will not affect the readability or other things.
Generation-based: In generation based chatbot (Such as seq2seq that you mentioned), the response of the chatbot is completely rely on what you feed in for training. If you remove punctuation or stop-word, Yes it impact on your response and you can not see those things in your Chatbot response.
Of course it degrades readability. Many of those words and all of the punctuation exist to guide the reader to the intended parsing of the sentence. To put it another way:
course degrades readability many words
punctuation guide reader intended parsing
sentence put another way
There are many examples of phrases, sentences, and paragraphs that require punctuation to disambiguate the intended meaning.
Removal of "syntactic sugar" is only to promote using certain (most) techniques for quickly determining likely relevance to a similarly processed document. Your bot design has to separate this process from the user interface. Whatever you return to the user should be in human language, not the internal word soup you employ for information retrieval.
Related
I'm currently working on an internationalisation project for a large web application - initially we're just implementing French but more languages will follow in time. One of the issues we've come across is how to display adjectives.
Let's take "Active" as an example. When we received translations back from the company we're using, they returned "Actif(ve)", as English "Active" translates to masculine "Actif" or feminine "Active". We're unsure of how to display this, and wondered if there are any well established conventions in the web development world.
As far as I see it there are three possible scenarios:
We know at development time which noun a given adjective is referring to. In this case we can determine and use the correct gender.
We're referring to a user, either directly ("you") or in the third person. Short of making every user have a gender, I don't see a better approach than displaying both, i.e. "Actif(ve)"
We are displaying the adjective in isolation, not knowing which noun it's referring to. For example in a table of data, some rows might be dealing with a masculine entity, some feminine.
Scenarios 2 and 3 seem to be the toughest ones. Does anyone have any experience handling these issues? Any tips would be appreciated!
This is complex, because we cannot imagine all the cases, and there is risk to go in "opinion based" answer, so I keep it short and generic.
Usually I prefer to give context in translation (for translator), e.g. providing template: _("active {user_name}" (so also the ordering will be correct if languages want different ordering).
Then you may need to change code and template into _("active {first_name_feminine}") and _("active {first_name_masculine}") (and possibly more for duals, trials, plurals, collectives, honorific, etc.). Note: check that the translator will not mangle the {} and the string inside. Usually you need specific export/import scripts. Or I add a note inside the string, and I quickly translate into English removing the note to the translator). Also this can be automated (be creative on using special Unicode characters which should not be used in normal text, to delimit such text).
But if you cannot know the gender, the Actif(ve) may be the polite version used in such language. You need a native speaker test, and changes back and forth.
I trained my luis model to recognize an intent called "getDefinition" with example utterances such as: "What does BLANK mean" or "Can you explain BLANK to me?". It recognizes the intent correctly. I also added an entity called "topic" and trained it to recognize what topic the user is asking about. The problem is that luis only recognizes the exact topic the user is asking about if I used that specific term in one of the utterances before.
Does this mean I have to train it with all the possible terms a user can ask about or is there some way to have it recognize it anyway?
For example when I ask "What does blockchain mean" it correctly identifies the entity (topic) as blockchain because the word blockchain is in the utterance. But if I ask the same version of the question about another topic such as "what does mining mean", it doesn't recognize that as the entity.
Using a list or phrase list doesn't seem to be solving the problem. I want to eventually have thousands of topics the bot responds to, entering each topic in a list is tedious and inconvenient. Is there a way LUIS can recognize that its a topic just from the context?
What is the best way to go about this?
Same Doubt, Bit Modified. Sorry for Reposting this here.
At the moment LUIS cannot extract an entity just based on the the intent. Phrase lists will help LUIS extract tokens that don't have explicit training data. For example training LUIS with the utterance "What does blockchain mean?" does not mean that it will extract "mining" from "What does mining mean?" unless "mining" was either included in a phrase list, or a list entity. In addition to what Nicolas R said about tagging different values, another thing to consider is that using words not commonly found (or found at all) in the corpuses that LUIS uses for each culture will likely result in LUIS not extracting the words without assistance (either via Phrase list or list entity).
For example, if you created a LUIS application that dealt with units of measurement, while you might not be required to train it with units such as inch, meter, kilometer or ounce; you would probably have to train it with words like milliradian, parsec, and even other cultural spellings like kilometre. Otherwise these words would most likely not be extracted by LUIS. If a user provided the tokens "Planck unit", LUIS might provide a faulty extraction where it returns "unit" as the measurement entity instead of "Planck unit".
I have a string with urls in them.
eg: Under maintenance. Try again after sometime
I want to pass the complete string to the translators.
I see that there are two options
1. label.maintenance = Under maintenance. {0}Try again{1} after sometime (pass with anchor tags as variables)
2. label.maintenance = Under maintenance. Try again after sometime (pass the string with html)
In the first case if the translator misplaces {0} and {1} that may spoil the format of my page.
In the second case, the translator has to understand html and can possibly inject incorrect links.
What is the best way to achieve this?
I had this problem multiple times before and I came to conclusion that second option is best. Although it requires that translator know what HTML is all about, it turns out it gives the best results.
It communicates clearly what is a link here and if translator needs to re-order the sentence (which happens just too often), (s)he knows what part belongs to link.
In the first example, translator wouldn't know why the sentence contains placeholders and what to do with them. At best (s)he will ask. At worst, translator could re-order them or move translated "Try again" somewhere else and put another word between placeholders.
The conclusion is: leave HTML in translatable resources. At the same time, try to hire experienced software translators as oppose to regular translators. Chances are high they would understand Localization concepts and they would use common glossary of software terms (i.e. the one from Microsoft). That would make your software easier understandable for end users and decrease Localization defects that you will need to fix. In the end hiring software translator may cost less than "cheap" regular one...
I'm doing part-of-speech & morphological analysis project for Japanese sentences. Each sentence will have its own webpage. To make this page more visual, I want to show one picture which is somehow related to the sentence. For example, For the sentence "私は学生です" （"I'm a student"), the relevant pictures would be pictures of school, Japanese textbook, students, etc. What I have: part-of-speech tagging for every word. My approach now: use 2-3 nouns from every sentence and retrieve the first image from search results using Bing Images API. Note: all the sentence processing up to this point was done in Java.
Have a couple of questions though:
1) what is better (richer corpus & powerful search), Google Images API, Bing Images API, Flickr API, etc. for searching nouns in Japanese?
2) how do you select the most important noun from the sentence to do the query in Image Search Engine without doing complicated topic modeling, etc.?
Thanks!
Japanese WordNet has links to OpenClipart pictures. That could be another relevant source. They describe it in their paper called "Enhancing the Japanese WordNet".
I thought you would start by choosing any noun before は、が and を and giving these priority - probably in that order.
But that assumes that your part-of-speech tagging is good enough to get は=subject identified properly (as I guess you know that は is not always the subject marker).
I looked at a bunch of sample sentences here with this technique in mind and found it as good as could be expected. Except where none of those are used, which is rarish.
And sentences like this one, where you'd have to consider maybe looking for で and a noun before it in the case where there is no を or は. Because if you notice here, the word 人 (people) really doesn't tell you anything about what's being said. Without parsing context properly, you don't even know if the noun is person or people.
毎年 交通事故で 多くの人が 死にます
(many people die in traffic accidents every year)
But basically, couldn't you implement a priority/fallback type system like this?
BTW I hope your sentences all use kanji, or when you see はし (in one of the sentences linked to) you won't know whether to show a bridge or chopsticks - and showing the wrong one will probably not be good.
To make matter more specific:
How to detect people names (seems like simple case of named entity extraction?)
How to detect addresses: my best guess - find postcode (regexes); country and town names and take some text around them.
As for phones, emails - they could be probably caught by various regexes + preprocessing
Don't care about education/working experience at this point
Reasoning:
In order to build a fulltext index on resumes all vulnerable information should be stripped out from them.
P.S. any 3rd party APIs/services won't do as a solution.
The problem you're interested in is information extraction from semi structured sources. http://en.wikipedia.org/wiki/Information_extraction
I think you should download a couple of research papers in this area to get a sense of what can be done and what can't.
I feel it can't be done by a machine.
Every other resume will have a different format and layout.
The best you can do is to design an internal format and manually copy every resume content in there. Or ask candidates to fill out your form (not many will bother).
I think that the problem should be broken up into two search domains:
Finding information relating to proper names
Finding information that is formulaic
Firstly the information relating to proper names could probably be best found by searching for items that are either grammatically important or significant. I.e. English capitalizes only the first word of the sentence and proper nouns. For the gramatical rules you could look for all of the words that have the first letter of the word capitalized and check it against a database that contains the word and the type [i.e. Bob - Name, Elon - Place, England - Place].
Secondly: Information that is formulaic. This is more about the email addresses, phone numbers, and physical addresses. All of these have a specific formats that don't change. Use a regex and use an algorithm to detect the quality of the matches.
Watch out:
The grammatical rules change based on language. German capitalizes EVERY noun. It might be best to detect the language of the document prior to applying your rules. Also, another issue with this [and my resume sometimes] is how it is designed. If the resume was designed with something other than a text editor [designer tools] the text may not line up, or be in a bitmap format.
TL;DR Version: NLP techniques can help you a lot.