Hacker News Reader

Comment by marginalia_nu | Hacker News Reader

I think some NLP is strictly beneficial for a search engine. You may think "grep for the web" sounds like a good idea, but let me tell you, having tried this, manually going through every permutation of plural forms of words and manually iterating the order of words to find a result is a chore and a half.

Like, instead of trying

  PDP11 emulator
  PDP-11 emulator
  "PDP 11" emulator
  PDP11 emulators
  PDP-11 emulators
  "PDP 11" emulators
  PDP11 emulation
  PDP-11 emulation
  "PDP 11" emulation

Basic NLP can do that a lot faster without introducing a lot of problems.

I do think Google currently goes way overboard with the NLP. It often feels like the query parser is an adversary you need to outsmart to get to the good results, rather than something that's actually helpful. That's not a great vibe. However, I think the big problem isn't what they are doing, but how little control you have over the process.

View on HN

I get that for general-purpose searches this is a good idea, but it would be nice if there was an easy way to disable this when you know you don't want it - for example, for most programming searches, if I type SomeAPINameHere the most relevant results will always be those that include my search term verbatim. I don't need Google to helpfully suggest "Did you mean Some API Name Here?", which will virtually always return lower-quality search results.

Early Google was a breath of fresh air compared to the stemming that its competitors at the time did, but nowadays even putting search terms in quotes doesn't seem to return the same quality of results for these types of queries that Google used to have.

I feel your pain. Two workarounds when Google gets it wrong are to put the term in quotation marks, or to enable Verbatim mode in the toolbelt. (I know various people have come up with ways to add "Google Verbatim" as a search engine option in their browser, or use a browser extension to make Verbatim enabled by default.)

Disclaimer: I work on Google search.

Both of these options are disappointing, in my experience. Verbatim mode seems weirdly broken sometimes (maybe it's overly strict), and quoting things is rarely enough to convince Google that you really want to search for exactly that thing and not some totally different thing that it considers to be a synonym.

One porridge is too hot and the other is too cold. I know Google could find a happy compromise here if it wanted to. In fact, I bet there's some internal-only hacked-together version that works this way and actually gives an acceptable user experience for the kind of people who have shown up to this thread to show their dissatisfaction.

Try this, go to Google and type in "eggzackly this".

Two results not containing "eggz" at all. Two results containing "eggzackly<punctuation>this" Two results containing "eggzackly" but missing "this".

Google Search is broken. It no longer does what it's directed, it just takes a guess. I suspect part of this is because someone decided that "no results found" was the worst possible result a search engine could give.

Googling that with the brackets I get results containing "eggzackly this" ranked 3, 4, 6 (your comment) and 7 whereas the others contain just eggzackly (or with the 'this' preceded by punctuation as you mention).

Therefor I don't see how your last sentence is the explanation (there are results), I've also happened to land on no results found sometimes with overly precise quoted queries (for coding errors mostly IIRC). But it is annoying that it doesn't seem stricktly enforced even when you want it to.

Google does go way overboard with "NLP". Starting at least 5 years ago there was a trend toward "similar" matching and search result quality nose-dived.

You can search for, say, "cycling (insert product category here)" and get motorcycle related results. Why? Because to google "cycling" = "biking" and "motorcycles" are "bikes", bob's your uncle, now you're getting hits for motorcycle products.

Every time I try to do a very specific search I can see from the search results how google tries to "help", especially if the topic is esoteric. The pages actually about the esoteric thing I'm searching for get drowned in a sea of SEO'd bullshit about a word/topic that is 1-2 degrees of separation from each other in a thesaurus. I'm sure someone at google is very, very proud of this because it increases their measure for search user satisfaction X percent.

It does this thesaurus crap even with words in quotes, which is especially infuriating.

Yeah. It's one of those things where it's invisible where it works and enraging when it doesn't. That's generally not a failure mode that's desirable. It at least should require extremely low failure rates to motivate.

"Basic NLP can do that a lot faster without introducing a lot of problems."

This is called "stemming" and is not sensibly approached with machine learning.

Of course, but stemming is a fairly basic technique in NLP, as is POS-tagging. NLP is not machine learning.

Modern NLP basically is machine learning

You can still do NLP without machine learning though, and a lot of the sorts of computational linguistics a search engine needs for keyword extraction and query parsing doesn't require particularly fancy algorithms. What it needs is fast algorithms, and that's not something you're gonna get with ML.

Stemming is not meaningfully a natural language processing technique, any more than arithmetic is a technique of linear equations.

Is it not the processing of natural language?

Would you call addition a system of linear equations?

No, you don't use the college senior label for the highschool freshman topic. You use the smallest label that fits.

It's string processing.

NLP is actually understanding the language. Stemming is simple string matching.

Playing the technicality game to stretch fields to encompass everything you think even marginally related isn't being thorough or inclusive; it's being bloated, and losing track of the meaning of the term.

Splitting on spaces also isn't NLP.

Stemming is a task specific to a natural language. You can't run an English stemmer on French and get good results, for example.

All NLP is, strictly speaking, more or less elaborate string matching.

> Splitting on spaces also isn't NLP.

String splitting can be, but it's a bit borderline. I'll argue you're in NLP territory if it doesn't split "That FBI guy i.e. J. Edgar Hoover." into four "sentences".

> NLP is actually understanding the language.

That's actually not an accepted terminology. There's, indeed, this:

  https://en.wikipedia.org/wiki/Natural-language_understanding

Not sure why are you so adamant that yours is the "true meaning", when NLP existed long before machine learning and AI were used for it. And even if not, every term can be defined differently, so it should be normal to have different institutions/people define NLP differently.

At the very least, https://en.wikipedia.org/wiki/Natural_language_processing seems to disagree.

(So do I: NLP does not have to be machine learning/AI based)

Semantic search requires NLP. So does the Q&A format the OP is complaining about. People conflate all things NLP to the latter, and forget about the former.

Maybe I'm not using the right qualifiers around the term NLP. The kind of NLP I was referring to is something like "Hey google, what is natural language processing?" and orienting the search around people asking questions in standard(ish) English like they would to another person.

NLP is very heavily integrated into search, so I don't think it's really possible to decouple them. But I agree the whole BonziBuddy thing they've got going now is annoying and it's especially unfortunate how it's replaced the search functionality. I'd have a lot more patience with it if I could choose this functionality when I wanted to ask a question.

That's known as Open Domain Question Answering[1] and is only a subset of NLP.

[1] https://www.pinecone.io/learn/question-answering/

Reply on news.ycombinator.com