R/derekovecs_apicall_syn_nlc.R

# Get classic and w2v based collocates from DeReKoVecs API 
#
# Have a look at the web gui column header mouse overs to figure out what the columns mean:
# http://corpora.ids-mannheim.de/openlab/derekovecs?word=triftiger&cutoff=500000&n=100&N=2000#tabs-3

library(httr)
library(tidyverse)

stopwords <- readLines(con = "../data/stopwords.txt",encoding="UTF-8")

# ngramme <- read.csv("../data/gold03_anno_ml.csv", quote="", header = TRUE, sep = "\t", dec=",", encoding="UTF-8")

ngramme <- read.csv("../data/goldstandard01_anno_ml.tsv", quote="", header = TRUE, sep = "\t", dec=",", encoding="UTF-8")


ngramme[, c(7:23)] <- sapply(ngramme[, c(7:23)], as.numeric)

DeReKoVecsCall <- function(wordform = "Grund", cutoff = 500000, n = 100) {
  params = list(word = wordform, cutoff = cutoff, n = n, json = 1)
  response <- tryCatch({httr::GET(url = 'http://korap-worker-04:5673', query = params,timeout(10))},
                       error=function(cond) return(NA))
  if(!is.na(response)) {
    content(response, as = 'parsed', type = 'application/json', simplifyDataFrame = TRUE)
  }
  else {
    return(NA)
  }
}

getW2VCollocates <- function(wordform = "Grund", ...) {
  ret<-DeReKoVecsCall(wordform, ...)
  if(!is.na(ret)) ret$collocators
  else NA
}

getClassicCollocates <- function(wordform = "Grund") {
  response <- tryCatch({httr::GET(url = 'http://korap-worker-04:5673/getClassicCollocators', query = list(w=wordform),timeout(10))},
                       error=function(cond) return(NA))
  if(!is.na(response)) {
    content(response, as = 'parsed', type = 'application/json', simplifyDataFrame = TRUE)$collocates
  }
  else {
    return(NA)
  }
}


getBoth <- function(wordform = "Schmetterlinge") {
  w2v <- getW2VCollocates(wordform)
  classic <- getClassicCollocates(wordform)
  if (length(w2v)>0 && length(classic>0) && !is.na(w2v) && !is.na(classic)) {
    merge(classic,w2v,by="word",all=TRUE)
  }
  else {
    return(NA)
  }
}

getRanks <- function(collocates,NApenalty=1000) {
  if (is.na(collocates)) {
    return(NA)
  }
  both <<-
    collocates %>%
    arrange(desc(cprob)) %>%
    mutate(w2v.rank = 1:nrow(.)) %>%
    mutate(w2v.rank = ifelse(is.na(cprob), NApenalty, w2v.rank)) %>%
    arrange(desc(prob)) %>%
    mutate(w2v.rank1 = 1:nrow(.)) %>%
    mutate(w2v.rank1 = ifelse(is.na(prob), NApenalty, w2v.rank1)) %>%
    arrange(desc(ld)) %>%
    mutate(classic.rank = 1:nrow(.)) %>%
    mutate(classic.rank = ifelse(is.na(ld), NApenalty, classic.rank))
  return(both)
}

# rankDiff with precalculated collocates

rankDiff1 <- function(wordform = "Schmetterlinge", collocate = "Bauch", both) {
  if (is.na(both)) {
    return(NA)
  }
  if (! collocate %in% both$word)
    return(NA)
  w2vRank <- both[both$word==collocate,]$w2v.rank
  classicRank <- both[both$word==collocate,]$classic.rank
  if (is.numeric(w2vRank) && is.numeric(classicRank)) {w2vRank - classicRank}
  else {NA}
}

featurenames<-c("af", "dice","ld","lfmd","llfmd","llr","lnpmi" ,"md","npmi","pmi","rlfmd","rnpmi","average","cprob","max","overall","prob","w2v.rank","w2v.rank1", "classic.rank")

getFeatures <- function(wordform = "Schmetterlinge", collocate = "Bauch", both) {
  if (is.na(both)) {
    return(as.numeric(rep(c(featurenames[0], NA), length(featurenames)+1)))
  }
  if (! collocate %in% both$word)
    return(as.numeric(rep(c(featurenames[0], NA), length(featurenames)+1)))
  features<-both[both$word==collocate,]
  w2vRank <- features$w2v.rank
  classicRank <- features$classic.rank
  rankDiff<-NA
  if (is.numeric(w2vRank) && is.numeric(classicRank)) {rankDiff<-w2vRank - classicRank}
  return(as.numeric(unlist(c(rankDiff,features[1,featurenames]))))
}

#w2v <- getW2VCollocates("triftiger")
#classic <- getClassicCollocates("triftiger")
#rankDiff("Schmetterlinge", "Bauch")

firstup <- function(x) {
  substr(x, 1, 1) <- toupper(substr(x, 1, 1))
  x
}

# rankDiffCase with precalculated collocates

rankDiffCase1<-function(wordform="Schmetterlinge",collocate="Bauch",collocates1,collocates2) {
  if (wordform=="" || collocate=="") {
    return(NA)
  }
  ret<-rankDiff1(wordform,collocate,collocates1)
  if (is.na(ret)) {
    ret<- rankDiff1(firstup(wordform),collocate,collocates2)
    if (is.na(ret)) {
      ret <- rankDiff1(wordform, firstup(collocate),collocates1)
      if (is.na(ret)) {
        ret <- rankDiff1(firstup(wordform),firstup(collocate),collocates2)
      }
    }
  }
  return(ret)
}

getFeaturesCase<-function(wordform="Schmetterlinge",collocate="Bauch",collocates1,collocates2) {
  if (wordform=="" || collocate=="") {
    return(NA)
  }
  ret1<-getFeatures(wordform,collocate,collocates1)
  ret2<-getFeatures(firstup(wordform),collocate,collocates2)
  ret3<-getFeatures(wordform,firstup(collocate),collocates1)
  ret4<-getFeatures(firstup(wordform),firstup(collocate),collocates2)
  ret<-rbind(ret1,ret2,ret3,ret4)
  colnames(ret)<-c("rankDiff",featurenames)
  colMeans(ret,na.rm=TRUE)
}

# get rid of some "offending" characters

cleanUp<-function(str) {
  ret<-str_replace_all(str,"[^[:alnum:]]"," ")
  ret<-str_replace_all(ret,"[\n\r\t]"," ")
  ret<-str_replace_all(ret,"[ ]+"," ")
  ret<-str_replace_all(ret,"^ ","")
  ret<-str_replace_all(ret, " $","")
  ret
}

cleanUpV<-Vectorize(cleanUp)

deleteStopwords =  function(wl, stopwords = NULL) {
  wl[!(wl %in% stopwords)]
}

ngramme$tokens <- cleanUpV(paste(ngramme$CO_TOKEN1, ngramme$CO_TOKEN2,ngramme$CO_TOKEN3,ngramme$CO_TOKEN4,ngramme$CO_TOKEN5,ngramme$CO_TOKENS6,sep=" "))

# retrieve collocates only once per token.

avgRankDiffCaseLinear<-function(wordlist) {
  wl<-deleteStopwords(tolower(unlist(strsplit(wordlist," "))),stopwords)
  sum<-0
  count<-0
  if (length(wl)>1) {
  for (i in 1:length(wl)) {
    collocates1<-getRanks(getBoth(wl[i]))
    collocates2<-getRanks(getBoth(firstup(wl[i])))
    for(j in 1:length(wl)) {
      if (i!=j) {
        rd<-rankDiffCase1(wl[i],wl[j],collocates1,collocates2)
        if (!is.na(rd)) {
          sum=sum+rd
          count=count+1
        }
      }
    }
  }
  }
  ret<-ifelse(count>0,sum/count,NA)
  print(paste(wordlist, ": ", ret))
  return(ret)
}

avgRankDiffCaseLinearV<-Vectorize(avgRankDiffCaseLinear)


avgFeaturesCaseLinear<-function(wordlist) {
  wl<-deleteStopwords(tolower(unlist(strsplit(wordlist," "))),stopwords)
  ret<- data.frame(matrix(ncol = length(featurenames)+1, nrow = 0))
  if (length(wl)>1) {
    for (i in 1:length(wl)) {
      collocates1<-getRanks(getBoth(wl[i]))
      collocates2<-getRanks(getBoth(firstup(wl[i])))
      for(j in 1:length(wl)) {
        if (i!=j) {
          ret<-rbind(ret,getFeaturesCase(wl[i],wl[j],collocates1,collocates2))
        }
      }
    }
  }
  colnames(ret)<-c("rankDiff",featurenames)
  mret<-colMeans(ret,na.rm=TRUE)
  print(paste(c(wordlist, ": ", mret)))
  return(mret)
}

ngramme[,c("rankDiff",featurenames)]<-t(sapply(ngramme$tokens,avgFeaturesCaseLinear))
write.table(ngramme,file="../data/goldstandard01_anno_ml_synfeat_nstop1.tsv", sep = "\t",quote=F)