Text Analysis : Recommendation to identify cause of loss from claim narrative documents

Question

I am trying to analyze auto claims narrative documents which contain description about the accident usually free text written by claims executives. Is there a nlp technique I could use to identify cause of loss like : drunk and driving , negligence , bad weather , etc ?

I have used TF-IDF technique to rank order words per claim but it does really help to identify the exact cause of loss easily even if I concentrate on top 20% words. I am aware of word embedding but really not sure if they could help .

Also tried LDA topic modeling in the past but that gives topics for entire corpus rather than claim level.

Below is the code I used to calculate tf-idf scores for the unigrams.

library(dplyr)
library(tidytext)
library(janeaustenr)
library(tidyr) # for separate
library(qdapDictionaries)




#generate unigrams
loss_narrative_by_file_1 <- loss_narrative_by_file %>%
  unnest_tokens(unigram, whole_text_proxy, token = "ngrams", n = 1)



# removing punctuation and numbers ( removing punct / digits faster on unigrams than whole text)
loss_narrative_by_file_1$unigram <- gsub("[[:punct:]]|[[:digit:]]|^http:\\/\\/.*|^https:\\/\\/.*","",loss_narrative_by_file_1$unigram)


#retaining non blank unigrams
loss_narrative_by_file_1 <- loss_narrative_by_file_1[unigram!="",]


# retain meaningful words only
loss_narrative_by_file_1 <- loss_narrative_by_file_1[unigram %in% GradyAugmented]


# remving stopwords


unigrams_filtered <- loss_narrative_by_file_1 %>%
  filter(!unigram %in% stop_words$word)



#check coutns 

unigrams_filtered %>% 
  count(unigram,sort=TRUE)

# after cleansing
unigram       n
<chr>     <int>
  1 claim     56247
2 loss      55068
3 policy    50394
4 insured   47512
5 date      36879
6 coverage  32167
7 plaintiff 31358
8 vehicle   26156
9 umbrella  25055
10 company   24479



# tf- idf ###############################################

# concat claim & filename
unigrams_filtered$loss_file <- paste(unigrams_filtered$loss_Loss,unigrams_filtered$filename_Trimmed, sep = "_")


unigram_tf_idf <- unigrams_filtered %>%
  count(loss_file, unigram) %>%
  bind_tf_idf(unigram, loss_file, n) %>%
  arrange(desc(tf_idf))

### validating results of tf-idf shows top 30% words with highest tf-idf score tend to contain the words associated with cause of loss. How can I narrow down further ?enter code here