what does the radical.txt mean?

[reid3290]

what does the radical.txt mean?

[hankcs]

Let's take '胖' as an example.

胖	胖肉半肉	月	肉	肉半

The third column is the simplified semantic radical, whereas the fourth column is the traditional form. The last column is the combination of traditional semantic radical and other components. They are not used in this paper.

[reid3290]

Gotcha, thanks.

[mzj123456]

what does the radical.ngram.vec and radical.vec mean?

[mzj123456]

How to pre-train character embedding and radical embeddings jointly using fastText?Thank you!

[hankcs]

radical.vec is actually the vector of a character in the form of its radical list.
radical.ngram.vec is the vector of each radical.

For joint pre-training, you need to

1. Modify the source code of fastText, see attached patch.
2. Convert your corpus to a sequence of characters in the form of their radical lists.
3. Set -minn 1 -maxn 1, so every individual radical vector will be outputted, no more no less.

Index: src/fasttext.cc
<+>UTF-8
===================================================================
--- src/fasttext.cc	(date 1500458693000)
+++ src/fasttext.cc	(date 1526813653000)
@@ -50,7 +50,48 @@
     ofs << word << " " << vec << std::endl;
   }
   ofs.close();
+
+  saveNgramVectors();
 }
+
+    void FastText::saveNgramVectors() {
+      std::ofstream ofs(args_->output + ".ngram.vec");
+      if (!ofs.is_open()) {
+        std::cerr << "Error opening file for saving vectors." << std::endl;
+        exit(EXIT_FAILURE);
+      }
+
+      std::unordered_map<std::string, Vector> w2v; // collect all word, subword
+      for (int32_t i = 0; i < dict_->nwords(); i++) {
+        std::string word = dict_->getWord(i);
+        std::vector<int32_t> ngrams;
+        std::vector<std::string> substrings;
+        Vector vec(args_->dim);
+        dict_->getSubwords(word, ngrams, substrings);
+        for (int32_t i = 1; i < ngrams.size(); i++) { // skip the word its self
+          vec.zero(); // reuse memory
+          if (ngrams[i] >= 0) {
+            vec.addRow(*input_, ngrams[i]);
+          }
+//          w2v[substrings[i]] = vec;
+          auto entry = w2v.emplace(::std::piecewise_construct // special to enable forwarding
+                  , ::std::forward_as_tuple(substrings[i]) // arguments for key constructor
+                  , ::std::forward_as_tuple(vec.size()) // arguments for value constructor
+          );
+          if (entry.second)
+          {
+            entry.first->second.zero();
+            entry.first->second.addVector(vec);
+          }
+        }
+      }
+      ofs << w2v.size() << " " << args_->dim << std::endl;
+      for(const auto &p : w2v)
+      {
+        ofs << p.first << " " << p.second << std::endl;
+      }
+      ofs.close();
+    }
 
 void FastText::saveOutput() {
   std::ofstream ofs(args_->output + ".output");
Index: src/fasttext.h
<+>UTF-8
===================================================================
--- src/fasttext.h	(date 1500458693000)
+++ src/fasttext.h	(date 1526813653000)
@@ -89,6 +89,8 @@
 
     void loadVectors(std::string);
     int getDimension() const;
+
+    void saveNgramVectors();
 };
 
 }

[mzj123456]

Thank you very much!

[mzj123456]

Is radical.ngram.vec the vector of each radical,not character?

[hankcs]

Yes, radical embedding. Unigram inside character is radical.

[mzj123456]

get,thanks! I have one more question.For example,the 㐥 in radical.txt not in character.vec and character-bi.vec .Isn't character.vec the vector of each character?What does the character-bi.vec and character.vec mean?

[hankcs]

radical.txt is a voluminous dictionary. What exists in a dictionary doesn't necessarily to appear in your text.

character.vec is embedding of char pre-trained without subword information (i.e. word2vec).

character-bi.vec is never used. I shouldn't have uploaded it. Now it's been removed.

[Crescentz]

hello， do you know 'python3 model.py --dataset dataset/$1-character/dataset.pkl'
where is the 'dataset'