MITIE ner model

Question

I have been exploring on using pretrained MITIE models for named entity extraction. Is there anyway I can look at their actual ner model rather than using a pretrained model? Is the model available as open source?

Answer 1

Setting things up:

For starters, you can download the English Language Model which contains Corpus of annotated text from a huge dump in a file called total_word_feature_extractor.dat.

After that, download/clone the MITIE-Master Project from their official Git.

If you are running Windows O.S then download CMake.

If you are running a x64 based Windows O.S, then install Visual Studio 2015 Community edition for the C++ compiler.

After downloading, the above, extract all of them into a folder.

Open Developer Command Prompt for VS 2015 from Start > All Apps > Visual Studio, and navigate to the tools folder, you will see 5 sub-folders inside.

The next step is to build ner_conll, ner_stream, train_freebase_relation_detector and wordrep packages, by using following Cmake commands in the Visual Studio Developer Command Prompt.

Something like this:

For ner_conll:

cd "C:\Users\xyz\Documents\MITIE-master\tools\ner_conll"

i) mkdir build ii) cd build iii) cmake -G "Visual Studio 14 2015 Win64" .. iv) cmake --build . --config Release --target install

For ner_stream:

cd "C:\Users\xyz\Documents\MITIE-master\tools\ner_stream"

i) mkdir build ii) cd build iii) cmake -G "Visual Studio 14 2015 Win64" .. iv) cmake --build . --config Release --target install

For train_freebase_relation_detector:

cd "C:\Users\xyz\Documents\MITIE-master\tools\train_freebase_relation_detector"

i) mkdir build ii) cd build iii) cmake -G "Visual Studio 14 2015 Win64" .. iv) cmake --build . --config Release --target install

For wordrep:

cd "C:\Users\xyz\Documents\MITIE-master\tools\wordrep"

i) mkdir build ii) cd build iii) cmake -G "Visual Studio 14 2015 Win64" .. iv) cmake --build . --config Release --target install

After you build them you will get some 150-160 warnings, don't worry.

Now, navigate to the "C:\Users\xyz\Documents\MITIE-master\examples\cpp\train_ner"

Make a JSON file "data.json" using Visual Studio Code for annotating text manually, something like this:

{
  "AnnotatedTextList": [
    {
      "text": "I want to travel from New Delhi to Bangalore tomorrow.",
      "entities": [
        {
          "type": "FromCity",
          "startPos": 5,
          "length": 2
        },
        {
          "type": "ToCity",
          "startPos": 8,
          "length": 1
        },
        {
          "type": "TimeOfTravel",
          "startPos": 9,
          "length": 1
        }
      ]
    }
  ]
}

You can add more utterances and annotate them, the more the training data the better is the prediction accuracy.

This annotated JSON can also be created via front-end tools like jQuery or Angular. But for brevity, I have created them by hand.

Now, to parse the our Annotated JSON file and pass it to ner_training_instance's add_entity method.

But C++ doesn't support reflection to deserialize JSON, that's why you can use this library Rapid JSON Parser. Download the package from their Git page and place it under "C:\Users\xyz\Documents\MITIE-master\mitielib\include\mitie".

Now we have to customize the train_ner_example.cpp file so as to parse our annotated custom entities JSON and pass it to MITIE to train.

#include "mitie\rapidjson\document.h"
#include "mitie\ner_trainer.h"

#include <iostream>
#include <vector>
#include <list>
#include <tuple>
#include <string>
#include <map>
#include <sstream>
#include <fstream>

using namespace mitie;
using namespace dlib;
using namespace std;
using namespace rapidjson;

string ReadJSONFile(string FilePath)
{
    ifstream file(FilePath);
    string test;
    cout << "path: " << FilePath;
    try
    {
        std::stringstream buffer;
        buffer << file.rdbuf();
        test = buffer.str();
        cout << test;
        return test;
    }
    catch (exception &e)
    {
        throw std::exception(e.what());
    }
}

//Helper function to tokenize a string based on multiple delimiters such as ,.;:- or whitspace
std::vector<string> SplitStringIntoMultipleParameters(string input, string delimiter)
{
    std::stringstream stringStream(input);
    std::string line;

    std::vector<string> TokenizedStringVector;

    while (std::getline(stringStream, line))
    {
        size_t prev = 0, pos;
        while ((pos = line.find_first_of(delimiter, prev)) != string::npos)
        {
            if (pos > prev)
                TokenizedStringVector.push_back(line.substr(prev, pos - prev));
            prev = pos + 1;
        }
        if (prev < line.length())
            TokenizedStringVector.push_back(line.substr(prev, string::npos));
    }
    return TokenizedStringVector;
}

//Parse the JSON and store into appropriate C++ containers to process it.
std::map<string, list<tuple<string, int, int>>> FindUtteranceTuple(string stringifiedJSONFromFile)
{
    Document document;
    cout << "stringifiedjson : " << stringifiedJSONFromFile;
    document.Parse(stringifiedJSONFromFile.c_str());

    const Value& a = document["AnnotatedTextList"];
    assert(a.IsArray());

    std::map<string, list<tuple<string, int, int>>> annotatedUtterancesMap;

    for (int outerIndex = 0; outerIndex < a.Size(); outerIndex++)
    {
        assert(a[outerIndex].IsObject());
        assert(a[outerIndex]["entities"].IsArray());
        const Value &entitiesArray = a[outerIndex]["entities"];

        list<tuple<string, int, int>> entitiesTuple;

        for (int innerIndex = 0; innerIndex < entitiesArray.Size(); innerIndex++)
        {
            entitiesTuple.push_back(make_tuple(entitiesArray[innerIndex]["type"].GetString(), entitiesArray[innerIndex]["startPos"].GetInt(), entitiesArray[innerIndex]["length"].GetInt()));
        }

        annotatedUtterancesMap.insert(pair<string, list<tuple<string, int, int>>>(a[outerIndex]["text"].GetString(), entitiesTuple));
    }

    return annotatedUtterancesMap;
}

int main(int argc, char **argv)
{

    try {

        if (argc != 3)
        {
            cout << "You must give the path to the MITIE English total_word_feature_extractor.dat file." << endl;
            cout << "So run this program with a command like: " << endl;
            cout << "./train_ner_example ../../../MITIE-models/english/total_word_feature_extractor.dat" << endl;
            return 1;
        }

        else
        {
            string filePath = argv[2];
            string stringifiedJSONFromFile = ReadJSONFile(filePath);

            map<string, list<tuple<string, int, int>>> annotatedUtterancesMap = FindUtteranceTuple(stringifiedJSONFromFile);


            std::vector<string> tokenizedUtterances;
            ner_trainer trainer(argv[1]);

            for each (auto item in annotatedUtterancesMap)
            {
                tokenizedUtterances = SplitStringIntoMultipleParameters(item.first, " ");
                mitie::ner_training_instance *currentInstance = new mitie::ner_training_instance(tokenizedUtterances);
                for each (auto entity in item.second)
                {
                    currentInstance -> add_entity(get<1>(entity), get<2>(entity), get<0>(entity).c_str());
                }
                // trainingInstancesList.push_back(currentInstance);
                trainer.add(*currentInstance);
                delete currentInstance;
            }


            trainer.set_num_threads(4);

            named_entity_extractor ner = trainer.train();

            serialize("new_ner_model.dat") << "mitie::named_entity_extractor" << ner;

            const std::vector<std::string> tagstr = ner.get_tag_name_strings();
            cout << "The tagger supports " << tagstr.size() << " tags:" << endl;
            for (unsigned int i = 0; i < tagstr.size(); ++i)
                cout << "\t" << tagstr[i] << endl;
            return 0;
        }
    }

    catch (exception &e)
    {
        cerr << "Failed because: " << e.what();
    }
}

The add_entity accepts 3 parameters, the tokenized string which can be a vector, the custom entity type name ,the start index of a word in a sentence and the range of the word.

Now we have to build the ner_train_example.cpp by using following commands in Developer Command Prompt Visual Studio.

1) cd "C:\Users\xyz\Documents\MITIE-master\examples\cpp\train_ner" 2) mkdir build 3) cd build 4) cmake -G "Visual Studio 14 2015 Win64" .. 5) cmake --build . --config Release --target install 6) cd Release

7) train_ner_example "C:\\Users\\xyz\\Documents\\MITIE-master\\MITIE-models\\english\\total_word_feature_extractor.dat" "C:\\Users\\xyz\\Documents\\MITIE-master\\examples\\cpp\\train_ner\\data.json"

On successfully executing the above we will get a new_ner_model.dat file which is a serialized and trained version of our utterances.

Now, that .dat file can be passed to RASA or used standalone.

For passing it to RASA:

Make the config.json file as follows:

{
    "project": "demo",
    "path": "C:\\Users\\xyz\\Desktop\\RASA\\models",
    "response_log": "C:\\Users\\xyz\\Desktop\\RASA\\logs",
    "pipeline": ["nlp_mitie", "tokenizer_mitie", "ner_mitie", "ner_synonyms", "intent_entity_featurizer_regex", "intent_classifier_mitie"], 
    "data": "C:\\Users\\xyz\\Desktop\\RASA\\data\\examples\\rasa.json",
    "mitie_file" : "C:\\Users\\xyz\\Documents\\MITIE-master\\examples\\cpp\\train_ner\\Release\\new_ner_model.dat",
    "fixed_model_name": "demo",
    "cors_origins": ["*"],
    "aws_endpoint_url": null,
    "token": null,
    "num_threads": 2,
    "port": 5000
}