Unsupervised-Cross-lingual-alignment-of-Knowledge-BaseTriples-with-Sentences

This project aims to utilise a high resource language ( English) to create an aligned corpus of sentences in a low resource language (Hindi) in a cross-lingual sentting. We experiment with various static embedding based methods and dynamic contextualised representations using multilingual transformer based models. These lead to strong baselines. We also propose two novel methods for improving on these baselines. We use a gold standard human annotated dataset from multple domains for our evaluation purposes. We report significant improvement on evaluating our novel methods on these datasets. This leads to the conclusion that our introduced methods are form an effective basis for cross-lingual alignment task. Utilising these alignment models, we create a cross-lingual aligned corpus of English fact triples aligned with Hindi sentences. This corpus, we believe, can be utilised for several downstream tasks like multilingual data-to-text generation, question answering, knowledge base population, knowledge graph completion etc. We make our code and final aligned corpus publicly available.

Figure 1: A sample data instance which contains a Hindi sentence mapped with relevant English triples. The English translation of the Hindi sentence has been provided for the reader’s convenience

Aligned Corpus

The aligned dataset is in the follwing json format :

{
  "sentence":"",
  "triples":[[]]
  }
Example of a data instance from the aligned corpus: 
{
  "sentence": "आशीष कक्कड़ ( 21 मई 1971 - 2 नवंबर 2020 ) एक भारतीय फिल्म निर्देशक , लेखक , अभिनेता और आवाज कलाकार थे ।",
  "triples": [
              ["Ashish Kakkad", "occupation", "film director"], 
              ["Ashish Kakkad", "occupation", "actor"],
              ["Ashish Kakkad", "occupation", "screenwriter"],
              ["Ashish Kakkad", "occupation", "artist"],
              ["Ashish Kakkad", "country of citizenship", "India"]
              ]
  }

Our aligned model predictes a set of relevant fact triples for a Hindi sentence. The statistics of the aligned corpus and gold testset are displayed below:

Domain	Entity Count	Sentence Count	Sentence Count (test data)	Avg Sentence length (test data)	Avg Fact Count (test data)
Actors	2106	5469	50	14.32	3.60
Cricketers	2316	4694	100	21.19	4.70
Politicians	3906	8916	100	18.64	3.47
Writers	2755	6629	50	15.65	1.78
Singers	739	1944	25	18.04	2.92
Journalists	607	1572	25	17.32	2.12
Total	12429	29224	350	17.52	3.08

The link to the aligned corpus and gold standard evaluation dataset is available here