LINKING.md

This README page is aimed to explain the rule-based parser and entity linker we used in the system

• Entity Linking System: You can test with the entity-linking system with following command

  python parse_programs.py --preprocess
  

  This command will output the linked sentence and their corresponding columns. The procedure is described as follows:

    sent, pos_tags = self.normalize(sent)
    raw_sent = " ".join(sent)
    linked_sent, pos = self.entity_link(table_name, sent, pos_tags)
    masked_sent, mem_str, mem_num, head_str, head_num, mapping = self.initialize_buffer(table_name, linked_sent, pos, raw_sent)
  

  We first normalize the sentence to get its tokenized form and the POS-TAG, and then we run greedy search to find out the longest string match of each n-gram chunk of the sentence in the table using self.entity_link function. Finally, we will distinguish the types of the linked entities. Specifically, there are three types, namely , and .

  1. contains the names, location, numbers, etc directly seen in the table.
  2. contains the counting numbers like "2, 3, 4, ..." not directly seen in the table.
  3. contains the numeric values which are neither count nor seen in the table, like average, difference, comparison number, etc. These fine-grained three types of slots are used in the following stage. For example, when generating adversarial examples, you can replace the or or with some confusing entities from nowhere. When generating our train/val/test_lm.py, we use a coarse-grained linking to replace all these three types of entities with simply [ENT], which turns out to perform slightly better than using the fine-grained version.

• Program Searching: You can run the following command to generate the parses for the training data, which can be further used for the program ranker model (BERT-style program/text matching).

  python parse_programs.py --gen_prog
  python parse_programs.py --rank_prog_bert
  

  The searching procedure is mainly through breath-first search using the following code, which takes as input the initialized buffer from the previous entity linking system

  result = self.run(table_name, raw_sent, masked_sent, pos, mem_str, mem_num, head_str, head_num)
  

  The maximum steps of search is constrained to 7 steps, and the timeout is set to 50 seconds. The searched programs need to consume all the buffer values and return True/False in the final step.