Multi-hop evidence pursuit
A method for neural network training is provided. The method inputs a training set of textual claims, lists of evidence including gold evidence chains, and claim labels labelling the evidence with respect to the textual claims. The claim labels include refutes, supports, and not enough information (NEI). The method computes an initial set of document retrievals for each of the textual claims. The method also includes computing an initial set of page element retrievals including sentence retrievals from the initial set of document retrievals for each of the textual claims. The method creates, from the training set of textual claims, a Leave Out Training Set which includes input texts and target texts relating to the labels. The method trains a sequence-to-sequence neural network to generate new target texts from new input texts using the Leave Out Training Set.
1. A computer-implemented method for neural network training, comprising:
inputting a training set of textual claims, lists of evidence including gold evidence chains, and claim labels labelling the evidence with respect to the textual claims, the claim labels including refutes, supports, and not enough information (NEI);
computing an initial set of document retrievals for each of the textual claims;
computing an initial set of page element retrievals including sentence retrievals from the initial set of document retrievals for each of the textual claims;
creating, from the training set of textual claims, a Leave Out Training Set which includes input texts and target texts, where
(a) if any evidence in a gold evidence chain comes from a document not in the initial set, setting a target text to be the claim label plus a first piece of missing evidence, and setting the input text to be a list of retrieved document titles from the initial set plus a corresponding textual claim plus a concatenation of all other evidence in the gold evidence chain; and
(b) for a fraction of remaining documents, randomly selected in the initial set, including all labeled NEI, setting the claim label plus a random piece of evidence as the target text, and setting the input text to be the list of retrieved document titles plus the corresponding textual claim plus a concatenation of all other evidence in a corresponding gold evidence chain; and
training, by a processor device, a sequence-to-sequence neural network to generate new target texts from new input texts using the Leave Out Training Set.
2. The computer-implemented method of claim 1 , wherein said creating step further comprises: (c) for other documents in the initial set, setting the target to be the claim label plus the word none, and setting the input text to be the list of retrieved document titles plus the corresponding textual claim plus the concatenation of all evidence in the corresponding gold evidence chain.
3. The computer-implemented method of claim 1 , wherein computing an initial set of document retrievals for each of the textual claims in the training set comprises using a Term Frequency-Inverse Document Frequency (TFIDF) document retriever queried by the claims.
4. The computer-implemented method of claim 1 , further comprising performance an inference method subsequent to the training method, the inference method comprising:
inputting a claim with a list of retrieved evidence from a previous hop, and a list of retrieved documents;
computing an input text as a list of retrieved document titles plus the claim plus a concatenation of retrieved evidence; and
applying a trained sequence-to-sequence neural network to compute a predicted label, an imagined title, and imagined evidence.
5. The computer-implemented method of claim 4 , wherein said inference method further comprises:
retrieving a document with the imagined title if possible, or else use previous document retrievals;
finding one or more sentences or one or more other page elements including table cells in the retrieved document or document retrievals that best match the imagined evidence according to a counting-based or machine-learning based sentence matcher, ignoring existing evidence that is already in the current set of retrieved evidence; and
forming a list of retrieved evidence for a next hop by adding sentences or page elements to a top of the list of retrieved evidence from the previous hop.
6. The computer-implemented method of claim 5 , wherein said forming step further comprising removing a same number of elements from the bottom of the list of retrieved evidence from the previous hop.
7. The computer-implemented method of claim 5 , further comprising:
classifying the retrieved evidence in the list with a final label classification; and
displaying on a display device the final label classification with the retrieved evidence in the list.
8. The computer-implemented method of claim 1 , wherein the sequence-to-sequence neural network is a Unified Text to Text Transformer.
9. The computer-implemented method of claim 1 , wherein (a) if any evidence in a gold evidence chain comes from the document not in the initial set, setting the target text to be the label, except supports if the label is NEI, plus the first piece of missing evidence.
10. The computer-implemented method of claim 1 , wherein (b) for the random half of remaining documents in the initial set, including all labeled NEI, setting the label, except supports if the label is NEI, plus the random piece of evidence as the target text.
11. The computer-implemented method of claim 1 , wherein computing an initial set of page element retrievals for each of the textual claims in the training set corpus comprises using a Term Frequency-Inverse Document Frequency retriever queried by the claims.
12. The computer-implemented method of claim 1 , wherein computing an initial set of page element retrievals for each of the textual claims in the training set corpus comprises using a neural network based ranking module queried by the claims.
13. A computer program product for neural network training, the computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computer to cause the computer to perform a method comprising:
inputting, by a user interface of the computer, a training set of textual claims, lists of evidence including gold evidence chains, and claim labels labelling the evidence with respect to the textual claims, the claim labels including refutes, supports, and not enough information (NEI);
computing, by a processor device of the computer, an initial set of document retrievals for each of the textual claims;
computing, by the processor device, an initial set of page element retrievals including sentence retrievals from the initial set of document retrievals for each of the textual claims;
creating, by the processor device from the training set of textual claims, a Leave Out Training Set which includes input texts and target texts, where
(a) if any evidence in a gold evidence chain comes from a document not in the initial set, setting a target text to be the claim label plus a first piece of missing evidence, and setting the input text to be a list of retrieved document titles from the initial set plus a corresponding textual claim plus a concatenation of all other evidence in the gold evidence chain; and
(b) for a fraction of remaining documents, randomly selected in the initial set, including all labeled NEI, setting the claim label plus a random piece of evidence as the target text, and setting the input text to be the list of retrieved document titles plus the corresponding textual claim plus a concatenation of all other evidence in a corresponding gold evidence chain; and
training, by the processor device, a sequence-to-sequence neural network to generate new target texts from new input texts using the Leave Out Training Set.
14. The computer program product of claim 13 , wherein said creating step further comprises: (c) for other documents in the initial set, setting the target to be the claim label plus the word none, and setting the input text to be the list of retrieved document titles plus the corresponding textual claim plus the concatenation of all evidence in the corresponding gold evidence chain.
15. The computer program product of claim 13 , wherein computing an initial set of document retrievals for each of the textual claims in the training set comprises using a Term Frequency-Inverse Document Frequency (TFIDF) document retriever queried by the claims.
16. The computer program product of claim 13 , further comprising performance an inference method subsequent to the training method, the inference method comprising:
inputting a claim with a list of retrieved evidence from a previous hop, and a list of retrieved documents;
computing an input text as a list of retrieved document titles plus the claim plus a concatenation of retrieved evidence; and
applying a trained sequence-to-sequence neural network to compute a predicted label, an imagined title, and imagined evidence.
17. The computer program product of claim 16 , wherein said inference method further comprises:
retrieving a document with the imagined title if possible, or else use previous document retrievals;
finding one or more sentences or one or more other page elements including table cells in the retrieved document or document retrievals that best match the imagined evidence according to a counting based or machine-learning based sentence matcher, ignoring existing evidence that is already in the current set of retrieved evidence; and
forming a list of retrieved evidence for a next hop by adding sentences or page elements to a top of the list of retrieved evidence from the previous hop.
18. The computer program product of claim 17 , wherein said forming step further comprising removing a same number of elements from the bottom of the list of retrieved evidence from the previous hop.
19. A computer processing system for neural network training, comprising:
a memory device for storing program code;
a processor device operatively coupled to the memory device for running the program code to:
input a training set of textual claims, lists of evidence including gold evidence chains, and claim labels labelling the evidence with respect to the textual claims, the claim labels including refutes, supports, and not enough information (NEI);
compute an initial set of document retrievals for each of the textual claims;
compute an initial set of page element retrievals including sentence retrievals from the initial set of document retrievals for each of the textual claims;
create, from the training set of textual claims, a Leave Out Training Set which includes input texts and target texts, where
(a) if any evidence in a gold evidence chain comes from a document not in the initial set, set a target text to be the claim label plus a first piece of missing evidence, and set the input text to be a list of retrieved document titles from the initial set plus a corresponding textual claim plus a concatenation of all other evidence in the gold evidence chain; and
(b) for a fraction of remaining documents, randomly selected in the initial set, including all labeled NEI, set the claim label plus a random piece of evidence as the target text, and set the input text to be the list of retrieved document titles plus the corresponding textual claim plus a concatenation of all other evidence in a corresponding gold evidence chain; and
train a sequence-to-sequence neural network to generate new target texts from new input texts using the Leave Out Training Set.
20. The computer processing system of claim 19 , wherein said processor device further runs the program code such that creating the Leave Out Training Set further comprises: (c) for other documents in the initial set, setting the target to be the claim label plus the word none, and setting the input text to be the list of retrieved document titles plus the corresponding textual claim plus the concatenation of all evidence in the corresponding gold evidence chain.