Extractive method for speaker identification in texts with self-training
AssigneeTencent America LLC
InventorsDian Yu, Dong Yu
Life of this patent
Estimating the term…
Abstract
A method, computer program, and computer system is provided for identifying a speaker in at text based work. Labeled and unlabeled instances corresponding to one or more speakers are extracted. Pseudo-labels are inferred for the extracted unlabeled instances based on the labeled instances. One or more of the unlabeled instances are labeled based on the inferred pseudo-labels.
Claims15
A method of identifying a speaker in a text-based work, executable by a processor, comprising:
- extracting labeled instances and unlabeled instances corresponding to one or more speakers;
- inferring pseudo-labels for the extracted unlabeled instances based on the labeled instances;
- and
- labeling one or more of the unlabeled instances based on the inferred pseudo-labels,
- wherein the labeled and unlabeled instances correspond to a class token, tokens in a first piece of text containing an utterance, a separator token, and tokens in a second piece of text that covers the first piece of text,
- wherein constructing an input sequence comprises concatenating the class token, the tokens in the first piece of text containing the utterance, the separator token, and the tokens in the second piece of text that covers the first piece of text, and
- wherein two vectors correspond to estimated probabilities of each of the tokens being a starting token or an ending token of an answer span that appears in the second piece of text.
A computer system for identifying a speaker in a text-based work, the computer system comprising:
- one or more computer-readable non-transitory storage media configured to store computer program code;
- and
- one or more computer processors configured to access said computer program code and operate as instructed by said computer program code, said computer program code including:
- extracting code configured to cause the one or more computer processors to extract labeled instances and unlabeled instances corresponding to one or more speakers;
- inferring code configured to cause the one or more computer processors to infer pseudo-labels for the extracted unlabeled instances based on the labeled instances;
- and
- labeling code configured to cause the one or more computer processors to label one or more of the unlabeled instances based on the inferred pseudo-labels,
- wherein the labeled and unlabeled instances correspond to a class token, tokens in a first piece of text containing an utterance, a separator token, and tokens in a second piece of text that covers the first piece of text,
- wherein constructing an input sequence comprises concatenating the class token, the tokens in the first piece of text containing the utterance, the separator token, and the tokens in the second piece of text that covers the first piece of text, and
- wherein two vectors correspond to estimated probabilities of each of the tokens being a starting token or an ending token of an answer span that appears in the second piece of text.
A non-transitory computer readable medium having stored thereon a computer program for identifying a speaker in a text-based work, the computer program configured to cause one or more computer processors to:
- extract labeled instances and unlabeled instances corresponding to one or more speakers;
- infer pseudo-labels for the extracted unlabeled instances based on the labeled instances;
- and
- label one or more of the unlabeled instances based on the inferred pseudo-labels,
- wherein the labeled and unlabeled instances correspond to a class token, tokens in a first piece of text containing an utterance, a separator token, and tokens in a second piece of text that covers the first piece of text,
- wherein constructing an input sequence comprises concatenating the class token, the tokens in the first piece of text containing the utterance, the separator token, and the tokens in the second piece of text that covers the first piece of text, and
- wherein two vectors correspond to estimated probabilities of each of the tokens being a starting token or an ending token of an answer span that appears in the second piece of text.
Dependencies are inferred from the claim text.