Information acquisition method and apparatus, device, and medium
The present disclosure relates to an information obtaining method and apparatus, a device, and a medium. The method includes setting a label table corresponding to each sample sentence in a sample set, wherein row characters and column characters in the label table are set identically in accordance with an order of characters of the corresponding sample sentence; and marking cells composed of the row characters and the column characters in the label table with corresponding information category labels; taking each sample sentence in the sample set as input information to a model to be trained and the label table corresponding to each sample sentence as output information of the model to be trained, and performing model training according to a preset target function; and generating an information extraction model based on parameters of the trained model to extract target sentence information by the information extraction model.
1 . An information obtaining method, comprising:
setting a label table corresponding to each sample sentence in a sample set, wherein row characters and column characters in the label table are set identically in accordance with an order of characters of the corresponding sample sentence;
marking cells composed of the row characters and the column characters in the label table as having or not having entity category labels or relationship category labels;
taking each sample sentence in the sample set as input information to a model to be trained and the label table corresponding to each sample sentence as output information of the model to be trained, and performing model training according to a preset target function;
processing characters of target sentence according to an entity relationship extraction model to generate a target label table corresponding to the target sentence, wherein the target label table comprises predicted scores corresponding to the cells composed of the row characters and the column characters;
calculating the predicted scores corresponding to the cells composed of the row characters and the column characters in the target label table to determine segmentation positions related to the entity category labels in the target label table;
determining whether the predicted scores for the cells in a diagonal line in the target label table meet the entity category labels according to the segmentation positions, and if yes, marking with the corresponding entity category labels;
determining whether the predicted scores for the cells not in the diagonal line in the target label table meet the relationship category labels according to the segmentation positions, and if yes, marking with the corresponding relationship category labels; and
generating entities of the target sentence and a relationship between entities of the target sentence according to the entity category labels and the relationship category labels marked in the target label table.
2 . The information obtaining method according to claim 1 , further comprising:
if the sample sentence is an English sentence, identically setting the row characters and the column characters of the label table in accordance with an order of words of the corresponding English sentence; and
if the sample sentence is a Chinese sentence, identically setting the row characters and the column characters of the label table in accordance with the order of characters of the corresponding Chinese sentence.
3 . The information obtaining method according to claim 1 , wherein the marking the cells composed of the row characters and the column characters in the label table as having or not having the entity category labels or the relationship category labels comprises:
marking the cells composed of the row characters and the column characters in the diagonal line of the label table as having or not having the entity category labels; and
marking the cells composed of the row characters and the column characters not in the diagonal line of the label table as having or not having the relationship category labels, wherein the relationship category labels are distributed symmetrically or asymmetrically about the diagonal line of the label table.
4 . The information obtaining method according to claim 1 , wherein the preset target function corresponding to the entity relationship extraction model comprises a first loss function which is determined by:
determining a label classification probability representing the entity category label or the relationship category label corresponding to each cell in the label table corresponding to each sample sentence according to a biaffine classification algorithm; and
calculating the label classification probability representing the entity category label or the relationship category label corresponding to each cell to determine the first loss function.
5 . The information obtaining method according to claim 4 , wherein the preset target function corresponding to the entity relationship extraction model comprises a second loss function which is determined by:
calculating entity category label probabilities in the diagonal line of the label table and relationship category label probabilities distributed symmetrically about the diagonal line to determine the second loss function.
6 . The information obtaining method according to claim 5 , wherein the preset target function corresponding to the entity relationship extraction model comprises a third loss function which is determined by:
calculating entity category label probabilities in the diagonal line of the label table according to a preset algorithm to determine the third loss function.
7 . An electronic device, comprising:
a processor; and
a memory configured to store instructions executable by the processor,
wherein the processor is configured to read the executable instructions from the memory and execute the instructions to implement an information obtaining method, wherein the method comprises:
setting a label table corresponding to each sample sentence in a sample set, wherein row characters and column characters in the label table are set identically in accordance with an order of characters of the corresponding sample sentence;
marking cells composed of the row characters and the column characters in the label table as having or not having entity category labels or relationship category labels;
taking each sample sentence in the sample set as input information to a model to be trained and the label table corresponding to each sample sentence as output information of the model to be trained, and performing model training according to a preset target function;
processing characters of target sentence according to an entity relationship extraction model to generate a target label table corresponding to the target sentence, wherein the target label table comprises predicted scores corresponding to the cells composed of the row characters and the column characters;
calculating the predicted scores corresponding to the cells composed of the row characters and the column characters in the target label table to determine segmentation positions related to the entity category labels in the target label table;
determining whether the predicted scores for the cells in a diagonal line in the target label table meet the entity category labels according to the segmentation positions, and if yes, marking with the corresponding entity category labels;
determining whether the predicted scores for the cells not in the diagonal line in the target label table meet the relationship category labels according to the segmentation positions, and if yes, marking with the corresponding relationship category labels; and
generating entities of the target sentence and a relationship between entities of the target sentence according to the entity category labels and the relationship category labels marked in the target label table.
8 . The electronic device according to claim 7 , wherein the method further comprises:
if the sample sentence is an English sentence, identically setting the row characters and the column characters of the label table in accordance with an order of words of the corresponding English sentence; and
if the sample sentence is a Chinese sentence, identically setting the row characters and the column characters of the label table in accordance with the order of characters of the corresponding Chinese sentence.
9 . The electronic device according to claim 7 , wherein the marking the cells composed of the row characters and the column characters in the label table as having or not having the entity category labels or the relationship category labels comprises:
marking the cells composed of the row characters and the column characters in the diagonal line of the label table as having or not having the entity category labels; and
marking the cells composed of the row characters and the column characters not in the diagonal line of the label table as having or not having the relationship category labels, wherein the relationship category labels are distributed symmetrically or asymmetrically about the diagonal line of the label table.
10 . The electronic device according to claim 7 , wherein the preset target function corresponding to the entity relationship extraction model comprises a first loss function which is determined by:
determining a label classification probability representing the entity category label or the relationship category label corresponding to each cell in the label table corresponding to each sample sentence according to a biaffine classification algorithm; and
calculating the label classification probability representing the entity category label or the relationship category label corresponding to each cell to determine the first loss function.
11 . The electronic device according to claim 10 , wherein the preset target function corresponding to the entity relationship extraction model comprises a second loss function which is determined by:
calculating entity category label probabilities in the diagonal line of the label table and relationship category label probabilities distributed symmetrically about the diagonal line to determine the second loss function.
12 . The electronic device according to claim 11 , wherein the preset target function corresponding to the entity relationship extraction model comprises a third loss function which is determined by:
calculating entity category label probabilities in the diagonal line of the label table according to a preset algorithm to determine the third loss function.
13 . A non-transitory computer-readable storage medium, storing a computer program configured to perform an information obtaining method, wherein the method comprises:
setting a label table corresponding to each sample sentence in a sample set, wherein row characters and column characters in the label table are set identically in accordance with an order of characters of the corresponding sample sentence;
marking cells composed of the row characters and the column characters in the label table as having or not having entity category labels or relationship category labels;
taking each sample sentence in the sample set as input information to a model to be trained and the label table corresponding to each sample sentence as output information of the model to be trained, and performing model training according to a preset target function;
processing characters of target sentence according to an entity relationship extraction model to generate a target label table corresponding to the target sentence, wherein the target label table comprises predicted scores corresponding to the cells composed of the row characters and the column characters;
calculating the predicted scores corresponding to the cells composed of the row characters and the column characters in the target label table to determine segmentation positions related to the entity category labels in the target label table;
determining whether the predicted scores for the cells in a diagonal line in the target label table meet the entity category labels according to the segmentation positions, and if yes, marking with the corresponding entity category labels;
determining whether the predicted scores for the cells not in the diagonal line in the target label table meet the relationship category labels according to the segmentation positions, and if yes, marking with the corresponding relationship category labels; and
generating entities of the target sentence and a relationship between entities of the target sentence according to the entity category labels and the relationship category labels marked in the target label table.
14 . The non-transitory computer-readable storage medium according to claim 13 , wherein the method further comprises:
if the sample sentence is an English sentence, identically setting the row characters and the column characters of the label table in accordance with an order of words of the corresponding English sentence; and
if the sample sentence is a Chinese sentence, identically setting the row characters and the column characters of the label table in accordance with the order of characters of the corresponding Chinese sentence.
15 . The non-transitory computer-readable storage medium according to claim 13 , wherein the marking the cells composed of the row characters and the column characters in the label table as having or not having the entity category labels or the relationship category labels comprises:
marking the cells composed of the row characters and the column characters in the diagonal line of the label table as having or not having the entity category labels; and
marking the cells composed of the row characters and the column characters not in the diagonal line of the label table as having or not having the relationship category labels, wherein the relationship category labels are distributed symmetrically or asymmetrically about the diagonal line of the label table.
16 . The non-transitory computer-readable storage medium according to claim 13 , wherein the preset target function corresponding to the entity relationship extraction model comprises a first loss function which is determined by:
determining a label classification probability representing the entity category label or the relationship category label corresponding to each cell in the label table corresponding to each sample sentence according to a biaffine classification algorithm; and
calculating the label classification probability representing the entity category label or the relationship category label corresponding to each cell to determine the first loss function.
17 . The non-transitory computer-readable storage medium according to claim 16 , wherein the preset target function corresponding to the entity relationship extraction model comprises a second loss function which is determined by:
calculating entity category label probabilities in the diagonal line of the label table and relationship category label probabilities distributed symmetrically about the diagonal line to determine the second loss function.
18 . The non-transitory computer-readable storage medium according to claim 17 , wherein the preset target function corresponding to the entity relationship extraction model comprises a third loss function which is determined by:
calculating entity category label probabilities in the diagonal line of the label table according to a preset algorithm to determine the third loss function.