Generative event extraction method based on ontology guidance

Patent metadata

Field	Value
Publication number	US-12222970-B2
Application number	US-202218280655-A
Country	US
Kind code	B2
Filing date	Sep 23, 2022
Priority date	Sep 28, 2021
Publication date	Feb 11, 2025
Grant date	Feb 11, 2025

Abstract

The present invention is a generative event extraction method based on ontology guidance, including: (1) constructing an event ontology knowledge base; (2) designing an event trigger word extraction template and an event argument extraction template; mapping an input event text to a first input sequence, and mapping an input event text integrating an event ontology to a second input sequence; (3) designing a class label mapping function that maps multi-word labels to event types and/or role types; (4) extracting the event ontology corresponding to the input event from the event ontology knowledge base, and constructing the first input sequence and the second input sequence according to the event trigger word extraction template and the event argument extraction template; and (5) predicting, by the event extraction model, the event type and the role type according to the class label mapping function and a processing mechanism thereof, and outputting an event trigger word span and an event argument span.

First claim

The invention claimed is:

1. A generative event extraction method based on ontology guidance, comprising the following steps:
step 1: constructing an event ontology knowledge base based on a domain knowledge base and an event annotation framework;
step 2: designing an event trigger word extraction template and an event argument extraction template for the generative event extraction; the event trigger word extraction template maps an input event text to a first input sequence of an event extraction model; and the event argument extraction template maps the input event text integrating an event ontology to a second input sequence of the event extraction model;
step 3: designing a class label mapping function that handles a mapping of multi-word labels to event type or role type;
step 4: for the input event text, extracting the event ontology corresponding to the input event from the event ontology knowledge base, and according to the input event text and the event ontology, constructing the first input sequence and the second input sequence according to the event trigger word extraction template and the event argument extraction template; and
step 5: inputting the first input sequence and the second input sequence into the event extraction model, and predicting, by the event extraction model, the event type and the role type according to the class label mapping function and a processing mechanism thereof, and outputting an event trigger word span and an event argument span.

2. The generative event extraction method based on ontology guidance according to claim 1 , wherein in step 1, a process of constructing an event ontology knowledge base is as follows:
step 1.1: taking an automatic content extraction pre-defined event framework as a target event ontology;
step 1.2: extracting an event framework related to the target event ontology in FrameNet as an expanded event ontology; and
step 1.3: integrating the target event ontology and the expanded event ontology, and carrying out de-duplication and manual inspection, so as to obtain the event ontology knowledge base.

3. The generative event extraction method based on ontology guidance according to claim 1 , wherein in step 2, the designed event trigger word extraction template is as follows:
[first marker]<Pseudo template > <Input event text> [second marker] The event trigger word is [MASK], a trigger word token is, the corresponding English word is [CLS]<pseudo template><input sentence> [SOS] The trigger word is [MASK], trigger token is;
wherein the pseudo template takes unused dummy tags embedded in pre-trained words.

4. The generative event extraction method based on ontology guidance according to claim 1 , wherein in step 2, the designed event argument extraction template is as follows:
[first marker]<event ontology> <Input event text> [second marker] argument type is [MASK], the argument token is, the corresponding English word is [CLS]<Event ontology> <input sentence> [SOS] The argument type is [MASK], argument token is.

5. The generative event extraction method based on ontology guidance according to claim 1 , wherein in step 3, the designed class label mapping function is as follows:
Y ( r i )={ w 1 , w 2 , . . . , w n }

in the event type prediction, Y(r i ) represents the mapping function between an i-th event type and the multi-word label, and Y(r i ) represents a word embedding vector of an n-th word label of the event type; and
in the role type prediction, Y(r i ) represents the mapping function between an i-th role type and the multi-word label, and w n represents a word embedding vector of an n-th word label of the role type.

6. The generative event extraction method based on ontology guidance according to claim 1 , wherein in step 5, the event extraction model adopts a codec Transformer framework;
when making prediction of the event type or role type using the event extraction model, the following formula is used to obtain prediction probabilities of the event type or role type:

p
⁡
(

r
i

)

=

∑

w
∈

r
i

⁢

exp
⁡
(

w
·

h

[
MASK
]

)

∑

w
′

∈
R

⁢

exp
⁡
(

w
′

·

h

[
MASK
]

)

wherein p(r i ) represents a prediction probability of an ri event type or a role label p(r i ), h [MASK] represents an output vector corresponding to the event extraction model at [MASK], w represents a word embedding vector of a word label of a target event type/role type, w′ represents a word embedding vector of word labels of all event types/role types, and R represents a set of all event types/role labels;
when the event extraction model is used to predict the event trigger word span and/or the event argument span, predictions of the event trigger word span and/or the event argument span are modeled as a sequence generation task. For an event text set S, the input event text is A and an associated event ontology is O, a conditional distribution p(H|A,O;θ E s ,θ G s ) of integrated event ontology is learned by training Z=E s (A,O) and H=G s (z), wherein Z is a potential source domain representation obtained by learning in an encoder E, H represents a potential source domain representation obtained by learning in a decoder G, H and θ G s represent model parameter sets of the encoder and decoder in the source domain, p(H|A, O;θ E s ,θ G s ) represents an overall probability of an output sequence H generated by the given input event text A and the associated event ontology O, wherein

p
⁡
(

H
❘
A

,

O
;

Assignees

• Univ Zhejiang

Inventors

• Chen Huajun
• Ye Hongbin
• Zhang Ningyu
• Deng Shumin
• Bi Zhen

Classifications

• G06N5/02

  Knowledge representation; Symbolic representation · CPC title

• G06F16/35

  Clustering; Classification · CPC title

• G06N20/00

  Machine learning · CPC title

• G06N3/08

  Learning methods · CPC title

• G06F40/289

  Phrasal analysis, e.g. finite state techniques or chunking · CPC title