Term extraction in highly technical domains

1. A computer-implemented method for fine-tuning of a transformer-based language model by extracting terminology terms from a text document, said method comprising:

identifying a text snippet in a text document;

identifying candidate multi-word expressions in the identified text snippet using part of speech tags;

determining a specificity score value for each of the candidate multi-word expressions, wherein the specificity score value is indicative of how unique the candidate multi-word expression is within the text snippet;

determining a topic similarity score value for each of the candidate multi-word expressions as a distance between an embedding vector corresponding to the identified text snippet and an embedding vector corresponding to the candidate multi-word expression;

selecting remaining expressions from the candidate multi-word expressions using a function of a specificity value and a topic similarity value of each of the candidate multi-word expressions;

adding a noun comprised in the text snippet to the remaining expressions depending on a function of a correlation of the noun with a same lemma as one of the remaining expressions and a noun morphology score value of the noun to be added;

labeling the remaining multi-word expressions; and

fine-tuning an existing pre-trained transformer-based language model using as training data the identified text snippet marked with the labeled remaining expressions.

2. The method according to claim 1 , wherein selecting remaining expressions using a function of the specificity threshold value and a topic similarity value comprises selecting remaining multi-word expressions from the candidate multi-word expressions if the specificity threshold value is greater than a predefined specificity threshold value and if the topic similarity value is greater than a topic similarity threshold value.

3. The method according to claim 1 , wherein selecting remaining expressions using a function of the specificity value and a topic value comprises selecting a noun of the identified text snippet if the noun has a score value that is greater than the lowest related multi-word expression score value.

4. The method according to claim 3 , wherein the selecting the noun if is score value is greater than the lowest related multi-word expression score value is used as a probabilistic approach to noun selection.

5. The method according to claim 1 , wherein the part of speech tags relate to adjectives, followed by nouns or related proper nouns.

6. The method according to claim 1 , wherein the specificity score value is determined by using a pre-trained static embedding matrix or by using context dependent embeddings originating from a transformer based-system.

7. The method according to claim 1 , wherein the adding a noun comprised in the text snippet to the remaining expressions comprises adding the noun to the remaining expressions only if a number of sub-tokens required for representing the noun is greater than a predefined sub-token threshold value.

8. The method according to claim 1 , wherein the using the remaining candidate expressions as training data also comprises adding to the training data manually labeled data.

9. The method according to claim 1 , wherein the using the remaining candidate expressions as training data also comprises adding to the training data annotated data from other sources.

10. The method according to claim 1 , wherein the pre-trained transformer-based language model is based on Bidirectional Encoder Representations from Transformers or a Generative Pre-trained Transformer.

11. A terminology extraction system for fine-tuning of a transformer-based language model by extracting terminology terms from a text document, the system comprising a processor and a memory, communicatively coupled to the processor, wherein the memory stores program code portions that when executed, enable the processor, to

identify a text snippet in the text document;

identify candidate multi-word expressions in the identified text snippet using part of speech tags;

determine a specificity score value for each of the candidate multi-word expressions, wherein the specificity score value is indicative of how unique the candidate multi-word expression is within the text snippet;

determine a topic similarity score value for each of the candidate multi-word expressions as a distance between an embedding vector corresponding to the identified text snippet and an embedding vector corresponding to the candidate multi-word expression;

select remaining expressions from the candidate multi-word expressions using a function of a specificity value and a topic similarity value of each of the candidate multi-word expressions;

add a noun comprised in the text snippet to the remaining expressions depending on a function of a correlation of the noun with a same lemma as one of the remaining expressions and a noun morphology score value of the noun to be added;

label the remaining multi-word expressions; and

fine-tune an existing pre-trained transformer-based language model using as training data the identified text snippet marked with the labeled remaining expressions.

12. The system according to claim 11 , wherein the program code instructions enable the processor during the selecting remaining expressions using a function of the specificity threshold value and a topic similarity value also to select remaining multi-word expressions from the candidate multi-word expressions if the specificity threshold value is greater than a predefined specificity threshold value and if the topic similarity value is greater than a topic similarity threshold value.

13. The system according to claim 11 , wherein the program code instructions enable the processor during the selecting remaining expressions using a function of the specificity value and a topic value also to select a noun of the identified text snippet if the noun has a score value that is greater than the lowest related multi-word expression score value.

14. The system according to claim 13 , wherein the program code instructions enable the processor also to use selection the noun if its score value is greater than the lowest related multi-word expression score value as a probabilistic approach to noun selection.

15. The system according to claim 11 , wherein the part of speech tags relate to adjectives, followed by nouns or related proper nouns.

16. The system according to claim 11 , wherein the program code instructions enable the processor during the determination the specificity score value also to use a pre-trained static embedding matrix or to use context dependent embeddings originating from a transformer-based system.

17. The system according to claim 11 , wherein the program code instructions enable the processor during the adding a noun comprised in the text snippet to the remaining expressions also to add the noun only to the remaining expressions if a number of sub-tokens required for representing the noun is greater than a predefined sub-token threshold.

18. The system according to claim 11 , wherein the program code instructions enable the processor during the using the remaining candidate expressions as training data also to add to the training data manually labeled data.

19. The system according to claim 11 , wherein the program code instructions enable the processor during the using the remaining candidate expressions as training data also to add to the training data annotated data from other sources.

20. A computer program product for fine-tuning of a transformer-based language model by extracting terminology terms from a text document, the computer program product comprising a computer readable storage medium having program instructions embodied the therewith, the program instructions being executable by one or more computing systems or controllers to cause the one or more computing systems to:

identify a text snippet in the text document;

identify candidate multi-word expressions in the identified text snippet using part of speech tags;

determine a specificity score value for each of the candidate multi-word expressions, wherein the specificity score value is indicative of how unique the candidate multi-word expression is within the text snippet;

determine a topic similarity score value for each of the candidate multi-word expressions as a distance between an embedding vector corresponding to the identified text snippet and an embedding vector corresponding to the candidate multi-word expression;

select remaining expressions from the candidate multi-word expressions using a function of a specificity value and a topic similarity value of each of the candidate multi-word expressions;

add a noun comprised in the text snippet to the remaining expressions depending on a function of a correlation of the noun with a same lemma as one of the remaining expressions and a noun morphology score value of the noun to be added;

label the remaining multi-word expressions; and

fine-tune an existing pre-trained transformer-based language model using as training data the identified text snippet marked with the labeled remaining expressions.