Systems and methods for language feature generation over multi-layered word representation

1. A computer-implemented method for outputting at least one cross-layer pattern to identify a target semantic phenomenon in a text document, the method comprising:

using at least one hardware processor for executing a code for:

extracting, for each word of at least some words of each training text fragment of a plurality of training text fragments designated as representing a target semantic phenomenon, a plurality of feature-values defined by respective layers;

statistically analyzing the plurality of feature-values identified for the plurality of training text fragments to identify at least one cross-layer pattern comprising a plurality of layers representing a common pattern for the plurality of training text fragments, the common cross-layer pattern defining at least one feature-value of a respective layer of at least one word and at least another feature-value of another respective layer of another word; and

generating instructions to present a marked human-readable text document by using the identified at least one cross-layer pattern for automatically marking at least one text fragment representing the target semantic phenomenon in a human-readable text document.

2. The method of claim 1 , further comprising:

training a statistical classifier to identify the target semantic phenomenon by matching or correlating between feature-values extracted from a new text fragment and the at least one cross-layer pattern; and

wherein the trained statistical classifier is used for analyzing the human-readable text document to identify the at least one text fragment representing the target semantic phenomenon.

3. The method of claim 2 , wherein the extracting of the plurality of feature-values defined by respective layers is performed for training text fragments designated as not representing the target semantic phenomenon, and the classifier is trained based on the feature-values extracted from the training text fragments designated as not representing the target semantic phenomenon.

4. The computer-implemented method of claim 1 , wherein the cross-layer pattern includes at least one negative feature-value that does not appear in a text fragment that includes the target semantic phenomenon.

5. The computer-implemented method of claim 1 , wherein each layer of the plurality of layers of the at least one cross-layer pattern is a member selected from the group consisting of: semantic, syntactic, domain knowledge, injection of knowledge by task expert, part-of-speech (POS) tag of the word, hypernym of the word, a named entity represented by the word, sentiment represented by the word, word appearing in a predefined lexicon.

6. The computer-implemented method of claim 1 , wherein the cross-layer pattern includes at least one word in the text fragment associated with multiple different layers.

7. The computer-implemented method of claim 1 , wherein the multiple different layers are combined for the at least one word.

8. The computer-implemented method of claim 1 , wherein the cross-layer pattern includes at least two different words in the text fragment each associated with a different layer.

9. The computer-implemented method of claim 1 , wherein the different layers associated with the at least two different words are defined by an order within the cross-layer pattern.

10. The computer-implemented method of claim 1 , wherein the target semantic phenomenon is a member of the group consisting of: a definition, a statement providing evidence for or against a topic, a statement made by an entity that something is the case about a topic without evidence, and a sentiment expressed by an entity about a topic.

11. The computer-implemented method of claim 1 , wherein the cross-layer pattern includes at least one defined gap between at least two layers each from a different word.

12. The computer-implemented method of claim 1 , wherein the cross-layer pattern is created by iteratively combining features to generate longer cross-layer patterns.

13. The computer-implemented method of claim 12 , further comprising applying a greedy analysis at the end of each iteration to identify the top predefined number of cross-layer patterns ranked according to probability of accurate prediction.

14. The computer-implemented method of claim 13 , wherein the top predefined number of cross-layer patterns are selected based on a correlation requirement with other previously selected higher ranking features.

15. The computer-implemented method of claim 12 , wherein combining features is performed by adding another feature of another word in combination and in order.

16. The computer-implemented method of claim 12 , wherein combining features is performed by adding another feature of the same word in combination.

17. A computer-implemented method for applying at least one cross-layer pattern to at least one text fragment to identify a target semantic phenomenon, the method comprising:

extracting a plurality of feature-values from at least some words in each text fragment of a human-readable text, each feature-value defined by a respective layer;

matching or correlating the plurality of feature-values with at least one cross-layer pattern; and

outputting an indication of the target semantic phenomenon in each respective text fragment when a match or correlation is found.

18. The computer-implemented method of claim 17 , wherein the matching or correlating with at least one cross-layer pattern is performed by applying a trained statistical classifier to the plurality of feature-values.

19. A system that identifies a target semantic phenomenon in text, comprising:

a data interface for receiving a plurality of training text fragment representing a target semantic phenomenon;

a program store storing code; and

at least one hardware processor coupled to the data interface and the program store for implementing the stored code, the code comprising:

code to extract, for each word of at least some words of the plurality of training text fragment, a plurality of feature-values defined by respective layers;

code to statistically analyze the plurality of feature-values to identify at least one cross-layer pattern comprising a plurality of layers representing a common pattern for the plurality of training text fragments, the common cross-layer pattern defining at least one feature-value of a respective layer of at least one word and at least another feature-value of another respective layer of another word; and

code to generate instructions to present a marked human-readable text document by using the identified at least one cross-layer pattern for automatically marking at least one text fragment representing the target semantic phenomenon in the marked human-readable text document.