Question answering system influenced by user behavior and text metadata generation

1. A computer-implemented method for handling dialogs based on user behavior data, the computer-implemented method comprising:

receiving, by a system comprising one or more processors, an input paragraph comprising one or more factual sentences, wherein each of the one or more factual sentences includes one or more words;

receiving, by the system, an input question comprising one or more words;

performing, by the system, a word-level gaze prediction on the input paragraph to identify one or more predicted eye-gaze attributes for the input paragraph, wherein the word-level gaze prediction comprises a prediction of a human gaze that would be performed by a human reading the input paragraph;

extracting, by the system, an answer to the input question based, at least in part, on the input paragraph, the input question, and the one or more predicted eye-gaze attributes of the input paragraph;

transmitting, by the system, the extracted answer;

wherein performing the word-level gaze prediction on the input paragraph includes passing outputs of each timestamp of a first set of bidirectional long short-term memory (LSTM) encoders through a first SoftMax layer that predicts the one or more eye-gaze attributes for each word of the input paragraph; and

performing, by the system, via a second set of bidirectional LSTM encoders, word-level gaze prediction on the input question to identify one or more predicted gaze attributes for the input question;

wherein the performing of the word-level gaze prediction on the input question includes passing outputs of each timestamp of the second set of bidirectional LSTM encoders through a second SoftMax layer that predicts one or more gaze attributes for each word of the input question; and

wherein the extracting of the answer to the input question is further based on the one or more predicted gaze attributes of the input question.

2. The computer-implemented method of claim 1 , wherein the extracting of the answer includes passing the outputs of the first and second set of bidirectional LSTM encoders to a memory network.

3. The computer-implemented method of claim 2 , wherein the one or more sentences of the input paragraph are encoded through the first set of bidirectional LSTM encoders based on the one or more predicted eye-gaze attributes of the input paragraph to yield a set of vector representations, wherein the set of vector representations includes a vector representation for each of the one or more sentences of the input paragraph, wherein the input question is encoded through the second set of bidirectional LSTM encoders based on the one or more predicted eye-gaze attributes of the input question to yield a vector representation of the input question.

4. The computer-implemented method of claim 3 , wherein the memory network is configured to:

generate a probability vector via a SoftMax operation based on superposing the encoded vector representation of the input question with the encoded vector representations of the input paragraph; and

generate a weighted sum vector by multiplying each of the encoded vector representations of the input paragraph with a corresponding probability value of the probability vector and summing products of the multiplication, wherein the extraction of the answer is based, at least in part, on the weighted sum vector.

5. The computer-implemented method of claim 4 , wherein the memory network is further configured to concatenate the weighted sum vector with the encoded question vector and pass the concatenated weighted sum vector through a dense layer and a third SoftMax layer, wherein the extraction of the answer is based, at least in part, on the concatenated weighted sum vector.

6. A computer program product for handling dialogs based on human eye-gaze behavior data, the computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a system comprising one or more processors to cause the system to perform a method comprising:

receiving, by the system, an input paragraph comprising one or more factual sentences, wherein each of the one or more factual sentences includes one or more words;

receiving, by the system, an input question comprising one or more words;

performing, by the system, word-level gaze prediction on the input paragraph to identify one or more predicted gaze attributes for the input paragraph, wherein the word-level gaze prediction comprises a prediction of a human gaze that would be performed by a human reading the input paragraph;

extracting, by the system, an answer to the input question based, at least in part, on the input paragraph, the input question, and the one or more predicted gaze attributes of the input paragraph;

transmitting, by the system, the extracted answer;

wherein performing the word-level gaze prediction on the input paragraph includes passing outputs of each timestamp of a first set of bidirectional long short-term memory (LSTM) encoders through a first SoftMax layer that predicts the one or more eye-gaze attributes for each word of the input paragraph; and

performing, by the system, via a second set of bidirectional LSTM encoders, word-level gaze prediction on the input question to identify one or more predicted eye-gaze attributes for the input question, wherein the performing of the word-level gaze prediction on the input question includes passing outputs of each timestamp of the second set of bidirectional LSTM encoders through a second SoftMax layer that predicts one or more eye-gaze attributes for each word of the input question, wherein the extracting of the answer to the input question is further based on the one or more predicted eye-gaze attributes of the input question.

7. The computer program product of claim 6 , wherein the extracting of the answer includes passing the outputs of the first and second set of bidirectional LSTM encoders to a memory network.

8. The computer program product of claim 7 , wherein the one or more sentences of the input paragraph are encoded through the first set of bidirectional LSTM encoders based on the one or more predicted eye-gaze attributes of the input paragraph to yield a set of vector representations, wherein the set of vector representations includes a vector representation for each of the one or more sentences of the input paragraph, wherein the input question is encoded through the second set of bidirectional LSTM encoders based on the one or more predicted eye-gaze attributes of the input question to yield a vector representation of the input question.

9. The computer program product of claim 8 , wherein the memory network is configured to:

generate a probability vector via a SoftMax operation based on superposing the encoded vector representation of the input question with the encoded vector representations of the input paragraph; and

generate a weighted sum vector by multiplying each of the encoded vector representations of the input paragraph with a corresponding probability value of the probability vector and summing products of the multiplication, wherein the extraction of the answer is based, at least in part, on the weighted sum vector.

10. The computer program product of claim 9 , wherein the memory network is further configured to concatenate the weighted sum vector with the encoded question vector and pass the concatenated weighted sum vector through a dense layer and a third SoftMax layer, wherein the extraction of the answer is based, at least in part, on the concatenated weighted sum vector.

11. A system for handling dialogs based on user behavior data, the system comprising one or more processors configured to perform a method comprising:

receiving, by the system, an input paragraph comprising one or more factual sentences, wherein each of the one or more factual sentences includes one or more words;

receiving, by the system, an input question comprising one or more words;

performing, by the system, a word-level gaze prediction on the input paragraph to identify one or more predicted gaze attributes for the input paragraph, wherein the word-level gaze prediction comprises a prediction of a human gaze that would be performed by a human reading the input paragraph;

extracting, by the system, an answer to the input question based, at least in part, on the input paragraph, the input question, and the one or more predicted eye-gaze attributes of the input paragraph; and

transmitting, by the system, the extracted answer;

wherein performing the word-level gaze prediction on the input paragraph includes passing outputs of each timestamp of a first set of bidirectional long short-term memory (LSTM) encoders through a first SoftMax layer that predicts the one or more eye-gaze attributes for each word of the input paragraph; and

performing, by the system, via a second set of bidirectional LSTM encoders, word-level gaze prediction on the input question to identify one or more predicted eye-gaze attributes for the input question;

wherein the performing of the word-level gaze prediction on the input question includes passing outputs of each timestamp of the second set of bidirectional LSTM encoders through a second SoftMax layer that predicts one or more eye-gaze attributes for each word of the input question; and

wherein the extracting of the answer to the input question is further based on the one or more predicted eye-gaze attributes of the input question.

12. The system of claim 11 , wherein the extracting of the answer includes passing the outputs of the first and second set of bidirectional LSTM encoders to a memory network.

13. The system of claim 12 , wherein the one or more sentences of the input paragraph are encoded through the first set of bidirectional LSTM encoders based on the one or more predicted eye-gaze attributes of the input paragraph to yield a set of vector representations, wherein the set of vector representations includes a vector representation for each of the one or more sentences of the input paragraph, wherein the input question is encoded through the second set of bidirectional LSTM encoders based on the one or more predicted eye-gaze attributes of the input question to yield a vector representation of the input question.

14. The system of claim 13 , wherein the memory network is configured to:

generate a probability vector via a SoftMax operation based on superposing the encoded vector representation of the input question with the encoded vector representations of the input paragraph;

generate a weighted sum vector by multiplying each of the encoded vector representations of the input paragraph with a corresponding probability value of the probability vector and summing products of the multiplication, wherein the extraction of the answer is based, at least in part, on the weighted sum vector;

concatenate the weighted sum vector with the encoded question vector; and

pass the concatenated weighted sum vector through a dense layer and a third SoftMax layer, wherein the extraction of the answer is based, at least in part, on the concatenated weighted sum vector.