Hallucination detection and handling for a large language model based domain-specific conversation system
Systems and methods are described herein for detecting and reducing hallucinations and improving the reliability of an LLM-based QA system. Particularly, the disclosure provides a set of hallucination detection and handling approaches that address different types of hallucinations of the QA system. The hallucination detection approaches described herein include comparing a natural language answer with context information by way of sentence similarity estimation and keyword matching. The hallucination handling approaches described herein include removing hallucinated sentences from the natural language answer or regenerating the natural language answer using better context information, depending on a level of hallucination detected in the natural language answer. A hybrid framework is also provided that systematically combines the hallucination detection approaches and hallucination handling approaches into one system to achieve an optimal hallucination-reduction performance for the QA system.
1 . A method of answering natural language questions, the method comprising:
receiving, with a processor, a natural language question;
retrieving, with the processor, natural language context information from an information source using a first information retrieval technique based on the natural language question;
generating, with the processor, a natural language answer using a language model based on the natural language question and the natural language context information; and
detecting, with the processor, hallucinated content in the natural language answer based on the natural language context information, the detecting the hallucinated including (i) comparing the natural language answer with the natural language context information, (ii) determining similarities between sentences in the natural language answer and sentences in the natural language context information, and (iii) labelling at least one sentence in the natural language answer as the hallucinated content in response to the at least one sentence having less than a threshold similarity with every sentence in the natural language context information.
2 . The method according to claim 1 , the determining the similarities further comprising:
determining a sentence embedding for each sentence in the natural language answer;
determining a sentence embedding for each sentence in the natural language context information; and
determining, for each respective sentence in the natural language answer, a respective embedding similarity between the respective sentence in the natural language answer and each respective sentence in the natural language context information, based on the sentence embedding for the respective sentence in the natural language answer and the sentence embedding for the respective sentence in the natural language context information.
3 . The method according to claim 2 , the determining the respective embedding similarity further comprising:
determining a respective cosine similarity between the sentence embedding for the respective sentence in the natural language answer and the sentence embedding for the respective sentence in the natural language context information.
4 . The method according to claim 2 , the determining the similarities further comprising:
determining, for each respective sentence in the natural language answer, a respective word pattern overlap rate between the respective sentence in the natural language answer and each respective sentence in the natural language context information.
5 . The method according to claim 4 , the determining the respective word pattern overlap rate further comprising:
determining a respective mapping between the respective sentence in the natural language answer and the respective in the natural language context information; and
determining the respective word pattern overlap rate between the respective sentence in the natural language answer and the respective in the natural language context information based on the respective mapping.
6 . The method according to claim 5 , the determining the respective word pattern overlap rate further comprising:
determining a respective overlap length as a sum of overlapping words between the respective sentence in the natural language answer and the respective in the natural language context information; and
determining the respective word pattern overlap rate between the respective sentence in the natural language answer and the respective in the natural language context information by dividing the respective overlap length by a number of words in a shorter sentence of the respective sentence in the natural language answer and the respective in the natural language context information.
7 . The method according to claim 6 , the determining the respective overlap length further comprising:
determining sum of overlapping words between the respective sentence in the natural language answer and the respective in the natural language context information as only including words in a sequence of overlapping words having at least a predetermined minimum number of words.
8 . The method according to claim 4 , the labelling the at least one sentence in the natural language answer as the hallucinated content further comprising:
labelling the at least one sentence in the natural language answer as the hallucinated content in response to:
(i) the respective embedding similarity between the at least one sentence in the natural language answer and each sentence in the natural language context information being less than a first predetermined threshold; and
(ii) the respective word pattern overlap rate between the at least one sentence in the natural language answer and each sentence in the natural language context information being less than a second predetermined threshold.
9 . The method according to claim 1 , the detecting the hallucinated content further comprising:
identifying keywords in the natural language answer;
determining whether the keywords are also present in the natural language context information; and
labelling the natural language answer as including the hallucinated content in response an amount of the keywords in the natural language answer not being present in the natural language context information exceeding a third predetermined threshold.
10 . The method according to claim 1 further comprising:
generating, with the processor, a modified natural language answer that reduces the hallucinated content in the natural language answer.
11 . The method according to claim 10 , the generating the modified the natural language answer further comprising:
generating the modified natural language answer by deleting the hallucinated content from the natural language answer.
12 . The method according to claim 11 further comprising:
determining, with the processor, a hallucination level that quantifies an amount of the hallucinated content in the natural language answer,
wherein the deleting the hallucinated content from the natural language answer is performed in response to the hallucination level being less than a predetermined threshold.
13 . The method according to claim 10 , the generating the modified the natural language answer further comprising:
retrieving modified natural language context information using a second information retrieval technique based on the natural language question, the second information retrieval technique being different than the first information retrieval technique; and
generating the modified natural language answer using the language model based on the natural language question and the modified natural language context information.
14 . The method according to claim 13 further comprising:
determining, with the processor, a hallucination level that classifies an amount of the hallucinated content in the natural language answer,
wherein the retrieving modified natural language context information and the generating the modified natural language answer based on the modified natural language context information are performed in response to the hallucination level being greater than a predetermined threshold.
15 . The method according to claim 13 , wherein the retrieving modified natural language context information and the generating the modified natural language answer based on the modified natural language context information are repeated using different information retrieval techniques in each case, until the modified natural language answer includes less than a threshold amount of hallucinated content.
16 . The method according to claim 13 , wherein the retrieving modified natural language context information and the generating the modified natural language answer based on the modified natural language context information are repeated using different information retrieval techniques in each case, until no further information retrieval techniques result a reduced amount of the hallucinated content in the modified natural language answer.
17 . The method according to claim 10 further comprising:
outputting, with an output device, the modified natural language answer to a user.
18 . The method according to claim 17 further comprising:
identifying, with the processor, sentences in the natural language context information that provide support for sentences in the modified natural language answer; and
outputting, with the output device, indications of the sentences in the natural language context information that provide support for sentences in the modified natural language answer.
19 . A method of answering natural language questions, the method comprising:
receiving, with a processor, a natural language question;
retrieving, with the processor, natural language context information from an information source using a first information retrieval technique based on the natural language question;
generating, with the processor, a natural language answer using a language model based on the natural language question and the natural language context information; and
detecting, with the processor, hallucinated content in the natural language answer based on the natural language context information, the detecting the hallucinated including (i) comparing the natural language answer with the natural language context information, (ii) identifying keywords in the natural language answer, (iii) determining whether the keywords are also present in the natural language context information, and (iv) labelling the natural language answer as including the hallucinated content in response an amount of the keywords in the natural language answer not being present in the natural language context information exceeding a third predetermined threshold.
20 . A method of answering natural language questions, the method comprising:
receiving, with a processor, a natural language question;
retrieving, with the processor, natural language context information from an information source using a first information retrieval technique based on the natural language question;
generating, with the processor, a natural language answer using a language model based on the natural language question and the natural language context information;
detecting, with the processor, hallucinated content in the natural language answer based on the natural language context information; and
generating, with the processor, a modified natural language answer that reduces the hallucinated content in the natural language answer, the generating the modified the natural language answer further including (i) retrieving modified natural language context information using a second information retrieval technique based on the natural language question, the second information retrieval technique being different than the first information retrieval technique, and (ii) generating the modified natural language answer using the language model based on the natural language question and the modified natural language context information.