KEYword based Sampling (KEYS) for Large Language Models

Generating ¹¹1Authors contributed equallytext from Language Models requires us to decode (sample) words. To sample words at each time step, various text decoding algorithms are used such as greedy, beam search, temperature based, top-k, nucleus, etc. They have some downsides especially when the task is Question Answering. This is because when humans form answers to a question, they first gather keywords they know about that question and the subject using the knowledge they have and then formulate the answer. For example, If someone asks “What does Professor Bowman do?”, we’ll think of keywords such as ”Professor”, ”New York University”, ”Natural Language Processing”, ”GLUE”, etc and then use these keywords to finally formulate the answer. Mostly everything else in the answer such as grammar are just fillers.
Harabagiu et al. (2000) did information retrieval and generation using keywords for Question Answering task and this motivated us to explore keyword based decoding for human-like answer generation.
Our work shows that this decoding is done to simulate how humans generate answers to a question, which gives better scores for Q/A task. We focus on keywords, as language model gives distribution over vocabulary and to directly sample from that distribution is not an ideal approach when we are dealing factual Q/A. For the question, ”What is the capital of Australia?”. Language models would generate distribution over cities such as ”Sydney”, ”Perth”, ”Canberra”, etc. But the factually correct keyword that should be sampled is ”Canberra”. Keyword based sampling (KEYS) would re-weight the probability of ”Canberra” as it would be present in the context for that question which makes our answer factually correct.
This decoding method also introduces trustworthiness in the generated answer as it restricts the model to generate tokens that are allocated in it’s knowledge and domain and doesn’t include tokens which are not. If we carefully choose our knowledge base, then it will force the model to pick keywords from that knowledge base rather than generating keywords which are harmful, incorrect or biased. For example,for question ”What is the capital of Australia?”, knowledge base will have words such as ”Sydney”, ”Perth”, ”Canberra”, etc. But count of ”Canberra” in the in the article about ”capital of Australia” will have large number of keywords of ”Canberra” than ”Sydney”, ”Perth”. It ensures that ”Canberra” has higher weight than ”Sydney”, ”Perth” when the distribution is re-weighted. To ensure numerical stability, we normalized the count scores for each keywords by the highest frequency word.

There are number of sampling methods that are used in generative task such as greedy, beam search, temperature, top-k, nucleus sampling. Greedy sampling suffers from repetitions when decoding is done. Beam search is good but computationally expensive. Temperature based sampling Hinton et al. (2015) introduces a hyper-parameter in the distribution where higher values mean softer distribution. So the hyper parameter is meant to increase the probability for probable words for being picked and reduces for less probable words. The problem is that it doesn’t focus on special “keywords” we want to focus on while decoding. Nucleus sampling Holtzman et al. (2019) takes the idea of top-k sampling one step further such that the smallest possible ‘k’ words are chosen such that the summation of their probability is greater than some threshold. Hence, it focuses on very certain words model thinks is correct but it still fails to work for our hypothesis where focus should be given to special “keywords” too. Hence, we propose a sampling method which will help to prove/disprove the hypothesis.

As shown by Durmus et al. (2020), the generated text from a language model suffers from unfaithfulness which means that the generated text is inconsistent from the source text such as passage. These inconsistencies are not captured in existing metrics. Faithfulness is calculated by generating summaries of source text (human annotations). Language models suffer trade-off between attractiveness and faithfulness. In ideal case, we need both of them to be high. But they suffer due to systemic errors like random sampling & hallucinations. This is main idea in the introduction of Faithfulness Metric FEQA (Faithfulness Evaluation with Question Answering). This new metric gives us human annotated faithfulness scores.

Shuster et al. (2022) shows language models generate factually incorrect (but fluent) answers. Methods such as document aggregation to counter this problem were proposed but they result in factually incorrect response. SeeKeR Language Model uses a modular approach where one of the modules is to search the internet to generate the knowledge which is then used to generate the text. This approach outperforms state-of-the-art model according to human ratings of consistency which is an important method to generate answers which are more human-like. Hence, we will use one of the module to generate knowledge and sample keywords from this to generate outputs according to human ratings of consistency and thus will test our method to be more human-like.

Indurthi et al. (2017) shows a fascinating problem of generating question answer pairs from a given knowledge graph could have be used for several downstream tasks. To extract keywords from entities and relationships and finally generate a natural language question that has a unique answer; such language models which generate questions from keywords performed better (state-of-the-art) than template based methods for generating questions which motivates us to generate answer in this way which is in the form of keywords where attention is given to those keywords which is also is more human-like behavior.

The results from our experiments are shown above in Table 1. It shows that our method outperforms almost all other decoding/sampling methods when generating human-like answers to a question i.e., it performs better than 5 of the 7 decoding strategies. Most recent metrics such as BertScore and BartScore proves that KEYS decoding is better at generating human-like answers. As generating answers is similar to generating responses from chatbots, we will be applying KEYS to chatbots. It makes sense as chatbots should act like personal assistant which needs human-like behaviour and metrics which evaluates them should be proxies to human judgements Khayrallah et al. (2023).

We want to thank to Prof. Samuel Bowman for his invaluable guidance during the Natural Language Understanding course, where the initial conceptualization of this work was undertaken as a class project.