Querying Knowledge via Multi-Hop English Questions (2019)
huyi / December 2022
- 0. Abstract
- 1. Introduction
- 2. The KALM system
- 3. The MetaQA Dataset and Multi-Hop Questions
- 4. Constructing a KALM Parser via Structural Learning
- 5. Capturing the Meaning of Multi-Hop Questions in Logic
0. Abstract
difficulties to KRR
obstacles on the way to make AI based on the knowledge representation and reasoning (KRR):
- inherent difficulty of knowledge specification
- the lack of trained specialists
current work
用的方法不是 machine learning 范畴的
- based on KALM
- KALM-QA
1. Introduction
what does KRR need
- Knowledge can be described by logical facts and rules, and processed by intelligent knowledge representation and reasoning (KRR) systems.
- such knowledge is hard to extract
controlled natural languages (CNL)
pros of CNL:
- easy to understand
- usable as knowledge as subsets of NLs with restricted grammars and interpretation rules, their meaning can be captured in logic precisely
cons of CNL:
- restrictive grammars could be problematic
- knowledge authoring enabling humans to write in NL that machines could understand
- knowledge acquisition enabling machines to understand the NL that humans write
- CNLs do not assume any kind of background knowledge example: Zoe Saldana appears in Avatar and Zoe Saldana is an actress of Avatar would have different logical representations and a query like Avatar has which actresses will yield no answers
- 然而这种缺失的 background knowledge 过多，难以补全–> KALM希望可以把这个步骤做得 scalable and feasible while still relying on CNL
introduction of KALM
- supports knowledge authoring and simple question answering with very high accuracy compared to the SoTA machine learning approaches
- a sophisticated semantic layer on top of ACE
contributions
KALM-QA
2. The KALM system
- Kalm input: ACE (Attempto Controlled English)
- KALM ensures that semantically equivalent sentences have the same logical representation
- architecture
Syntatic parsing
Attempto Parsing Engine (APE) as the top level parser
- input and output
extracts:
- 每个词的词性
- grammatical relations
–> a list of stylized first-order terms: discourse representation structure, or DRS
- example:
frame-based parsing
-
frame represents one or more related semantic relationships among entities, where each entity plays a particular role
-
example: MOVIE frame
其中bn是同义词编号(based on BabelNet synsets) 一些role里面有严格的 type constraint
- Ivp (logical valence patterns)
a lexical unit & a set of grammatical patterns, each associated with a frame role
- A grammatical pattern represents a grammatical relation between the lexical unit and a frame role.
- An lvp is applicable to a sentence if
- the sentence contains the lexical unit
- the role-fillers for each frame role can be extracted based on the respective grammatical patterns.
- example:
- we call the extracted relation a candidate parse
Role-filler Disambiguation
一个verb可能对应很多不同的frame，要看两边的名词能不能fit frame 对应的 role filler
Constructing logical forms
sematic relations extracted via Ivps –>
- facts: unique logical representations (ULR)
- queries: unique logical representation for queries(ULRQ) for questions
details later
3. The MetaQA Dataset and Multi-Hop Questions
metaQA 对重复名字分辨不出来，里面有很多mislabeled questions
4. Constructing a KALM Parser via Structural Learning
这一节讲 frame-semantic parser 逐层的学习结构。
先看用来做 role-filler 的 grammer patterns 是怎么学出来的；再看 Ivp是怎么在 training sentence 上 generate 的
4.1 Learning Grammatical Patterns for Role -filler Extraction
用以下的 prolog extraction rules:
- grammatical patterns and the corresponding extraction rules: learned
- structure-learning process:
- ACE(一种 CNL) –> DRS
- pairs in DRS (lexical unit, role-filler) –$^{KALM}$–> extraction rules
- construction of the library of utility rules
- reachability reasoning in DRS
- construction of the actual extraction rules
utility predicates and rules
reachability reasoning in DRS
embed the DRS in a graph structure:
- For each term in DRS, create a node in the graph.
- For any pair of nodes n1, n2 that represent the DRS subterms p1, p2 that share a variable, create a directed labeled edge n1 → n2 and back. The label of the edge n1 → n2 is the position of the shared variable in p1 and the edge n2 → n1 is labeled with the position of that variable in p2.